Frozen Language Model

Follow this topic Ask about this topic

Momentum

7 papers in the last four weeks, down 12% on the four weeks before. 0.1% of all new papers.

Jul 6Week of Sep 21

Latest papers 53

Sep 29, 2026eess.AS

Voices as Handles: Reasoning about Speaker Identity with Frozen Text LLMs

Multi-user voice agents must track who said what across dialogue sessions. Text LLMs are attractive backbones for such agents, but transcripts alone do not expose acoustic speaker identity, leaving the model without a persistent reference for linking information to speakers across sessions. We address this gap by introducing Speaker Handles, soft-token representations that expose acoustic speaker identity to a frozen text LLM for cross-session speaker-dependent reasoning. A three-stage curriculum trains a lightweight projector, with fewer than 0.1% of the backbone's parameters, to map speaker embeddings into these handles. Establishing whether the resulting handles truly support cross-session speaker-dependent reasoning is challenging with existing benchmarks because textual cues can partially reveal fact ownership. We therefore present SpeakerBind, a controlled shared-agent benchmark in which overlapping facts across users require correct cross-session speaker attribution. Speaker Handles achieve 97.40-98.36% accuracy on VoxCeleb1 and 70.40% on SpeakerBind, close to the 71.88% topline. These results show that the proposed Speaker Handles provide an efficient way to integrate acoustic speaker identity into frozen text LLMs for speaker-content reasoning.

Runqiu Xu, Zhisheng Zheng, David Harwath

The University of Texas at Austin, Austin, TX, USA

Voice Agents Speaker Frozen Language Model Strong Large Language Model Soft-Token Representations

Sep 29, 2026cs.AI

Locating Answer-Correctness Signals in Frozen Large Language Models

Language models expose internal signals that predict whether an answer is correct, readable from a single forward pass of a frozen model without additional generations. Yet existing probes often commit to one signal family or layer and can be brittle under distribution shift; in retrieval-augmented settings, many specialized detectors instead target passage faithfulness, which can diverge from correctness when retrieved evidence is unhelpful or conflicting. We therefore ask where answer correctness is readable, which internal signal families carry it, and how they should be combined. We search over hidden states, token probabilities, residual-stream features, attention, and their fusion, treating the selected readouts as a predictive measurement rather than a mechanistic localization. We run this analysis separately in closed-book and with-context settings, since context can change which readouts are informative. A consistent anatomy emerges: correctness concentrates in the answer span, recovered from the answer tokens even under retrieval, and the families carry it complementarily, so fusing them helps most out of distribution, where a single signal is weakest. The protocol is effective across two backbones and gates a retrieval controller as one downstream use.

Yuansen Liu, Yixuan Tang, Anthony Kum Hoe Tung

School of Computing, National University of Singapore

Frozen Language Model Correctness Answer Private Signals

Sep 29, 2026cs.CL

NinaXander: Feasibility and Limits of Composing Frozen Language Models Across Architecture Families via a Shared Latent Space

In this paper we propose NinaXander, a series of composed language models obtained by connecting layers of frozen language models from different architecture families with a single trained shared-latent adapter. A composed model runs the first layers of one model, converts the resulting intermediate representation once with the adapter, and then runs the remaining layers of the other model. Once the adapter is trained, several composed models that connect at different layers are obtained without retraining. Using the recurrent RWKV-4-Raven-7B and the Transformer-based Tulu-Pythia-6.9b, abbreviated as RWKV and Pythia, this study examines whether frozen models from different families can be recombined post hoc. The composed models answered multiple-choice questions, and those whose generations we examined produced syntactically well-formed text. The configuration that combines the first 5 layers of Pythia with the remaining 27 layers of RWKV reduced the Transformer key-value (KV) cache by 84.4% with accuracy not significantly different from that of RWKV alone. In multiple-choice accuracy, however, no composed model matched the parent model Pythia, and language-modeling performance decreased sharply on WikiText, a corpus of Wikipedia articles outside the training domain. The correspondence between intermediate representations was also obtained in one favorable case, with a shared tokenizer, the same depth, and the same hidden width, and does not show that the models share a general semantic space.

Takanori Kotama, Shun-ichiro Hayashi, Daichi Mukunoki +2

Graduate School of Informatics Nagoya University · Information Technology Center Nagoya University

Frozen Language Model

Sep 29, 2026quant-ph

SQUARE: Structured Quantum Representation Adapters as Compact Quadratic Feature Maps for Frozen Language Models

Frozen language models (LMs) are increasingly used as fixed feature extractors for downstream reranking, scoring, and preference modeling, raising a practical question: how should a compact module represent interactions among features in a fixed low-dimensional bottleneck? Common linear and low-rank adapters remain linear at the adaptation module itself, whereas explicit second-order alternatives introduce pairwise interactions through direct parameterization or predefined factorizations. We propose SQUARE, a Structured QUAntum REpresentation adapter that amplitude-encodes the bottleneck vector, applies a parameterized quantum circuit, and measures the resulting state. We show that each basis-probability feature is exactly a normalized quadratic form in the bottleneck coordinates, while the additional Pauli-

Z

readouts are signed linear combinations of these probabilities. The measured map can therefore parameterize interactions over

O(d^2)

coordinate pairs through a small set of shared circuit parameters, where

d

is the bottleneck dimension. It provides a structured parameterization within, rather than beyond, the classical normalized-quadratic feature class. In a disjoint same-pipeline evaluation over eight GLUE-derived controlled interaction tasks and five shared seeds, SQUARE achieves an average test accuracy of

0.7565

, compared with

0.7355

for an affine normalized-quadratic predictor,

0.7271

for the evaluated parameter-matched Givens mixing model,

0.6817

for an MLP, and

0.6155

for a frozen-circuit control. Under reduced supervision, it also shows consistent gains over the strongest evaluated classical comparator, with the same qualitative pattern across multiple frozen LM backbones. All circuit experiments use simulation, while the learned feature map can be evaluated exactly in batched PyTorch without quantum hardware.

Emily Jimin Roh, Hyojun Ahn, Hoyeong Lee +4

Korea University · Sookmyung Women’s University · Ulsan National Institute of Science and Technology +2

Large Language Model Quantization Frozen Language Model Frozen

Sep 27, 2026cs.LG

BOReFT: Manifold Steering of Language Models for Black-box Optimization

Language models are increasingly used as proposal models for black-box search, from program optimization to molecular design. Existing approaches typically improve proposals through iterative prompting or parameter updates, offering limited control over how completely and efficiently the model's search space is explored. Continuous optimization methods, such as Bayesian optimization, provide a principled way to search but require a suitable domain to operate over. To address this, we introduce BOReFT, which learns a compact, low-dimensional space of hidden-state interventions in a frozen language model, and uses this space as the search domain for Bayesian optimization with an external scoring function. Empirically, we find that the learned domain spans semantic regions and exhibits smoothness properties that support search. Theoretically, we show that semantic coverage and interpolation control the best score available in the learned space, and that decoding from this space yields a standard stochastic-bandit observation model for adaptive search. We evaluate BOReFT on the interpretable word search task "Semantle" and on three more real-world discovery tasks in de novo molecule property optimization. Compared to strong LLM baselines, BOReFT finds in Semantle a higher number of hidden targets and, on two out of three molecular objectives, achieves higher property scores. Consequently, our method provides a principled new bridge between discrete proposal spaces of LLM-based search and continuous black-box optimization.

Dhruv Agarwal, Rico Angell, Kavitha Srinivas +4

University of Massachusetts Amherst · New York University · IBM Research +1

Bayesian Optimization Black-Box Optimization Frozen Language Model Molecular Optimization Black Box Steering

Sep 24, 2026cs.LG

Minimally Invasive Steering of Language Models

Pre-logit steering adapts a frozen language model to a test-time reward by adding vectors to its final hidden states. Unregularized reward optimization can substantially alter the output distribution and degrade generation quality. We propose Minimally Invasive Steering Vector Optimization (MISVO), which penalizes interventions using the local KL geometry of the induced token distribution. The resulting Fisher quadratic measures distributional sensitivity and admits an analytic gradient computed through matrix--vector products with the frozen language-model head. We derive an exact decomposition of the sequence-level KL gradient into an analytic Fisher term and a suffix score-function term. For a fixed generation horizon, we show that the suffix term is second order in the steering magnitude and that three Fisher surrogates agree with the full KL gradient to first order. MISVO uses the frozen-reference surrogate to optimize position-specific interventions without updating model parameters. Across preference and code-generation tasks on models with approximately 1B--14B parameters, MISVO achieves the highest mean reward in six of seven model--task settings, with diversity and coherence scores close to those of Best-of-N.

Taha Entesari, Jingyu Zhang, Daniel Khashabi +1

Johns Hopkins University

Steering Frozen Language Model Kullback-Leibler Divergence

Sep 24, 2026eess.AS

Learning New Words from Unlabeled Test Data in Automatic Speech Recognition

New words are invented every day. A human listener can learn a new word by hearing it clearly once and inferring its usage from sentence context. This paper proposes granting ASR a similar ability to learn the contextual representations and spellings of new words from unlabeled test data at test time. A frozen CTC acoustic model provides spellings, a frozen language model provides contextual evidence for out-of-vocabulary (OOV) word detection, and an adaptation module expands the vocabulary by learning the lexical token representations with distributions over CTC-generated candidates. The spelling model of each token is optimized by minimizing a Kullback-Leibler divergence (KLD) objective. We demonstrate that the CTC-weighted language model log likelihood ratio can be interpreted as the KLD between the unknown correct ASR and the unsupervised learned ASR, and that, using a Pinsker bound, the square root of KLD can be interpreted as an upper bound on the total variation distance between the true and estimated spelling of the unknown word. Experiments show relative OOV character-error-rate reductions of up to 14.97% on LibriSpeech and 6.67% on dysarthric Speech Accessibility Project data for recurring OOV words, relative to the corresponding rescoring system.

