Efficient Language Model Inference

Feed-forward networks (FFNs) dominate the parameter count and computation of modern language models, yet existing pruning methods often struggle to convert sparsity into hardware-friendly inference efficiency gains. We introduce PrunePath, a budget-adaptive structured sparsification framework for FFN layers. Built on MoEfication, PrunePath replaces independent expert-wise thresholding with a softmax-normalized routing distribution and activates important experts under a cumulative-mass threshold. This formulation imposes a token-level probability budget, enabling adaptive expert counts and a direct inference-time sparsity knob from a single checkpoint. Across NLU, NLG, and instruction-tuning evaluations, PrunePath achieves a favorable sparsity--performance trade-off compared with existing static pruning and MoEfication-based methods. We further implement Triton kernels for KV-cache decoding to translate the resulting structured sparsity into practical memory savings and measurable decoding-speed improvements. These results demonstrate the superior performance of PrunePath for building highly sparse, deployment-friendly large language models.

Inference-time sampling can elicit strong reasoning abilities from language models without additional training. Existing power-sampling methods do so by sharpening the distribution over full generated outputs, favoring completions that are individually likely under the model. We argue that this is the wrong object to target for reasoning: a completion entangles a reasoning trace with a final answer, whereas what matters is whether an answer is supported by many plausible reasoning paths. We therefore shift the target from the full-output distribution to the sharpened answer marginal, making self-consistency an inference-time objective rather than a post-hoc voting criterion. Surprisingly, this marginal target admits an efficient approximation: we propose a simple, purely autoregressive parallel sampling algorithm that approximately samples from the sharpened answer marginal, eliciting stronger performance than standard power sampling on mathematics and coding benchmarks while being orders of magnitude faster.

Reasoning models often generate long traces in which useful self-correction and unproductive revision are hard to distinguish. We study this distinction through backtracking dynamics: local reconsideration, retraction, or re-derivation inside long-form reasoning traces. On 6{,}000 Qwen3-8B AIME traces, we annotate segment-level backtrack severity and analyze event timing, normalized depth, and local burst structure. We find that early isolated repair is often compatible with correct reasoning, whereas incorrect traces more often show moderate-to-severe backtracks that persist and cluster late. Cross-corpus checks show the same qualitative asymmetry across additional model/domain pairs. Filtering analyses instantiate the signal as a prefix-causal selective early-exit policy: at shallow and intermediate depths, burst-aware filtering outperforms fixed length-based filtering while using only prefix-available features. Moderate length cutoffs remain strong completed-trace baselines, but burst-aware control provides a deployable mechanism for separating recoverable repair from likely instability.

Mixture-of-Experts (MoE) has become the de facto architecture for hundred-billion-parameter language models, yet its advantages at sub-billion scales for on-device deployment remain largely unexplored. To close this gap, we present MobileMoE, a family of on-device MoE language models with sub-billion active parameters (0.3-0.9B active and 1.3-5.3B total) that establish a new Pareto frontier for on-device LLMs. We first formulate an on-device MoE scaling law that jointly optimizes MoE architecture under mobile memory and compute constraints, identifying an on-device sweet spot - moderate sparsity with fine-grained and shared experts - that is simultaneously memory and compute-optimal. Building on the derived architectures, we train MobileMoE with a four-stage recipe covering pre-training, mid-training, instruction fine-tuning, and quantization-aware training, all on open-source datasets. Across 14 benchmarks, MobileMoE matches or exceeds leading on-device dense LLMs with 2-4$\times$ fewer inference FLOPs, and matches or surpasses the state-of-the-art MoE OLMoE-1B-7B with up to 60% fewer parameters. To bridge the last mile to mobile deployment, we provide the first efficient MoE inference on commodity smartphones with comprehensive on-device profiling. At comparable INT4 weight memory, MobileMoE-S delivers $1.8$-$3.8\times$ faster prefill and $2.2$-$3.4\times$ faster decode than the dense baseline MobileLLM-Pro.