Mengqi Wang, Mark A. Hasegawa-Johnson, Haolong Zheng +1

University of Illinois at Urbana-Champaign, USA · Korea Advanced Institute of Science and Technology, South Korea

Automatic Speech Recognition Accents Unlabeled Data Frozen Language Model

Sep 23, 2026cs.LG

Learning from Failures: Heterogeneous Graph Memory for Small Language Model Tool-Using Agents

Small and medium-sized language models offer cost-effective executors for tool-using agents, making them attractive for local and large-scale deployment. However, in long-horizon and stateful environments, they often make structural errors such as missing required observations, performing premature writes, repeating failed calls, and violating action preconditions. These errors can lead to incorrect state updates, policy violations, and costly or irreversible consequences, making reliable tool execution a critical deployment challenge. Existing fine-tuning approaches require substantial data and computation, while flat memory may retrieve failed actions without preserving their causal context or safety conditions. In this paper, we propose FRESH, a Failure-aware Retrieval framework over Experience-Structured Heterogeneous graphs, which transforms historical successes and failures into structured external experience for tool-using agents. By explicitly modeling the dependencies among tasks, actions, errors, repairs, and execution conditions, FRESH helps frozen language models reuse reliable strategies, avoid recurring failures, and make safer decisions in stateful tool interactions. Experiments on

τ

-Bench and AppWorld with multiple open-source models show that FRESH consistently improves task success and tool-use reliability over no-memory agents and representative memory-based baselines.

Jiaxing Li, Lei Song, Rui Dong +1

School of Computer Science and Engineering, Southeast University, Nanjing, China

Large Language Model Agents Frozen Language Model Strategic Agents

Sep 17, 2026cs.AI

GraphSkillAA: Attribution-Guided Skill-Graph Updating with Targeted Validation and Rollback

External skills provide domain knowledge and procedures without updating model parameters, but flat collections obscure skill applicability, dependencies, and composition. Graphs organize skills into addressable nodes and explicit relations, supporting selection and composition. Yet existing skill-graph methods use this structure mainly for retrieval, rather than to localize updates, scope retesting, or precisely roll back rejected changes. We introduce GRAPHSKILLAA (GraphSkill Abductive Attribution), which uses one addressable graph for skill selection, execution, failure attribution, targeted updating, validation, and rollback. Nodes separate applicability, execution, and exclusion conditions; typed edges encode prerequisite and enhancement relations. The frozen student records used nodes and edges, while the teacher contrasts related successes and failures to route each supported repair to the smallest relevant field or relation; execution lapses or insufficient evidence leave the graph unchanged. A Local Gate retests affected examples, while a Big Gate evaluates the merged graph on the complete update pool; rejected changes are rolled back. With GPT-5.6-sol, GRAPHSKILLAA reaches 81.5%, 66.7%, and 91.2% on SearchQA, LiveMath, and DocVQA, respectively, and attains the highest observed mean in every main setting. These results show that object-level attribution and graph-scoped validation make a skill graph a locally optimizable, testable, and reversible external state.

Ziqiao Shang, Ling-Yue Ge, Lan-Zhe Guo

National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China · School of Intelligence Science and Technology, Nanjing University, Suzhou, China

Unsupervised Skill Discovery Frozen Language Model Search-Augmented Reasoning Rollback Validation

Sep 14, 2026cs.AI

Safe Error Correction for Language Models: Frozen-Base Adjustment with Capability Preservation

We study a practical question: can a small correction module fix errors in a frozen language model's outputs without degrading its base capabilities? We propose CRN v2, a lightweight logit-level correction module (~34M trainable parameters, 0.73% of the 4.65B text module) that sits atop a fully frozen Gemma 4 E2B model. The base model is never updated; only the correction module learns, via supervised fine-tuning followed by reference-free DPO on 83,400 error-correction pairs. On a 60-question domain exam (CEHRI: Certified Human-Robot Intelligence, covering facts, arithmetic, and implicit-goal reasoning), CRN v2 corrects 53.3% of base-model errors (reworded variant: 43.3%) while showing no degradation on tested capability benchmarks (MMLU/BoolQ N=200; car-wash N=8). A LoRA baseline at the matched CRN v1 budget (6.6M params, rank 19) achieves 83.3% correction but suffers 30-75% capability loss on the same benchmarks -- the correction-capability tradeoff. An ablation shows that the KL preservation term (lambda=0.1) is critical: lowering it to 0.01 degrades correction to 35.0%. A hidden-state injection variant at earlier layers (1.6M params, SFT-only) reaches 50.0%/55.8% but does not exceed logit correction; shallower injection (layer 4) drops to 30.0%/28.3%; multi-depth logit correction (~35M) reaches only 40%; and longer training (5,000 SFT + 2,000 DPO) stays at 53.3% -- none of the alternative configurations we tested exceeded the rank-128 logit result, consistent with a best-achieved result of ~53% rather than a floor. This is a study of a design principle (frozen base + logit correction + KL anchoring), not a claim of architectural novelty. All code, main-result weights, and evaluation scripts are released (deep variant as code only -- no trained deep checkpoints).

Gautam Kishore

Eulogik

Frozen Language Model Correction Error Correction Logit Lens

Sep 14, 2026cs.LG

The Router Within: Eliciting Native Skill Routing from a Frozen LLM

Skills extend an LLM agent beyond its parametric knowledge, and the gain they promise rests on picking the right one. Deployed harnesses route by preloading every skill's metadata into the context, which disperses the agent's attention and caps the library size. Retrieval pipelines move the selection out of the context, but also out of the agent's capability. We show that the frozen agent LLM already carries the routing signal in its own forward passes, and that two linear maps suffice to read it out with no skill text in the context. Our Gavel (Glance And Verdict from a frozen LLM) reads it in two steps. A glance scores the full library by matching the task's mid-layer states against a compact bank that one forward pass builds for each skill at installation, with the two maps as the only trained parameters. A verdict then resumes each shortlisted skill's forward pass, reads the model's own likelihood and yes/no judgment, and fuses both with the glance as a product of experts. Trained once, Gavel transfers zero-shot to three public benchmarks and SkillTraj, our new benchmark of 372 simulated agent trajectories. On Qwen3-32B it outperforms progressive disclosure and retrieve-and-rerank pipelines that add 1.2B to 16B external parameters, by up to 13.4 points on written tasks and up to 21.9 when the need for a skill arises mid-rollout. Routing accuracy improves as the backbone does, and in a bash-agent harness Gavel lets the 32B trigger the right skill on Skill-Use more often than models of up to 1.6T parameters in Codex.

Ruishuo Chen, Xun Wang, Yu Chen +2

Institute for Interdisciplinary Information Sciences, Tsinghua University

Skills Large Language Model Routing Large Language Model Agents Frozen Language Model

Aug 17, 2026cs.CL

Frozen Memory Is Not Enough: Rethinking External Memory as Extraction

Methods for improving knowledge use in large language models typically fall into two regimes. Non-parametric retrieval offers flexible access to external knowledge, but adds retrieval latency, context overhead, and only shallow integration with the backbone. Parametric adaptation is efficient at inference time, but entangles knowledge with model weights and can be hard to update, audit, or transfer. Engram-style hashed memory occupies a middle regime: it stores learned information in an external, addressable table, yet consumes that table through a small learned reader. This raises a basic question: when such a memory is moved across backbones, what matters more, the frozen memory itself or the target-side reader? We study this question through cross-model frozen-memory extraction, in which a memory trained on a source model is frozen and attached to a different target model, with only a lightweight reader trained. Ablations show that learned memory content and correct addressing both matter, but the transferred table becomes useful only through a reader aligned to the target model. In downstream question answering tasks, a dual-layer, four-branch reader nearly closes the gap between same-model and cross-model reuse, achieving an average score of 38.8 under our controlled evaluation protocol. Moreover, when the provider reader is directly compatible with the target interface, the frozen artifact can provide substantial utility without target-side training, while optional reader adaptation yields further improvement. These results suggest that Engram can serve as a reusable external knowledge artifact, provided that the target has access to a compatible reader interface; target-side adaptation can further improve alignment when direct reader reuse is insufficient.

Mingyuan Li, Guangsheng Yu, Xu Wang +1

ELLIS Institute Finland · University of Turku · University of Technology Sydney

Peak Memory Cross-Encoder Frozen Language Model Cross-Task Knowledge Transfer Extraction

Aug 13, 2026cs.CL

Prompts in the Wild: A Large Analyzed Collection of Transactional Prompts in Code

The behavior of contemporary generative Large Language Models (LLMs) is directly shaped by prompts, unstructured texts that describe the desired output and model behavior. In this paper we argue that prompts are linguistic objects that merit investigation in their own right. To this end, we collect 57.5K unique samples of prompts from GitHub. Specifically, we focus on transactional prompts: reproducible natural language instructions that are integrated into software. To enable the empirical, quantitative study of prompts, we introduce a structured ontology, capturing the properties of prompts as well as their formal and semantic components. Based on this ontology, we transform prompts from unstructured raw texts into richly structured linguistic objects. Analysis of these structured data reveals significant diversity of usage patterns across languages, domains, tasks, and modalities, in a typical Zipf-like distribution where some clearly prevail and others, more diverse, appear in the long tail. To validate the reliability of the ontology-based annotation of the prompts, we perform a comprehensive error analysis across all fields, providing a detailed assessment of annotation quality. We release the dataset together with a browsing and exploration interface (https://github.com/OnlpLab/transactionalPromptsCollection ).

Victoria Basmov, Yoav Goldberg, Reut Tsarfaty

Bar-Ilan University · Allen Institute for Artificial Intelligence

Prompt Engineering Frozen Language Model Unstructured Data Instruction

Aug 12, 2026cs.CL

DIVE: Unlocking Self-Improvement in Frozen Language Models Through Diversity-Driven Skill Evolution

Large language models (LLMs) cannot retain post-deployment experience without parameter updates. We introduce DIVE, a diversity-driven framework that enables frozen LLMs to improve by evolving persistent natural-language skills from task experience and verifier feedback. These skills encode reusable reasoning procedures, verification strategies, common failure modes, and output constraints and are both executed and revised by the same underlying model without access to a teacher model. Since natural-language skill evolution is a stochastic, non-convex search process, optimizing a single skill trajectory can overfit to sampled experience or converge to a suboptimal solution. DIVE mitigates this optimization variance by independently evolving multiple skill populations from bootstrapped experience, adaptively refining them through diverse transformations, and jointly selecting a complementary set of skills. Across six mathematical and logical reasoning tasks and multiple model families, DIVE consistently outperforms existing reasoning methods, prompt-optimization approaches, skill-development frameworks, and memory-based baselines. It achieves rapid self-improvement from accumulated experience, obtaining substantially larger performance gains with fewer rollouts than parameter-based methods such as SFT and GRPO, and prompt optimization with GEPA. Further, the resulting skills transfer across model scales and families, enabling smaller models such as GPT-5-nano to match or outperform larger counterparts, i.e., GPT-5, under conventional prompting. These results establish diversity-driven skill evolution as an effective, interpretable, and parameter-free approach to LLM self-improvement.

Siheng Xiong, Ali Payani, Oguzhan Gungordu +1

Georgia Institute of Technology · Cisco Research

Self-Evolving Skill Libraries Frozen Language Model Recursive Self-Improvement Diversity Skills Natural Language

Aug 12, 2026cs.CL

LODESTAR: Trustworthy Entropy Is Navigated, Not Merely Measured -- Reinforced Polarizer Keeps a Frozen LLM from Being Confidently Misled by the Wrong Evidence

Predictive-distribution entropy makes a strong selection rule in retrieval-augmented question answering: across five QA benchmarks, keeping the candidate answer that a frozen respondent LLM produces with the lowest answer-token entropy lifts mean answer

F_1

from 0.4769 to 0.5148 over the retriever's top-ranked passage, with no gold answers. Yet this lowest-entropy rule, which prior entropy-based selectors adopt, fails in a specific and consequential way: a misleading passage makes the respondent confidently wrong, driving its entropy down precisely where the signal looks most trustworthy. We show that the failure comes from the passage the respondent reads -- and the context that passage is read in is an input we can intervene on. We introduce LODESTAR, to our knowledge the first method to score a text intervention by the uncertainty it induces in a third-party frozen respondent, compared across one question's candidates. LODESTAR uses reinforcement learning to train, once and offline, a polarizer -- a short fixed natural-language string inserted into the respondent's prompt and never into its weights; its training labels are built offline from gold answers and two LLM judges, and inference reads neither. Evaluating every competing selector under the same frozen respondent and the same candidate pools on 5,008 questions, LODESTAR attains the highest mean

F_1

of any inference-ready selector (0.5148 to 0.5339), the highest exact match (0.4136), and the highest GPT-4o judge score of the frozen-respondent configurations judged (0.6435); its three-seed mean wins all 70 method-by-dataset

F_1

cells against fourteen published configurations while remaining paired-significant against every one. The gain holds both in-domain and out-of-domain, and ablating the polarizer shows it is what makes the respondent read a misleading passage less often (26.0% against 30.3%).

Po-Jen Ko, Che-Cheng Wu, Hung-Chun Hsu +2

1Research Center for Information Technology Innovation, Academia Sinica, Taiwan

Semantic Entropy Large Language Model Responses Token-Level Entropy Natural Language Inference Frozen Language Model

Aug 10, 2026cs.CL

ZetaGPT: A Reference Implementation of Positional--Encoding--Free State--Space--Attention Language Models

Transformer-based language models rely on self-attention, whose computation is permutation-equivariant and therefore lacks an intrinsic mechanism for representing token order. Existing architectures address this limitation by explicitly incorporating positional information through learned positional embeddings or hand-crafted positional encodings, such as rotary positional encoding (RoPE), treating positional information as an architecturally acquired capability rather than an inherent property of the model. Motivated by the pursuit of positional-encoding-free architectures, this work explores a language model architecture that integrates causal state-space equations to implicitly encode positional information before attention computation. Specifically, each model block applies a causal state-space equation before self-attention, allowing recurrent state dynamics to encode sequential information into token representations. Consequently, subsequent attention layers operate on position-aware representations without requiring explicit positional encodings while retaining the expressive modeling capacity of self-attention. We present \textsc{ZetaGPT}, a compact hybrid language model designed for research, rapid prototyping, algorithm verification, and educational applications. In addition to the proposed architecture, \textsc{ZetaGPT} provides a fully open-source, end-to-end training pipeline encompassing dataset construction, tokenizer training, pretraining, supervised fine-tuning, reinforcement learning from human feedback (RLHF), and chain-of-thought (CoT) reasoning via pure reinforcement learning. To the best of our knowledge, \textsc{ZetaGPT} is the first open-source small language model without explicit positional encoding and establishes a compact, reproducible reference implementation for the development and empirical study of positional-encoding-free language models.

Róisín Luo

University of Galway, Ireland

Positional Encoding Linear Attention Frozen Language Model Recurrent State Openai

Aug 10, 2026cs.AI

ChronoState: Hidden Elapsed-Time Conditioning for Temporal-State Action Selection in Frozen-Backbone Language Models

Temporal decisions in language-model systems often depend on both symbolic task state and elapsed wall-clock time, such as cache expiration, job completion, quota resets, deadlines, or stale sessions. We study whether elapsed time can be supplied as a non-token, system-side scalar and composed with visible symbolic state by a frozen-backbone language model. We introduce ChronoState, a compositional temporal-state benchmark in which symbolic state appears in the prompt, elapsed seconds tau are supplied through a hidden chronometric-injection channel, and the model selects a forced-choice temporal action. Here, "hidden" means hidden from the user-visible token sequence, not from model computation. Using Qwen2.5-3B-Instruct as a frozen bf16 backbone with a 31-dimensional sinusoidal-plus-log time encoding, gated FiLM residual modulation, and a rank-8 LoRA action surface, hidden-time CI reaches 0.9305 +/- 0.0134 accuracy and 0.9410 +/- 0.0103 balanced accuracy. No-time and shuffled-time controls fall to 0.5511 +/- 0.0042 and 0.3323 +/- 0.0097, respectively, with high shuffled-time wrong-state consistency supporting causal dependence on the injected scalar within the trained distribution. Generalization remains strong for held-out templates, durations, and multi-constraint compositions, but held-out quota-family transfer is weak at 0.5065 +/- 0.0559, while a fair prompt+LoRA timestamp baseline reaches 0.9893 +/- 0.0052. Thus, ChronoState supports a narrow conclusion: hidden elapsed time can be composed with symbolic task state under direct supervision, but does not establish autonomous time tracking, broad unseen-family abstraction, or superiority over prompt-injected timestamps.

Sam Siavoshian, Omar Ramadan, Amir K. Saeed +3

Independent Researcher, California, USA · Johns Hopkins University, Whiting School of Engineering, Baltimore, MD, USA

Chronos Frozen Language Model Temporal Reasoning Inference-Time Action Selection Confidence Intervals

Aug 9, 2026cs.CL

Deployable Per-Instance Multi-Layer Activation Steering for Large Language Models

Activation steering edits the behaviour of a frozen language model by adding a learned vector to its residual stream, and current practice fixes the injection layers globally per task. We argue that the best layers are an instance-level decision, and we make per-instance, multi-layer selection both well understood and deployable. On two open-weight 8B models and six binary persona traits, a per-instance oracle over layer subsets shows that the best layers vary from one input to the next: on most trait-model pairs, no fixed global layer set recovers the per-instance benefit. A greedy rule that ranks layers by single-layer marginal effect recovers nearly all of the oracle's benefit, but both must score candidate layers against the gold answer, so neither can run at deployment; the rule instead becomes the target a prompt-only predictor is trained to reproduce. Our deployable recipe needs no label at inference: a per-instance layer ranker read off the prompt embedding, a classifier that infers the steering direction, and an adaptive gate that scores short steered passes against that inferred direction and steers no more layers than necessary. The recipe recovers most of the oracle's lift (the bulk on the stronger model, a clear majority on the harder one), never drives any trait-model pair below its unsteered alignment baseline on average, and largely avoids the fluency collapse that strong global selection incurs at higher layer counts. A mechanistic account, "direction over magnitude", explains the behavioural flip under a mis-directed global set, the output collapse from steering too many layers, and the ceiling of unsteerable inputs.

Muhammad Faishal Adly Nelwan, Alfan Farizki Wicaksono

Faculty of Computer Science Universitas Indonesia

Linear Activation Steering Model Activations Frozen Language Model Steering Personality

Aug 5, 2026cs.LG

KV-Skill: Forging Expertise in the Model's Native Language

Task knowledge is commonly stored either as text in the prompt or as an update to model weights. Text is modular but must be interpreted on every use, while weight adaptation makes the resulting capability difficult to load, remove, or share independently. We introduce KV-Skill, a design space of external factorized operators that a frozen language model reads through a lightweight interface. KV-Skill supports two complementary paths. Registration converts an authored text skill into a text-derived operator and trains a shared per-backbone interface. Reward learning develops a compact latent operator directly from task outcomes, with or without an authored skill. Neither path adds positions to the prompt. Across ten benchmarks and four backbones from three model families, converting text to a KV-Skill consistently makes the same procedural knowledge more effective. On Qwen3.5-4B LiveMath, registration reaches 77.2 accuracy, compared with 23.4 for the source text skill, 52.0 for SkillOpt, and 64.5 for SoftSkill. Under matched reward training and parameter budgets, KV-Skill gives the best result in seven of eight matched settings against soft prefixes, prefix tuning, and LoRA. A post-hoc rank analysis further shows that text-derived operators retain nearly all of their benefit with one task-aligned direction per injection layer, while matched random directions fail. Finally, one shared interface retains three independently loadable KV-Skills without measurable forgetting. These results show that task knowledge can be acquired from text or experience, compressed into an external operator, and deployed separately from the backbone. Code is available at: https://github.com/shawnzhg/KV-Skill

Zhaowei Han, Xiang Zhang, Bing Han +3

University of Michigan, Ann Arbor, MI, USA

Qwen3 Frozen Language Model Logical Connectives

Jul 28, 2026cs.CL

IRIS: Reusable Identity Representations from Frozen LLMs for Entity Alignment

Entity alignment (EA) identifies entities across knowledge graphs (KGs) that refer to the same real-world object. Conventional EA methods mainly exploit explicit graph structures and textual fields, which often provide insufficient semantic understanding to recognize the same entity under heterogeneous descriptions and distinguish it from semantically similar entities. Although large language models (LLMs) offer deeper entity understanding, existing LLM-based EA methods largely use this capability for auxiliary generation or candidate-conditioned decisions. Consequently, such understanding is not distilled into a stable and directly comparable identity space, leaving alignment tied to specific KG pairs or candidate sets and requiring repeated processing as the matching context changes. To address these limitations, we propose IRIS (Identity Representations from Internal States), a training-free framework that constructs for each entity an iris-like signature encoding its distinctive and stable identity characteristics. IRIS derives these signatures by eliciting identity-oriented contextual representations from a frozen LLM, thereby forming a shared space in which each entity is encoded once and can be aligned across different KGs through direct similarity comparison, without pair-dependent representation construction or candidate-wise LLM inference. Across four established EA benchmarks and two frozen LLM backbones, the best IRIS variants achieve Hits@1 scores of 100.00, 99.38, 98.31, and 97.99 on D-Y-15K V2, DBP-WIKI, ICEWS-WIKI, and ICEWS-YAGO, respectively.