Large Language Models (LLMs) and Multimodal LLMs (MLLMs) have demonstrated immense potential in autonomous driving (AD) by offering human-like reasoning and open-world generalization. However, the excessive computational overhead and high inference latency of these massive models severely hinder their deployment in resource-constrained AD systems. To address this challenge, we propose a novel decision-making framework utilizing a lightweight confidence-aware language model, which bridges the gap between complex multimodal intention reasoning and efficient inference. Specifically, we design a multi-agent collaborative workflow, comprising action voting, confidence assessment, and summarization agents, to generate high-quality, confidence-annotated decision demonstrations via explicit Chain-of-Thought (CoT) reasoning. These demonstrations are then distilled into a lightweight language model featuring a dual-head architecture, enabling the joint prediction of decision probabilities and the generation of textual rationales. The distillation is realized via a confidence-aware fine-tuning strategy coupled with Retrieval Augmented Generation (RAG) to enhance the model's adaptability and data efficiency. Comprehensive closed-loop experiments on the nuPlan benchmark demonstrate that our approach achieves state-of-the-art (SOTA) success rates in both regular and long-tail scenarios while maintaining low inference latency.

Deploying Large Language Model (LLM) applications, particularly those relying on Retrieval-Augmented Generation (RAG), remains challenging due to high computational demands, outdated knowledge bases, and the need to manually select optimal pipeline components. In this work, we propose a modular framework for benchmarking and guiding the efficient development of RAG applications by focusing on resource telemetry and component recommendation, suggesting the best components for a domain-specific dataset. Our approach leverages core techniques in LLM applications, including document chunking, vector databases, embedding models, and retrievers, to evaluate trade-offs among accuracy, efficiency, and scalability. By directly correlating retrieval and generation quality with underlying hardware constraints, RAGe supports researchers to identify the most effective, domain-specific RAG setups for their specific operational needs, facilitating rapid prototyping even on consumer-grade hardware.

Although large language model (LLM) conversational systems process millions of multi-turn dialogues daily, they remain fundamentally reactive: they respond only after the user types a query. A key step toward proactive interaction is next-query prediction, which anticipates the user's subsequent query based solely on the preceding dialogue. Progress on this task is hindered by the lack of dedicated benchmarks and a fundamental efficiency--quality trade-off: naively concatenating full dialogue history incurs linearly growing token consumption, while truncating to the latest turn discards crucial cross-turn context. Our key insight is that accurate prediction does not require re-reading raw history; it suffices to track the user's evolving intent trajectory across topics, unresolved needs, and interest shifts. We propose OnePred, which maintains a recursively updated memory as its sole cross-turn context, bounding the per-turn cost independently of conversation length. We train the model via a two-stage reinforcement learning pipeline that first teaches what to predict, then what to compress, shaping the memory into a prediction-oriented intent chain. To establish a rigorous testbed, we introduce NQP-Bench, spanning three diverse subsets. Experiments demonstrate that OnePred reduces per-turn token consumption by up to 22$\times$ compared to full-history inputs while consistently exceeding all baselines in prediction quality, with larger gains on longer conversations. Our code is publicly available at https://github.com/ZBWpro/OnePred.

In the competitive landscape of sponsored search, balancing retrieval quality with production latency is a critical challenge. While large retrieval models based on Small Language Models (SLMs) such as Qwen3-Embedding-4B/8B set strong upper bounds on public benchmarks, their deployment in high-throughput, latency-sensitive environments remains impractical. In this paper, we present HARNESS-LM (HLM), a three-phase training framework for transferring the capabilities of large-scale retrievers into compact, cost-efficient models. The approach comprises: (1) training a high-performance reference ("teacher") retriever by fine-tuning a billion-parameter-scale SLM; (2) aligning query representations via an L2 objective to distill knowledge into a sub-600M parameter student encoder; and (3) applying a final contrastive refinement stage to optimize the student for retrieval performance. We also present a comprehensive empirical study of key design choices, including alignment objectives, embedding dimensionality, model scale, architecture, and optimization strategies, to identify configurations that are most effective in production settings. On a real-world Bing Ads evaluation benchmark, HLM recovers over 98% of the reference retriever's precision across multiple settings, while delivering up to 27x lower online query-encoder latency and 20x higher throughput on NVIDIA A100 GPUs. Online A/B testing on Bing Ads further shows a +1% Revenue, +0.6% Impression, and +0.4% Click uplift over the current ensemble of retrievers running in production with the deployed 190M parameter model, clearly highlighting the practical efficacy of the HLM recipe in a real-world sponsored search setting.