Xinran Liu, Shengtao Li, Shouqian Shi +2

Hierarchical Semantic Alignment Entities Knowledge Graphs Frozen Language Model LLM Inference Optimization

Jul 15, 2026cs.CL

Smarter and Cheaper at Once: Byte-Exact KV-Cache Grafting Turns a Frozen Small Model into a Verified-Knowledge Flywheel

We report a way to make a frozen small language model both more capable and dramatically cheaper at once, without changing any weights. Verified knowledge is deposited once as a byte-exact key-value (KV) state artifact and later restored, by graft, into a fresh inference context. The restore is bit-exact: under a pinned deterministic configuration, the grafted logits are byte-for-byte identical to a fresh computation (SHA-256 equality), with zero KL divergence and 100% argmax agreement over fifty samples. We show that own-position graft is the unique numerically exact operating point on a model with floating-point rotary encoding, and we verify byte-exactness on two model scales (12B, 31B) and two GPU targets, one through a pre-registered replay. On AIME 2025, a frozen Gemma-4-12B moves from 80.0% to 93.3% once a verified solution library is grafted, above its own 77.5% and its 31B sibling's 89.2% published anchors. On the recurring case, eight problems the base model never solves within a 401,026-token budget are answered from cached verified solutions in 61 total decode tokens, a factor of 6,574 fewer tokens and about 8,700x less energy; the capability claim proper rests on held-out transfer (7 of 7 at 31B). The same byte-exact store widens usable context from 32,768 to 2,854,766 tokens at zero extra accelerator memory, and moves byte-identical between machines of the same architecture. We describe the system at the behavior level; the engine is proprietary, and every reported number is backed by committed input and output hashes so the scoring can be re-checked without it.

Sietse Schelpe

Corbenic AI

Frozen Language Model Kv-Cache Management

Jul 14, 2026cs.SE

Form, Not Content? A Preregistered, Placebo-Controlled Evaluation of Learned Error-Conditioned Self-Repair Through Prompts and Weights in Frozen Small Code Models

Frozen small code LLMs are deployed locally, yet the information guiding a retry after a failed attempt is still measured without placebo controls in the self-repair literature. We treat a failed program as a conjecture and an execution counterexample as an oracle-relative refutation, and introduce PoPE (Popperian Placebo-controlled Evaluation): a methodology for measuring whether evidence that falsifies LLM-generated code can be used operationally by that same model. In PoPE, error content is paired with channel-specific placebos that keep the predeclared scaffold while ablating task-relevant content or deranging the task-error assignment. Frozen small code models (0.5-1.5B) are evaluated under preregistered rules through a prompt channel and a weight channel (small-data adapter training), with four generations per arm-unit pair. In the prompt channel, public-tier screening unlocked 12 units under the content-ablated form placebo versus 10 under the live error-pattern arm on a 40-unit resistant band; the result was recorded as mechanism-null. In the weight channel, an 8-8 tie was observed between the error-content adapter and the intervention-free baseline (p=1.0), while the SHA-deranged placebo adapter stayed ahead with 10 unlocks; content-attributable superiority was not confirmed. These results do not constitute evidence of equivalence or non-inferiority. Equivalence was not tested separately. Findings are restricted to the public-tier screening endpoint; hidden-tier confirmation was deferred by design. We read this not as compiled criticism disappearing as information, but as the loss of its external role in testing a new conjecture: when a representation learned from the oracle is written back into the generation state, testing is replaced by conditioning. No working JEPA-RL controller is claimed. PoPE is presented as a placebo-controlled, retestable measurement standard.

Mehmet Iscan

PythaLab, Yıldız Technical University, Istanbul, Turkey

Controlled Study Frozen Language Model Prompt Engineering Self-Repair

Jul 7, 2026cs.LG

A Quiet Failure in Calibrated Virtual Screening: Marginal Conformal Prediction Under-Covers the Minority Class, and a Class-Conditional Fix Recovers It

Conformal prediction is being adopted in drug discovery to put an honest number on model reliability: pick an error rate alpha, and the method returns prediction sets containing the true label with probability at least 1 - alpha. We show this guarantee can be dangerous on imbalanced datasets. Across four datasets, standard (marginal) conformal prediction hits its global 90% coverage target while leaving the minority class badly exposed: realized minority coverage falls to 64.8% on blood-brain-barrier penetration and to 4.2% on clinical-trial toxicity, where the rare class is nearly abandoned. The failure is not tied to one model: a random forest, a graph network, and a frozen chemical language model all reproduce it (p < 0.001 in every case), with severity tracking baseline calibration on rare labels rather than architecture. A conservation identity explains the effect: the minority's shortfall equals the majority's surplus amplified by the imbalance ratio, predicting the measured gap to within one point and ordering severity across datasets. The failure survives realistic scaffold splits and a second conformal score, while aggregate accuracy and overall coverage stay reassuringly high, which is exactly why it is easy to miss. Class-conditional (Mondrian) conformal prediction closes the gap on every dataset, restoring minority coverage to target for a modest increase in prediction-set size. We localize the failures to generic molecular scaffolds - plain benzene and pyridine cores occurring in both classes - propose a one-number diagnostic, and show with a cost model that abstaining on affected compounds flips a screening campaign from net-negative to net-positive utility. Our contribution is demonstrating on real chemistry how severe and invisible this known conformal-theory gap becomes under imbalance, and laying out a practical protocol restoring per-class reliability.

Muhammadjon Tursunbadalov, Mustafojon Tursunbadalov

School of Science and Technology, Champions College Prep, United States

Online Conformal Prediction Minority Class Imbalance Drug Discovery Moleculenet Frozen Language Model

Jul 2, 2026cs.LG

Safe Inference-Time Alignment via Lagrangian Reward Augmentation

Inference-time alignment steers a frozen language model during decoding using auxiliary reward signals, avoiding the cost of repeated weight updates. However, existing inference-time alignment methods typically optimize a single scalar score, so explicit safety constraints must either be ignored or encoded through manually tuned penalties. We propose Lagrangian Reward Augmentation (LARA), a general inference-time alignment framework under safety constraints. Starting from a KL-regularized constrained objective with a reward model and a cost model, LARA dualizes the constraint and reduces the optimization problem to a one-dimensional convex problem over a nonnegative dual variable. Estimated on a small calibration set, this dual variable defines an augmented reward that can be used as a drop-in scoring signal within existing inference-time alignment methods. For sequence-level sampling methods, such as Best-of-N reranking, the calibrated dual variable corresponds to the solution of the expected-cost constrained problem. For token-level reward-guided decoding methods, the same construction yields a principled dual-calibrated heuristic rather than an exact constrained-policy guarantee. We evaluate LARA on both sequence-level and token-level inference-time alignment methods, and find that LARA improves the helpfulness-harmlessness tradeoff, with Best-of-N achieving the best performance among inference-time methods, approaching finetuning-based direct alignment baselines.

Yaswanth Chittepu, Ativ Joshi, Sohini Chintala +1

University of Massachusetts Amherst · Independent Researcher

Reward Gradients Frozen Language Model Minimax Rate

Jul 1, 2026cs.LG

Black-Box Inference of LLM Architectural Properties with Restrictive API Access

In practice, most commercial LLM providers do not publicly release details of underlying LLM architectures. However, prior work has shown that given limited API access to an LLM (namely, top-

k

logits and/or a logit bias function), one can recover certain architectural details of an LLM, such as the hidden dimension of the feed-forward network. Perhaps in response to these results, most commercial LLM providers have restricted their APIs to expose only the single logit for each decoded token, and they no longer give users the ability to bias logits. We show that even under current restrictive APIs, several architectural parameters are still recoverable. We present NightVision, an attack that uses restrictive black-box API access to estimate the hidden dimension, depth, and parameter count of an LLM. Algorithmically, NightVision relies on a novel common set prompting technique in which multiple prompts expose log probabilities for the same set of output tokens; a spectral analysis of these results is used to infer hidden dimension. NightVision additionally uses end-to-end time to first token (TTFT) measurements and the estimated hidden dimension to estimate depth and parameter count. We empirically evaluate NightVision on 32 open-source LLMs, recovering hidden dimension to within 23% average relative error across all models (9% on MoE models), and depth and parameter count to within 53% for models exceeding three billion parameters. We run extensive ablations to demonstrate how these accuracies scale with token budget and model properties. Overall, our results suggest that current LLM APIs are not sufficiently restricted to fully obfuscate the architectural details of their underlying models.

Christopher Ellis, Shreyas Chaudhari, Mei-Yu Wang +3

Carnegie Mellon University Pittsburgh, PA 15213 · Pittsburgh Supercomputing Center

Attacker Large Language Model Large Language Model Tool Use Black Box Frozen Language Model Feed-Forward Access Control

Jun 27, 2026cs.CV

Mural: Transferring LLM knowledge to image generation via Mixture-of-Transformers

Leveraging capabilities of large language models (LLMs) in text-to-image (T2I) synthesis is an important research direction. In this work we investigate whether the knowledge of a frozen LLM can be effectively utilized in T2I generation when trained exclusively on standard text-image pairs. We integrate a frozen, reasoning-capable LLM with a diffusion-based image generator via shared attention within the Mixture-of-Transformers (MoT) architecture. Our experiments span two critical questions: (1) what degree of the LLM's intrinsic knowledge remains accessible during T2I training, and (2) what novel capabilities emerge in the resulting system. Across established benchmarks, our models achieve strong performance among unified understanding-generation systems: 0.85 on GenEval, 86.75 on DPG-Bench, and 0.66 on WISE with inference-time reasoning, using only text-image data. Remarkably, we uncover emergent behaviors absent from training data, including cross-lingual image generation, color-guided composition, emoji / ASCII scene construction, and generation directed by world knowledge. These results demonstrate that pretrained LLM knowledge can guide image synthesis under standard text-to-image training paradigms, without interleaved multimodal signals or explicit reasoning supervision. Our findings open new avenues for harnessing frozen model capabilities in resource-constrained multimodal learning.

Achin Jain, Jie An, Siddharth Chaudhary +1

Amazon AGI

Modern Text-To-Image Models Image Generation Multimodal Generation Frozen Language Model Transformer Architectures Synthesis+1

Jun 23, 2026cs.CL

What Intermediate Layers Know: Detecting Jailbreaks from Entropy Dynamics

Jailbreak attacks reveal a persistent weakness in aligned Large Language Models: carefully crafted prompts can elicit policy-violating responses despite safety training. While most defenses operate at the prompt or output level, it remains unclear how harmful intent is encoded within the model's internal representations. We investigate this question by analyzing token-level predictive entropy trajectories across layers of a frozen LLM using the logit lens. We find that static aggregate statistics of prompt-level entropy (e.g., mean, variance) carry little discriminative signal, whereas features capturing how entropy evolves across token positions, such as monotonic rank-based trend scores, are substantially more informative. Importantly, this signal is not uniform across model depth: it is concentrated in intermediate layers and degrades at the final layer, indicating that jailbreak-relevant structure is most pronounced in mid-network representations rather than at the output head. Across multiple models (Llama, Qwen, Gemma) and adversarial benchmarks, these entropy dynamics provide architecture-consistent separation without additional training. Together, our findings show that jailbreak behavior is reflected in structured intermediate uncertainty dynamics, clarifying both which entropy-derived features encode harmful intent and where in the network that signal is most pronounced.

Sofiia Nikolenko, Michele Papucci, Mina Rezaei +1

LMU Munich, Munich, Germany · relAI – Konrad Zuse School of Excellence in Reliable AI · University of Pisa, Pisa, Italy +2

Large Language Model Jailbreaks Token-Level Entropy Jailbreak Success Rates Jailbreaks Frozen Language Model Adversarial Evaluation

Jun 18, 2026cs.AI

SoftSkill: Behavioral Compression for Contextual Adaptation

Natural-language skills let agents reuse task knowledge, yet deploying a long Markdown document makes the model interpret that knowledge anew on every call. We ask whether the behavior induced by a skill can be carried by a compact, trainable context. SoftSkill initializes virtual token embeddings from a skill document and optimizes a soft skill with next-token prediction while keeping the language model frozen. The resulting conditioning sequence can occupy the skill section or another supported prompt location. On Qwen3.5-4B, a 32-token soft skill improves over no skill by 7.6, 42.1, and 1.3 points on SearchQA, LiveMath, and DocVQA. It also exceeds the optimized textual skill on SearchQA and LiveMath by 4.5 and 12.5 points, respectively, while replacing skill documents of hundreds to thousands of tokens. The same approach improves multi-step agent execution: on Qwen3.6-35B-A3B, OfficeQA and ALFWorld rise by 8.2 and 14.1 points over no skill; on Qwen3.5-4B with aligned action decoding, ALFWorld success rises from 44/134 with the untrained initialization to 91/134 after training. These findings show that a skill document can serve as the starting point for a continuous control that improves both answers and actions without updating the backbone.

Xijia Tao, Yihua Teng, Xinyu Fu +6

The University of Hong Kong · Huawei Research · City University of Hong Kong +1

Frozen Language Model Natural Language Prior Knowledge Integration Data Compression Methods

Jun 16, 2026cs.LG

From Sparse Features to Trustworthy Proxies: Certifying SAE-Based Interpretability

Sparse autoencoders (SAEs) are increasingly used to extract interpretable features from language models (LMs), yet a central question remains: when can an SAE-based explanation be treated as a faithful view of an underlying frozen LM We study this through a post-hoc generalization framework that certifies the LM via a sparse proxy, obtained by replacing a native hidden activation with its pretrained SAE reconstruction. Our framework derives an upper bound on the base model's expected risk using four measurable quantities: proxy risk, SAE reconstruction gap, concept-pool mismatch, and sparse complexity. We interpret this certificate as an operational criterion for explanatory faithfulness. In particular, a non-vacuous bound indicates that the extracted sparse features retain meaningful predictive information, while small reconstruction and mismatch errors indicate that the proxy remains behaviorally close to the original model. Empirically, we show that the bound becomes non-vacuous on GPT-2 Small, Gemma-2B, and Llama-3-8B at practical sample sizes. A detailed layerwise analysis of Llama-3-8B reveals a strong depth dependence, with later layers becoming much easier to certify, associated with both stronger local fidelity and weaker downstream error amplification. Finally, through feature-shuffling ablations, we show that the decomposition distinguishes genuine semantic alignment from mere statistical sparsity, providing a useful diagnostic for when SAE-based explanations become less reliable.

Dibyanayan Bandyopadhyay, Asif Ekbal

Department of Computer Science and Engineering, Indian Institute of Technology Patna, Bihar, India.

Interpretability Model Activations Explainable AI Methods Frozen Language Model Certification

Jun 5, 2026cs.CL

SV-Detect: AI-generated Text Detection with Steering Vectors

Detecting AI-generated text is especially difficult under distribution shift, such as transfer across domains, source models, and editing attacks. We propose an AI-generated text detector based on steering vectors extracted from the hidden representations of a frozen language model. At each layer, we construct a direction that separates human-written from AI-generated text, and represent each input by its layer-wise alignment with these directions. A lightweight classifier trained on these projection features yields the final detection score. Our method achieves strong performance both in-distribution and under distribution shift, including across domains, source models, and machine-editing transformations such as polishing and rewriting. Interpretation analyses show that the learned directions align with recognizable stylistic cues while capturing substantial additional signal beyond surface features. These results position AI-generated text detection as a representation-space probing problem and show that steering vectors provide a simple and effective solution.

Mikhail Vishnyakov, Tatiana Gaintseva

Independent Researcher · Queen Mary University of London

Machine-Generated Text Detection Frozen Language Model Steering

Jun 4, 2026cs.CL

ReverseEOL: Improving Training-free Text Embeddings via Text Reversal in Decoder-only LLMs

Recent advances in Large Language Models (LLMs) have opened new avenues for generating training-free text embeddings. However, the causal attention in decoder-only LLMs prevents earlier tokens from attending to future context, leading to biased contextualized representations. In this work, we propose Reverse prompting with Explicit One-word Limitation (ReverseEOL), a simple yet effective method for enhancing the representational capability of frozen LLMs. ReverseEOL augments the standard forward embedding with an additional reversed embedding derived from the reversed input text. Since reversing the input exposes each token to context inaccessible in the original order, the resulting reversed embedding effectively provides complementary information to the original one. As a result, combining the forward and reversed embeddings yields a richer final representation. Comprehensive experiments on STS and MTEB benchmarks demonstrate that ReverseEOL significantly improves the performance of existing training-free baselines across a broad range of LLMs with diverse architectures and scales. Extensive ablations and analyses further confirm the necessity of our reversal mechanism.

Ailiang Lin, Zhuoyun Li, Yusong Wang +3

Institute of Science Tokyo · Tencent

Frozen Language Model Reverse

Jun 2, 2026cs.CL

When Does Complexity Conditioning Help a Frozen Sentence Embedding? A Controlled Study of Per-Sentence and Pair-Level Difficulty Adaptation

A common intuition is that sentence embeddings should adapt to the difficulty of the input. We test this intuition in a controlled, multi-seed setting: a lightweight post-encoder adapter attaches to a frozen Qwen3-Embedding-0.6B encoder, accessing only its final pooled embedding, and is evaluated on four paraphrase and semantic-similarity tasks (PAWS, MRPC, QQP, STS-B). The naive form of the idea fails: surface-based per-sentence complexity is nearly uncorrelated with frozen-baseline error (Pearson approximately 0.05) and provides no advantage over constant or shuffled controls, while degrading a saturated baseline. Even when the target is aligned to a non-circular pair-difficulty signal, the per-sentence gate still cannot reliably capture difficulty because difficulty is primarily a property of the pair, not the individual sentence. In contrast, a small pair-level residual gated by a held-out cross-encoder difficulty signal yields consistent gains on the larger and graded tasks, including +0.022 Spearman on STS-B and +0.037 on QQP, while remaining anchored to the frozen baseline across all seeds. Because this useful form operates on sentence pairs rather than individual sentences, the resulting model is best understood as a lightweight re-ranker over cached frozen embeddings, not a replacement single-vector embedding; we make no state-of-the-art claim. Our contribution is a controlled account of when difficulty-aware adaptation helps and when it fails, together with a pre-training diagnostic that predicts the available headroom.