Automated scoring of ESG narrative disclosures with large language models (LLMs) is gaining traction, yet whether reasoning-heavy frontier models add value commensurate with their cost remains empirically unsettled. We evaluate this question on a corpus of ten Japanese listed firms across three rubric axes -- quantitative targets, progress-tracking infrastructure, and external-standard alignment -- using a four-model consensus design that combines a reasoning-on frontier model with three reasoning-off contemporaries. Across 120 firm x axis x model scores, the pooled mean absolute deviation between the reasoning-on model and each reasoning-off counterpart is 0.38 on a 5-point scale; only 2% of pairwise comparisons reach a two-point deviation, and none exceeds two points. Per-firm cost accounting shows the reasoning-on arm alone costs roughly 5.6x as much as the three-provider reasoning-off ensemble, for outcomes that differ only within small margins. We conclude that in span-based ESG narrative scoring, reasoning-heavy deployment does not materially improve outcomes relative to reasoning-off consensus, while substantially increasing operational cost. We discuss implications for cost-effective ESG auto-scoring pipelines and LLM deployment governance in applied accountability settings. An earlier version of this work is available on SSRN (Abstract ID 6683303).

Large Reasoning Language Models (LRMs) tackle complex tasks through test-time scaling, generating long-form Chain-of-Thought (CoT) trajectories during inference. These trajectories contain reflection markers such as wait, but, and alternatively, accompanying hesitation, revision, and alternative explorations. Recent test-time control methods use these markers to steer reasoning, but treating them as one coarse-grained category can obscure their distinct intervention effects. We study how marker identity and reasoning context shape these effects. Type-wise suppression produces different accuracy--length tradeoffs, while fixed-prefix interventions show that the relative benefit of continuation and revision varies across prefixes and tasks. Motivated by these findings, we introduce PathCal, a training-free controller that adjusts the strength of a fixed continuation bias using instantaneous competition among continuation, revision, and alternative-opening markers. It applies finite logit adjustments without additional model forward passes. An exact log-odds identity and frozen-prefix probes establish its local directional effect. Experiments show that PathCal improves the accuracy--length tradeoff in many mathematical reasoning settings and also reduces generation length in coding and multilingual mathematics.

Aspect-Term Sentiment Analysis (ATSA) in multi-aspect sentences faces a fundamental tradeoff between efficiency and expressiveness. Existing models either re-encode the sentence for each aspect or rely on static use of deep representations, leading to redundant computation and limited adaptivity. We argue that Transformer depth is a costly, queryable resource, and propose DABS, a single-pass inference framework that encodes each sentence once to construct a reusable, depth-ordered substrate. Each aspect then queries this shared representation to selectively read relevant tokens and abstraction levels, without re-encoding. This decouples shared sentence encoding from lightweight, aspect-conditioned readout. Experiments on four ATSA benchmarks show that DABS achieves competitive performance while reducing end-to-end computation by up to 60% in multi-aspect settings (M >= 2). Further analyses indicate that adaptive depth querying is most beneficial for linguistically complex cases such as negation and contrast. Code is publicly available at https://github.com/panzhzh/acl-dabs

In-context learning (ICL) allows large language models (LLMs) to adapt to new tasks through demonstrations, yet it suffers from escalating inference costs as context length increases. While task vectors offer a promising alternative by compressing demonstrations into compact hidden-state representations, their quality has been evaluated only through downstream task accuracy. This indirect criterion provides limited insight into how to design more effective task vector extraction methods. In this paper, we posit that inference using task vectors should align their predictive distribution with that of ICL. To quantify this, we introduce $d_{\text{NTP}}$, a metric that measures the discrepancy in next-token probabilities between task vector-based and ICL-based inference. Our empirical analysis reveals that $d_{\text{NTP}}$ serves as a performance proxy, exhibiting a strong negative correlation with downstream accuracy. Motivated by this, we develop Linear Task Vector (LTV), a method designed to minimize $d_{\text{NTP}}$ via a closed-form linear mapping that estimates demonstration effects through regression. Across eight classification benchmarks and five LLMs, LTV consistently outperforms existing task vector baselines, improving average accuracy by 9.2% while reducing inference latency. We further show that LTV outperforms the baselines on regression tasks. Moreover, we investigate the transferability of LTV across different model scales; an aspect that has remained nascent in task vector research. Specifically, we empirically show that task vectors from a larger model can enhance a smaller model's performance by 6.4%, suggesting a new utility for extracted task representations.