Suhwan Hwang

Cross-Encoder Frozen Language Model

May 29, 2026cs.CL

Point-in-Time Financial RAG with Frozen LLMs and Market-Feedback Adaptive Retrieval

Financial retrieval-augmented generation (RAG) systems typically rank evidence by textual relevance, but in financial markets evidence utility depends on event type, forecast horizon, and market context. We study news-triggered event-impact prediction as a point-in-time financial RAG problem. For each company-news anchor, the system retrieves financial news and SEC filing passages, appends a pre-decision market-context card, and predicts multi-horizon residual-return signals. Our method keeps the LLM frozen and adapts retrieval through an external Bayesian source memory updated from matured residual-return feedback. On a fixed 89-stock Nasdaq-oriented universe derived from the FinRL-DeepSeek/FNSPID task, using original FNSPID news and point-in-time EDGAR filing passages, Frozen Reader with Source Memory improves held-out macro-F1 from 0.438 to 0.471 and downstream portfolio Sharpe from 0.52 to 0.84 relative to Frozen Reader with No Memory. Supervised LoRA gives modest gains under static retrieval, but after source-memory adaptation, the LoRA reader does not improve over the frozen reader. These results suggest that, for financial RAG systems, learning where to retrieve can be as important as learning how to read, offering a modular route to market-feedback adaptation.

Zijie Zhao, Roy E. Welsch

Massachusetts Institute of Technology · Cambridge, MA, USA

Financial Question Answering Frozen Language Model Deepseek

May 28, 2026cs.CL

COFT: Counterfactual-Conformal Decoding for Fair Chain-of-Thought Reasoning in Large Language Models

Large language models (LLMs) can reveal and amplify societal biases during chain-of-thought (CoT) generation. We present COFT (Chain of Fair Thought), a training-free decoding method that applies token-level fairness control at decode time, with distribution-free marginal validity guarantees (under exchangeability) for any frozen causal language model. COFT operates in three stages. First, it creates a masked counterfactual prompt by replacing sensitive spans with neutral tokens. Second, it compares the factual and masked logit distributions through lightweight logit fusion to attenuate attribute-driven biases. Third, it uses dual-branch split-conformal calibration to certify per-step candidate token sets at a user-chosen risk level. We evaluate COFT across six models and multiple bias benchmarks. Our method reduces standard bias metrics by 30-55% (median 38%) while preserving task utility and language quality. Reasoning accuracies remain unchanged within run-to-run noise margins. The computational overhead is modest, equivalent to one additional cached forward pass (<=11%). COFT offers a clear, auditable path to safer CoT generation with significant bias reduction, negligible utility loss, and no requirement for retraining, auxiliary classifiers, or weight access.

Arya Fayyazi, Mehdi Kamal, Massoud Pedram

Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, California, USA.

Large Language Model Bias Frozen Language Model Chain-of-Thought Reasoning Biases Algorithmic Fairness

May 27, 2026cs.LG

SYNAPSE: Neuro-Symbolic Visual Thought-to-Text Decoding via Topological Semantic Denoising

Recent advances in large language models have accelerated open-vocabulary EEG-to-imagined-text decoding, where non-invasive neural activity recorded during visual perception is translated into coherent natural language descriptions of viewed stimuli. However, existing systems remain highly vulnerable to biological noise, where corrupted neural projections induce hallucinated or semantically unstable generation in frozen language models. We introduce SYNAPSE (Symbolic Neural Alignment for Precise Semantic Extraction), a lightweight neuro-symbolic framework that stabilizes neural text generation through inference-time symbolic regularization. By purifying EEG-derived semantic candidates using commonsense graph structure and latent exemplars, SYNAPSE improves semantic stability without end-to-end LLM fine-tuning. Experiments across popular EEG decoding benchmarks and multiple frozen LLM backends demonstrate consistent gains over unconstrained prompting baselines, robustness under object-label ablation, and performance commensurate with substantially more resource-intensive fine-tuned systems, while preserving biometric privacy by localizing raw EEG processing entirely within the encoder stack.

Akshaj Murhekar, Abhijit Mishra

School of Information University of Texas at Austin

Electroencephalography Decoding Neuro-Symbolic Framework Frozen Language Model Iterative Denoising Methods

May 26, 2026cs.AI

Why LLMs Fail at Causal Discovery and How Interventional Agents Escape

Causal discovery is a cornerstone of scientific reasoning, yet whether large language models can perform it reliably remains an open question. Recent benchmarks show that even fine-tuned models plateau on simple causal graphs and degrade as complexity grows, but why they fail has not been established. We prove the failure is fundamental: supervised fine-tuning, direct preference optimization, and in-context learning all produce predictors that cannot distinguish between causal graphs generating similar observational data, and any attempt to do so requires the model's internal representations to grow unboundedly, violating the very conditions under which these methods work. We formalize this as a kernel obstruction theorem, establishing that the limitation is intrinsic to the learning paradigm, \emph{not any particular model or dataset}. We propose Agentic Causal Bayesian Optimization (A-CBO), wherein a frozen language model serves as an interventional oracle answering targeted queries about intervention effects, while an external Bayesian loop concentrates beliefs over candidate graphs in logarithmically many rounds. Because the decision operates outside the space where the obstruction applies, A-CBO provably converges while the underlying model remains unchanged. On Corr2Cause, A-CBO matches fine-tuned baselines without any training. On Extended Corr2Cause, a new benchmark scaling to 24 variables with 18K test samples, A-CBO significantly outperforms both fine-tuning and preference optimization, with the advantage growing

Amartya Roy, Sonali Parbhoo

SIRE, IIT Delhi and Robert Bosch GmbH, India · Imperial College London

Large Language Models Fail LLM Reasoning Strategies Frozen Language Model Model Fine-Tuning Multi-Objective Bayesian Optimization Algorithms

May 26, 2026cs.CL

LATTE: Forecasting Peer Anchored Preference Trajectories for Personalized LLM Generation

Personalized generation with frozen large language models requires a conditioning signal that is both compact and current. Existing personalization methods typically retrieve or summarize user histories in text, or compress them into static latent profiles and soft prompts. These approaches are efficient, but they treat a user's past behavior as an aggregate profile and therefore mix stable identity, recent drift, and item content in the same representation. We propose LAtent Trajectory Tracking and Extrapolation (LATTE), a framework that represents personalization as forecasting a peer anchored relative preference state. For each historical session, LATTE subtracts a time masked baseline formed from comparable users who responded to the same item, producing a state that measures how the target user differs from peers under a shared item context. A lightweight sequence predictor then forecasts the next state in this trajectory, and a State to Token Bridge injects the forecast into a frozen instruction tuned LLM through a single anchored soft token. We provide a latent factor analysis showing when peer anchoring cancels shared item variation and why temporal forecasting trades off stale averages against noisy recent states. Experiments on Amazon Reviews 2023 and MemoryCD show that LATTE consistently outperforms retrieval, summary memory, static latent profiles, difference aware latent profiles, and soft prompt compression baselines. On Amazon Reviews 2023, LATTE improves average ROUGE-L from 0.219 for a static latent profile and 0.245 for the strongest added latent compression baseline to 0.259. Additional pairwise comparisons and diagnostic analyses suggest that the improvement is mainly due to forecasting user-specific trajectory information, rather than merely adding a soft prompt interface.

Jinze Li, Xiaoyan Yang, Shuo Yang +5

The University of Hong Kong · Ant Healthcare, Ant Group

Large Language Model Personalization Frozen Language Model Large Language Model Generation Latent Variable Peers

May 22, 2026cs.LG

PaP-NF: Probabilistic Long-Term Time Series Forecasting via Prefix-as-Prompt Reprogramming and Normalizing Flows

Time series forecasting plays a central role in many real-world applications and has been extensively studied. Most existing approaches rely on deterministic models. However, real-world environments exhibit inherently uncertain and complex future behaviors, making single-point predictions insufficient. This highlights the need for probabilistic forecasting methods that can quantify and represent uncertainty. In this work, we propose PaP-NF, a probabilistic forecasting framework that aligns continuous time series representations with a frozen large language model (LLM) using a Prefix-as-Prompt mechanism, and conditions a normalizing flow decoder on the global context extracted by the LLM. The quality of the resulting predictive distributions is evaluated using the Continuous Ranked Probability Score (CRPS), a standard metric in probabilistic forecasting. Across a variety of long-term forecasting benchmarks, PaP-NF robustly captures multi-modal uncertainty while maintaining competitive point forecasting accuracy. The official implementation is available at: https://github.com/democracy04/PaP-NF

Minju Kim, Youngbum Hur

Department of Industrial Engineering, Inha University, Incheon, Republic of Korea

Probabilistic Forecasting Time Series Frozen Language Model

May 21, 2026cs.LG

PACE: Two-Timescale Self-Evolution for Small Language Model Agents

Deploying language-model agents in production often requires substantial compute and human effort to tune prompts, parsers, validators, and other components of the agent pipeline. Self-evolution offers a promising alternative, but most existing frameworks assume access to frontier models that can reliably diagnose failures, propose revisions, and judge their own updates. We study whether frozen small language models (SLMs) can serve as effective self-evolving agents under resource constraints. We propose PACE (Prompt And Control Logic Evolution), a two-timescale framework that coordinates low-risk prompt refinement with higher-risk control-logic updates. PACE evolves prompts under fixed control logic until prompt-level gains saturate, then considers constrained control-logic updates that are accepted through held-out validation. Across three frozen SLM backbones ranging from 4B to 14B parameters and four controlled benchmarks, PACE achieves the best performance on all 12 backbone--benchmark combinations, improving over vanilla SLM agents by up to +9.2% relative improvement and over the stronger single-mode evolution baseline by up to +5.4% relative improvement. A tau-bench case study further shows that PACE improves multi-turn tool-use success over vanilla and prompt-only evolution. These results suggest that reliable SLM agent self-evolution is possible without updating model weights or relying on frontier-model teachers, and that the key benefit is not any single final solver pattern but autonomous, validated discovery of task-appropriate inference strategies.