Chain-of-thought (CoT) reasoning can improve LLM performance, but its benefits vary across tasks, models, and individual queries. On some tasks, CoT provides marginal or negative accuracy gains while substantially increasing token consumption. We investigate whether early decoding dynamics can help identify when explicit reasoning is useful. Our analysis shows that tasks with larger CoT gains tend to exhibit decreasing early-stage entropy, while low-gain tasks more often show oscillatory or increasing trajectories. These patterns motivate an empirical regime interpretation of decoding behavior. We propose EDRM (Entropy Dynamics-based Reasoning Manifold), a lightweight routing framework whose default heuristic requires no router training. EDRM summarizes early entropy trajectories with a compact, interpretable representation and selects among Direct, Standard, and CoT decoding. Across 15 benchmarks and 7 LLMs, its global and instance-level variants offer 16--55% token savings at comparable accuracy. The Full instance-level configuration combines routing with a short Direct fallback branch; our heuristic-router ablation reports up to 4.9 percentage points of additional accuracy from oracle answer selection over routing without fallback. These results support selective reasoning based on early decoding signals and demonstrate the potential of entropy-based control for efficient LLM inference.

Chain-of-thought (CoT) is a standard approach for eliciting reasoning capabilities from large language models (LLMs). However, the common CoT paradigm treats thinking as a prerequisite for answering, which can delay access to plausible answers and incur unnecessary token costs even when the model is able to identify an answer before extended thinking, a behavior known as performative reasoning. In this paper, we introduce CopT, a reformulated reasoning pipeline that reverses the usual order of thinking and answering. Instead of thinking before answering, CopT first elicits a draft answer and then invokes subsequent on-policy thinking conditioned on its own draft answer for reflection and correction. To assess whether the draft answer should be trusted, CopT recasts continuous embeddings as inference-time contrastive verifiers. Specifically, it contrasts the model's support for the same generated tokens under discrete-token inputs and continuous-embedding inputs, yielding a sequence-level reverse KL estimator for answer reliability. Our analysis shows that under certain assumptions, the expected estimate equals the mutual information between the unresolved latent state and the emitted answer token, explaining why it captures answer-relevant uncertainty rather than arbitrary uncertainty in the latent state. When the answer is deemed insufficiently reliable, CopT performs further on-policy thinking, where a second KL estimator dynamically controls draft-answer visibility, preserving useful partial information while reducing the risk of being misled by unreliable content. Across mathematics, coding, and agentic reasoning tasks, CopT improves peak accuracy by up to 23% and reduces token usage by up to 57% at comparable or higher accuracy, without any additional training. The code is available at https://github.com/sdc17/CopT.

State-space models (SSMs) face a fundamental trade-off between efficiency and expressivity that is mainly dictated by the structure of the model's transition matrix. Unstructured transition matrices enable maximal expressivity, as measured by their ability to model finite-state automaton (FSA) transitions, but come at a prohibitively high compute and memory cost. In contrast, most structured transition matrix forms are highly efficient both in runtime and memory consumption, but suffer from limited expressivity. Building on recent work on structured sparse SSMs, we propose Flash PD-SSM, a novel SSM that achieves comparable throughput to widely-used structured SSMs with significantly better expressivity guarantees. Flash PD-SSM maintains a trainable set of structured sparse matrices, a single one of which is discretely selected at each time-step, enabling FSA expressiveness at the level of unstructured matrices while maintaining the efficiency required for training models at scale. First, we validate Flash PD-SSM against a suite of alternative models on synthetic mechanistic and state-tracking tasks, finding that its theoretical expressivity is achieved in practice. Second, on multivariate time-series tasks involving sequences of length over 17,000, we find that Flash PD-SSM defines a new state-of-the-art (SoTA) accuracy among competing SSM methods. Finally, we demonstrate that Flash PD-SSM is an effective drop-in replacement for hybrid LLMs, yielding improvements both in natural language state-tracking and in common language modeling scenarios. The model exhibits increased throughput and decreased memory consumption compared to SSMs widely used in frontier language models.

Autoregressive language models execute Transformer layers sequentially, creating a latency bottleneck that is not removed by conventional tensor or pipeline parallelism. We study whether this layerwise dependency can be relaxed by treating the hidden-state trace across layers as the solution of a nonlinear residual equation and solving it with parallel Newton-style updates. While this view is principled, exact Newton corrections require expensive Jacobian-vector products and naive fixed-point iterations are unstable on trained Transformers. We introduce Structured Newton Layer Parallelism (SNLP), a training and inference framework that replaces exact layer Jacobians with cheap architecture-induced surrogate dynamics. In residual Transformers, this yields Identity Newton (IDN), where the correction reduces to a prefix-sum-like update; in mHC-style architectures, HC Newton (HCN) uses the model's residual mixing matrix. We also study SNLP-aware training, including pretraining regularization and direct SNLP-forward SFT. Experiments on Nanochat-scale Transformers show that SNLP exposes a practical speed-quality frontier: on 0.5B models, it reaches up to 2.58x wall-clock speedup, and a less aggressive configuration reaches 1.40x speedup without increasing PPL. The useful tradeoff comes from the biased finite-iteration computation induced by IDN/HCN rather than exact recovery of the sequential trace. We further show that SNLP-forward SFT can preserve downstream task accuracy, and that SNLP can serve as a drafter for self-speculative decoding while a sequential verifier preserves output correctness.

Large Reasoning Models (LRMs) achieve strong performance by generating long chains of thought (CoT), but often overthink, continuing to reason after a solution has already stabilized and thereby wasting tokens and increasing latency. Existing inference-time early-exit methods rely primarily on answer-level signals, such as confidence or trial-answer consistency, to decide when to stop. However, these signals mainly reflect answer readiness rather than reasoning convergence: they may trigger before the model has finished exploring or self-correcting, causing premature exits that can degrade final-answer accuracy and leave the retained reasoning chain semantically incomplete. We identify reasoning-level semantic redundancy as a complementary signal for semantic-preserving early exit: when successive steps no longer add novel progress and instead revisit established conclusions, the reasoning trajectory has likely converged. Building on this insight, we propose PUMA, a plug-and-play framework that combines a lightweight Redundancy Detector with answer-level verification. The detector flags semantically redundant candidate exits, while verification confirms whether stopping is safe, allowing PUMA to remove redundant continuation while preserving both answer accuracy and a coherent reasoning prefix. Across five LRMs and five challenging reasoning benchmarks, PUMA achieves 26.2% average token reduction while preserving accuracy and retained CoT quality. Additional experiments on code generation, zero-shot vision-language reasoning, and learned stopping-policy internalization further demonstrate that reasoning-level redundancy is a robust, transferable, and learnable signal for efficient reasoning. Our code is available at https://github.com/giovanni-vaccarino/PUMA.

Sub-billion-parameter Transformer language models are increasingly deployed on edge devices, where the privacy, latency, and operating-cost advantages of on-device inference are constrained by tight memory-bandwidth, energy, and thermal budgets that make architectural choice and accelerator-specific cost central to efficient inference. We present LLMForge, a hardware-aware neural architecture search (NAS) framework whose three composable contributions together make edge-LM architecture search hardware-conditioned, since different substrates impose different hardware cost bottlenecks. Infinite-Head Attention (IHA) decouples the number of query heads, KV groups, and per-head query/key and value dimensions, expanding the feasible per-layer attention configuration space by approximately 400x over grouped-query attention within our search-space ranges. Forge-Former, an encoder-based surrogate for ranking architectural candidates, outperforms MLP and random-forest baselines. Forge-DSE, an NSGA-II-based design-space-exploration engine, pairs Forge-Former with a multi-backend hardware cost model spanning GPUs, systolic accelerators, and ring-dataflow edge accelerators. Across four different hardware substrates, the searches converge to visibly different architectures whose shapes track each substrate's cost bottleneck. On the multi-chip ring substrate, our co-search returns three 300M-scale deployment-aware variants on the Pareto front. Each is re-trained on FineWeb-Edu-10BT under matched recipe against SmolLM2-360M and Qwen-0.5B architecture baselines. The accurate variant has the lowest validation loss 2.798 and competitive benchmark performance with fewer parameters, the energy-optimized variant lowers energy per token by 40%, and the latency-optimized variant lowers TTFT and TPOT by 43%.