Chen Ling, Pei Chen, Albert Guan +4

Amazon

Self-Evolution Large Language Model Agents Frozen Language Model Language-Model Agents Timescale

May 21, 2026cs.AI

Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents

LLM agents are shaped not only by their language models, but also by the runtime harness that mediates observation, tool use, action execution, feedback interpretation, and trajectory control. While existing agent adaptation methods mainly update model parameters, many failures in deterministic, rule-governed domains stem from mismatches at the model--environment interface. We propose Life-Harness, a lifecycle-aware runtime harness that improves frozen LLM agents without changing model weights or evaluation environments. Life-Harness evolves from training trajectories by converting recurring interaction failures into reusable interventions across environment contracts, procedural skills, action realization, and trajectory regulation, and remains fixed for evaluation on unseen tasks. On seven deterministic environments from

τ

-bench,

τ^2

-bench, and AgentBench, Life-Harness improves 116 out of 126 model--environment settings across 18 model backbones, with an average relative improvement of 88.5%. Harnesses evolved only from Qwen3-4B-Instruct trajectories transfer to 17 other models, showing that Life-Harness captures reusable environment-side structure rather than model-specific behavior. These results position runtime interface adaptation as a complementary alternative to model-centric agent training. Code is available at https://github.com/Tianshi-Xu/Life-Harness.

Tianshi Xu, Huifeng Wen, Meng Li

Peking University

Agent Harness Large Language Model Agents Frozen Language Model

May 20, 2026cs.LG

EvoStruct: Bridging Evolutionary and Structural Priors for Antibody CDR Design via Protein Language Model Adaptation

Equivariant graph neural network (GNN) methods for antibody complementarity-determining region (CDR) design achieve the highest sequence recovery but suffer from severe vocabulary collapse. The current best GNN methods over-predict very few amino acids, such as tyrosine and glycine, while ignoring functionally important residues. We trace this failure to GNN encoders learning amino acid distributions de novo from limited structural data, discarding substitution patterns encoded in evolutionary databases. To resolve this, we propose EvoStruct, which bridges a frozen protein language model (PLM) with 3D structural context from an E(3)-equivariant GNN via a cross-attention adapter. Unlike prior PLM-structure adapters for general protein design, EvoStruct targets the vocabulary collapse problem specific to CDR design through progressive PLM unfreezing and R-Drop consistency regularization. On the CHIMERA-Bench dataset, EvoStruct achieves the highest amino acid recovery and lowest perplexity among several antibody design methods, improving sequence recovery by 16% and reducing perplexity by 43% relative to the best GNN baselines, while recovering 2.3x greater amino acid diversity and the highest binding-pair correlation with ground truth.

Mansoor Ahmed, Sujin Lee, Umar Khayaz +1

Georgia State University, Atlanta, USA · Georgia Institute of Technology, Atlanta, USA

Antibody Epitope Prediction Alphaevolve Graph Neural Networks Large Language Model Adaptation Frozen Language Model

May 19, 2026cs.LG

LLM Benchmark Datasets Should Be Contamination-Resistant

Benchmark datasets are critical for reproducible, reliable, and discriminative evaluation of LLMs. However, recent studies reveal that many benchmark datasets are included in pretraining corpora, i.e.,

\textit{contaminated}

, which diminishes their value as reliable measures of model generalization. In this paper, we argue that benchmark datasets should be

\textit{contamination-resistant}

, i.e.,

\textit{unlearnable}

, but support

\textit{inference}

. To accomplish this, we first highlight the wide prevalence of benchmark dataset contamination and outline the properties of contamination-resistant datasets. Second, we highlight how the asymmetry between the inference and training pipelines in the Transformer architecture can be leveraged to support contamination-resistance. Third, we outline mathematical advancements to make these datasets interoperable across various LLM architectures. Based on the above, we call on the community to ensure the reliability of LLM benchmarking by: (i) advancing novel contamination-resistant methodologies, (ii) developing supporting methods and platforms, and (iii) adopting contamination-resistant benchmarks into existing evaluation pipelines.

Ali Al-Lawati, Jason Lucas, Dongwon Lee +1

The Pennsylvania State University, University Park, PA, USA.

Benchmark Datasets Large Language Model Benchmarks Model Training Frozen Language Model Discriminative

May 11, 2026cs.CL

The Bicameral Model: Bidirectional Hidden-State Coupling Between Parallel Language Models

Existing multi-model and tool-augmented systems communicate by generating text, serializing every exchange through the output vocabulary. Can two pretrained language models instead coordinate through a continuous, concurrent channel? The Bicameral Model couples two frozen language models through a trainable neural interface on their intermediate hidden states. At every generation step, both models run in lockstep: a primary model drives the task while an auxiliary model operates tools, solves constraints, or executes code, with both conditioning on each other's activations through a translation network and a learned suppression gate (

\sim

1% of combined parameters). The gate learns a selective communication protocol from task loss alone, without a prescribed format. We demonstrate the mechanism across three tool backends. On arithmetic, coupling two 0.5B models with a calculator raises accuracy from 36% to 96%. On logic grid puzzles, coupling two 0.6B models with a Z3 solver achieves

1.7\times

the unaugmented baseline on ZebraLogic. On mathematical reasoning, coupling with a Python sandbox enables the auxiliary to generate problem-specific code from hidden-state signals alone, without ever seeing the problem text.

Cedric Flamant, Udaya Ghai, Kanna Shimizu

AWS Agentic AI

Frozen Language Model

May 8, 2026cs.CL

PYTHALAB-MERA: Validation-Grounded Memory, Retrieval, and Acceptance Control for Frozen-LLM Coding Agents

Local LLM-based coding agents increasingly work in settings where correctness is earned through execution feedback, persistent state, and bounded repair, not through a single fluent answer. Static retrieval, long-context prompting, self-refinement, execution-feedback repair, and reinforcement learning over model weights each address part of this setting, but they do not jointly provide validation-grounded episodic memory, adaptive retrieval-action selection, delayed credit assignment, and structural skill reuse around a frozen local model. We introduce PYTHALAB-MERA, a lightweight external controller for local validation-conditioned code generation. The frozen language model proposes complete source files; the controller decides which memory records and AST-derived skills should enter the next prompt, validates each candidate through a fail-fast pipeline, converts validation outcomes into bounded shaped rewards, and propagates delayed credit through TD(lambda)-style eligibility traces. We evaluate the implementation as a local CLI artifact on reinforcement-learning coding tasks with strict validation gates. In the measured hard RL setting with three tasks, three repetitions, and a three-attempt budget, PYTHALAB-MERA passed 8/9 strict validations; the self-refinement baseline and the investigated GRACE extension each passed 0/9. These results support a deliberately bounded claim: in this recorded setting, the external memory-and-retrieval controller improved validation success. They do not establish general-purpose code synthesis, state-of-the-art performance, formal program correctness, or formal safety.

Mehmet Iscan

PythaLab, Yildiz Technical University

Code Generation Frozen Language Model Episodic Memory Acceptance

May 4, 2026cs.AI

Stable Agentic Control: Tool-Mediated LLM Architecture for Autonomous Cyber Defense

Agentic systems involved in high-stake decision-making under adversarial pressure need formal guarantees not offered by existing approaches. Motivated by the operational needs of security operations centers (SOCs) that must configure endpoint detection and response (EDR) policies under adversarial pressure, we present a tool-mediated architecture: LLM agents use deterministic tools (Stackelberg best-response, Bayesian observer updates, attack-graph primitives) and select from finite action catalogs enforced at the tool-output interface. A composite Lyapunov function machine-checked in Lean 4 with zero sorry certifies controllability, observability from asymmetric sensor data, and Input-to-State Stability (ISS) robustness under intelligent adversarial disturbance, with two corollaries extending the certificate to any controller or adversary from the catalogs. On 282 real enterprise attack graphs, the claims hold with margin. On paired offensive/defensive telemetry, a tool-mediated Claude Sonnet 4 controller reduces the attacker's expected payoff (game value) by 59% relative to a deterministic greedy baseline, with zero variance across 40 runs at four temperatures. A Claude Haiku 4.5 controller converges to suboptimal game values but stays catalog-bounded over an additional 40 runs, demonstrating that architectural stability is not dependent on the controller capability. The LLM agent's non-determinism furthers creative exploration of strategies, while the tool-mediated architecture ensures system stability.

Kerri Prinos, Lilianne Brush, Cameron Denton +5

Horizon3.ai San Francisco, CA

Agentic Control Large Language Model Agents Multivariate Telemetry Data Frozen Language Model

May 1, 2026cs.LG

Borrowed Geometry: Cross-Distribution Head-Importance Fingerprints of Frozen Pretrained Gemma 4 31B

Frozen Gemma 4 31B weights pretrained exclusively on text, unmodified, transfer through a thin trainable interface to non-text modalities the substrate has never processed. On the L24--L29 slice (192 attention heads), an English-text TxtCopy attention probe (95 sentences) and per-head ablation impact on four non-language token-pattern tasks (binary copy, associative recall, 1D cellular automaton Rule 90, binary addition) jointly classify four heads -- L26.28, L27.28, L27.2, L27.3 -- as top-tier on both signals. The slice-level joint coincidence is significant under hypergeometric null (

P = 0.0013

,

N=192

,

K=38

,

n=4

) and survives multiplicity-aware permutation tests (

P_{V4} = 0.013

). Pretrained Gemma L26 reaches 60.22% on OGBench cube-double-play-task1 vs ~1% for random-init Gemma (

+59

pt at

n=3

); a FrozenRandom-GPT2 control with correct

1/\sqrt{d_k}

scaling also fails. Head-level causal validation: zeroing L26.28 in the trained cube-task1 IQL agent drops success

63.3\% \to 10.0\%

vs

46.7\%

for a layer-matched low-TxtCopy negative control (

3.2\times

specificity at

n=30

;

n=5

paired-

t

p=0.039

). A full L26 sweep places L26.28 at rank 4 of 32. Honest negatives: within-L26 Spearman

ρ(\text{TxtCopy, drop}) = +0.37

(opposite of within-layer causal reading); single-head activation patching does not transfer the matching variable; the 4 named heads alone do not suffice on any task; Walker2d-DT and scene-task1 recruit L24 outside the named slice and show null head-ablation specificity. We frame the contribution as a cross-distribution importance fingerprint at the slice level plus head-level causal evidence on one cross-modality target.

Abay Bektursun

Independent research

Frozen Language Model Ablation Cross-Modality Fingerprint

Apr 28, 2026cs.CL

Investigating Linguistic Steering: An Analysis of Adjectival Effects Across Large Language Model Architectures

Achieving reliable control of Large Language Models (LLMs) requires a precise, scalable understanding of how they interpret linguistic cues. We introduce a rigorous framework using Shapley values to quantify the steering effect of individual adjectives on model performance, moving beyond anecdotal heuristics to principled attribution. Applying this method to 100 adjectives across a diverse suite of models (including o3, gpt-4o-mini, phi-3, llama-3-70b, and deepseek-r1) on the MMLU benchmark, we uncover several critical findings for AI alignment. First, we find that a small subset of adjectives act as disproportionately powerful "levers," yet their effects are not universal. Cross-model analysis reveals a "family effect": models of a shared lineage exhibit correlated sensitivity profiles, while architecturally distinct models react in a largely uncorrelated manner, challenging the notion of a one-size-fits-all prompting strategy. Second, focused follow-up studies demonstrate that the steering direction of these powerful adjectives is not intrinsic but is highly contingent on their syntactic role and position within the prompt. For larger models like gpt-4o-mini, we provide the first quantitative evidence of strong, non-additive interaction effects where adjectives can synergistically amplify, antagonistically dampen, or even reverse each other's impact. In contrast, smaller models like phi-3 exhibit a more literal and less compositional response. These results suggest that as models scale, their interpretation of prompts becomes more sophisticated but also less predictable, posing a significant challenge for robustly steering model behavior and highlighting the need for compositional and model-specific alignment techniques.

Lars Malmqvist

Research and Implementation

Linguistics Frozen Language Model Artificial Intelligence Alignment Steering

Apr 24, 2026cs.CL

Learning Evidence Highlighting for Frozen LLMs

Large Language Models (LLMs) can reason well, yet often miss decisive evidence when it is buried in long, noisy contexts. We introduce HiLight, an Evidence Emphasis framework that decouples evidence selection from reasoning for frozen LLM solvers. HiLight avoids compressing or rewriting the input, which can discard or distort evidence, by training a lightweight Emphasis Actor to insert minimal highlight tags around pivotal spans in the unaltered context. A frozen Solver then performs downstream reasoning on the emphasized input. We cast highlighting as a weakly supervised decision-making problem and optimize the Actor with reinforcement learning using only the Solver's task reward, requiring no evidence labels and no access to or modification of the Solver. Across sequential recommendation and long-context question answering, HiLight consistently improves performance over strong prompt-based and automated prompt-optimization baselines. The learned emphasis policy transfers zero-shot to both smaller and larger unseen Solver families, including an API-based Solver, suggesting that the Actor captures genuine, reusable evidence structure rather than overfitting to a single backbone.

Shaoang Li, Yanhang Shi, Yufei Li +10

Stony Brook University · Meta AI

LLM Reasoning Strategies Frozen Language Model Evidence-Grounded Questions Downstream Reasoning

Apr 23, 2026cs.CV

Frozen LLMs as Map-Aware Spatio-Temporal Reasoners for Vehicle Trajectory Prediction

Large language models (LLMs) have recently demonstrated strong reasoning capabilities and attracted increasing research attention in the field of autonomous driving (AD). However, safe application of LLMs on AD perception and prediction still requires a thorough understanding of both the dynamic traffic agents and the static road infrastructure. To this end, this study introduces a framework to evaluate the capability of LLMs in understanding the behaviors of dynamic traffic agents and the topology of road networks. The framework leverages frozen LLMs as the reasoning engine, employing a traffic encoder to extract spatial-level scene features from observed trajectories of agents, while a lightweight Convolutional Neural Network (CNN) encodes the local high-definition (HD) maps. To assess the intrinsic reasoning ability of LLMs, the extracted scene features are then transformed into LLM-compatible tokens via a reprogramming adapter. By residing the prediction burden with the LLMs, a simpler linear decoder is applied to output future trajectories. The framework enables a quantitative analysis of the influence of multi-modal information, especially the impact of map semantics on trajectory prediction accuracy, and allows seamless integration of frozen LLMs with minimal adaptation, thereby demonstrating strong generalizability across diverse LLM architectures and providing a unified platform for model evaluation.

Yanjiao Liu, Jiawei Liu, Xun Gong +1

School of Artificial Intelligence, Jilin University, Changchun 130012, PR China.

Autonomous Driving Trajectory Forecasting Traffic Spatio-Temporal Reasoning Frozen Language Model Semantic Mapping

Apr 22, 2026cs.CL

LayerTracer: A Joint Task-Particle and Vulnerable-Layer Analysis framework for Arbitrary Large Language Model Architectures

Currently, Large Language Models (LLMs) feature a diversified architectural landscape, including traditional Transformer, GateDeltaNet, and Mamba. However, the evolutionary laws of hierarchical representations, task knowledge formation positions, and network robustness bottleneck mechanisms in various LLM architectures remain unclear, posing core challenges for hybrid architecture design and model optimization. This paper proposes LayerTracer, an architecture-agnostic end-to-end analysis framework compatible with any LLM architecture. By extracting hidden states layer-by-layer and mapping them to vocabulary probability distributions, it achieves joint analysis of task particle localization and layer vulnerability quantification. We define the task particle as the key layer where the target token probability first rises significantly, representing the model's task execution starting point, and the vulnerable layer is defined as the layer with the maximum Jensen-Shannon (JS) divergence between output distributions before and after mask perturbation, reflecting its sensitivity to disturbances. Experiments on models of different parameter scales show that task particles mainly appear in the deep layers of the model regardless of parameter size, while larger-parameter models exhibit stronger hierarchical robustness. LayerTracer provides a scientific basis for layer division, module ratio, and gating switching of hybrid architectures, effectively optimizing model performance. It accurately locates task-effective layers and stability bottlenecks, offering universal support for LLM structure design and interpretability research.

Yuhang Wu, Qinyuan Liu, Qiuyang Zhao +1

China Electronic Technology Nanhu Research Institute, Jiaxing, China

Layer-Wise Frozen Language Model Traces Joint Search

Apr 20, 2026cs.CL

FreezeEmpath: Efficient Training for Empathetic Spoken Chatbots with Frozen LLMs

Empathy is essential for fostering natural interactions in spoken dialogue systems, as it enables machines to recognize the emotional tone of human speech and deliver empathetic responses. Recent research has made significant progress in developing empathetic spoken chatbots based on large language models (LLMs). However, several challenges still exist when training such models, including reliance on costly empathetic speech instruction data and a lack of emotional expressiveness in the generated speech. Finetuning LLM with cross-modal empathetic instruction data may also lead to catastrophic forgetting and a degradation of its general capability. To address these challenges, we propose FreezeEmpath, an end-to-end empathetic spoken chatbot trained in a simple and efficient manner. The entire training process relies solely on existing speech instruction data and speech emotion recognition (SER) data, while keeping the LLM's parameters frozen. Experiments demonstrate that FreezeEmpath is able to generate emotionally expressive speech and outperforms other empathetic models in empathetic dialogue, SER, and SpokenQA tasks, demonstrating the effectiveness of our training strategy.

Yun Hong, Yan Zhou, Yang Feng

Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS) · State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences, Beijing, China

Speech Language Models Empathy Emotional Support Conversation Task Frozen Language Model Chatbots

Dec 26, 2025cs.LG

DuaDeep-SeqAffinity: Dual-Branch Deep Learning for Tri-Stream Sequence-Based Antibody--Antigen Affinity Prediction

DuaDeep-SeqAffinity is a sequence-only deep learning framework that predicts antibody--antigen binding affinity directly from primary amino acid sequences, avoiding the cost and scarcity of resolved three-dimensional structures. The antigen and the antibody heavy and light chains are processed as three independent streams, each embedded with a frozen ESM-2 protein language model and passed through parallel Transformer and convolutional neural network (CNN) branches before late fusion, a decoupled design intended to preserve local complementarity-determining region (CDR) signal that monolithic encoders can dilute. On a sequence-disjoint split of the AbRank benchmark, the model achieves a Pearson correlation of 0.683, an R^2 of 0.460, and a pairwise ranking AUC of 0.895, significantly outperforming single-branch ablations (paired t-test, p < 0.05). Attention-map and gradient-based saliency analyses further show that the model preferentially attends to CDR loops and candidate epitope residues, supporting its use as a scalable, structure-free tool for high-throughput antibody screening.

Aicha Boutorh, Soumia Bouyahiaoui, Manel Kara Laouar +3

National School of Artificial Intelligence (ENSIA) Sidi Abdallah Campus, Algiers, Algeria · Lamine Debaghine University Hospital Center, CHU Bab El Oued, Algiers, Algeria

Antibody Sequence Modeling Frozen Language Model Streaming

Date pendingcs.LG

Dead Weights, Live Signals: Feedforward Graphs of Frozen Language Models

We present a feedforward graph architecture in which heterogeneous frozen large language models serve as computational nodes, communicating through a shared continuous latent space via learned linear projections. Building on recent work demonstrating geometric compatibility between independently trained LLM latent spaces~\cite{armstrong2026thinking}, we extend this finding from static two-model steering to end-to-end trainable multi-node graphs, where projection matrices are optimized jointly via backpropagation through residual stream injection hooks. Three small frozen models (Llama-3.2-1B, Qwen2.5-1.5B, Gemma-2-2B) encode the input into a shared latent space whose aggregate signal is injected into two larger frozen models (Phi-3-mini, Mistral-7B), whose representations feed a lightweight cross-attention output node. With only 17.6M trainable parameters against approximately 12B frozen, the architecture achieves 87.3% on ARC-Challenge, 82.8% on OpenBookQA, and 67.2% on MMLU, outperforming the best single constituent model by 11.4, 6.2, and 1.2 percentage points respectively, and outperforming parameter-matched learned classifiers on frozen single models by 9.1, 5.2, and 6.7 points. Gradient flow through multiple frozen model boundaries is empirically verified to be tractable, and the output node develops selective routing behavior across layer-2 nodes without explicit supervision.

Marcus Armstrong, Navid Ayoobi, Arjun Mukherjee

Department of Computer Science University of Houston Houston, TX 77204

Frozen Language Model Feed-Forward Cross-Attention Latent Space Private Signals