Parallel reasoning, where a generator samples many candidate solutions and an aggregator selects the best, is one of the most effective forms of test-time scaling in large language models, and pairwise self-verification has become its strongest aggregation primitive. Yet pairwise verification carries a heavy cost: each judgment reads two complete solutions in full, and existing methods perform tens of such judgments per problem regardless of whether the comparison is informative. We introduce CAPS (Cascaded Adaptive Pairwise Selection), an inference-only framework that allocates verifier compute non-uniformly along two orthogonal axes: an evidence axis that adapts how much of each candidate the judge sees, and a distribution axis that adapts how comparisons are spread across the pool. CAPS instantiates these into a four-stage cascade with an optional rescue subroutine, and admits a closed-form verifier-token cost in which the per-candidate marginal cost is roughly halved relative to uniform full-evidence schedules. On four self-verifying models (Qwen3-14B, GPT-OSS-20B, Qwen3-4B-Instruct/Thinking) and five reasoning benchmarks spanning code (LiveCodeBench-v5/v6, CodeContests) and math (AIME 2025, HMMT 2025), CAPS outperforms the leading pairwise verifier on 14 of 20 suites while using 25.4% of its verifier-token budget on code, and outperforms pointwise self-verification on all 20. The trade-off suites admit an interpretable diagnostic in terms of the verifier's accuracy at partial versus full evidence, providing a concrete pre-deployment check for cascade suitability.

Multi-Level Intermediate Representation (MLIR) underlies modern ML compiler infrastructure (TensorFlow, JAX/StableHLO, PyTorch Inductor, IREE), yet appears only in trace amounts in code-LM pretraining corpora. MLIR is also extensible by design: new dialects ship per application domain, so a fine-tuned model per dialect does not scale. We ask whether inference-time priors derived mechanically from each dialect's Operation Definition Specification (ODS) can substitute for gradient-based adaptation. First, we release four natural-language-to-MLIR benchmarks across three dialects (MLIR-Spec-150, Linalg-Spec-30, StableHLO-Spec-30, StableHLO-Held-Out-200), totaling 410 in-scope NL-to-MLIR pairs, plus a 25-program out-of-grammar stress set, with functional references on 30 Spec prompts, under Apache-2.0 with Gebru datasheets and Croissant 1.1 metadata. Second, we build a three-layer schema-derived constraint stack: a CFG over op signatures (C1), type-domain splits from an ODS-extracted type lattice (C2), and an SSA-scope validator driving five-retry rejection sampling (C3). Porting it to StableHLO required no new constraint-layer code. On structurally dominated dialects, SmolLM2-1.7B with token-level constraints matches or exceeds 15B-34B code LMs running post-hoc rejection at a matched retry budget and 8-25x the per-generation speed: on linalg it reaches 80.0% verify-valid (three runs, n=125), 21 to 44 pp above CodeLlama-34B, Granite-Code-34B, and StarCoder2-15B with non-overlapping CIs. The comparison is between systems, not models: under the baselines' post-hoc protocol, SmolLM2 falls from 52.0% to 24.5% on arith+func. On arith+func and the templated StableHLO-Held-Out-200 the baselines match or beat the SLM; we scope these as non-win cells. We release benchmarks, decoder, every per-prompt generation, and a pinned reproducibility Dockerfile.

In current Large Language Models we can trust the production of smoothly flowing prose on the basis of the principles of machine learning. However, there is no comparably principled basis to justify trust in the content of the text produced. It appears to be conventional wisdom that addressing this issue by adding more principled reasoning is not computationally affordable. Here we propose a principled method of reasoning that is efficient enough to be practical for large language models. Further, the method allows the retention of much of the currently used software and hardware base. Our method for improving the functioning of large language models consists of a first stage of preprocessing that recodes the data to a Unary Relational Integracode that is more explicit about the relationships among the objects described in the text, followed as a second stage by a standard but possibly streamlined machine learning process that then also learns to predict these relationships. The method may be viewed as realizing a world model and applying beyond natural language, to vision and actions, for example, where the multiple properties of an object referred to in an input are brought together explicitly, rather than remaining distributed in the various references to it in the input. We articulate its advantages in terms of Robust Logic, a system for performing principled chaining on learned, and hence uncertain, information. We show that this recoding has the surprising and fortuitous property that, while succinct, it makes the task of learning a core subset of relational rules that hold in the world described in the training data polynomial time learnable in a defined sense, the polynomial depending on the complexity of the rule. This gives support for sound reasoning within each single call of the learned classifier as well as between multiple calls.

The continued improvements in language model capability have unlocked their widespread use as drivers of autonomous agents, for example in coding or computer use applications. However, the core of these systems has not changed much since early instruction-tuned models like ChatGPT. Even advanced AI agents function on message exchange formats, successively exchanging messages with users, systems, with itself (i.e. chain-of-thought) and tools in a single stream of computation. This bottleneck to a single stream in chat models leads to a number of limitations: the agent cannot act (generate output) while reading, and in reverse, cannot react to new information while writing. Similarly, the agent cannot act while thinking and cannot think while reading or acting on information. In this work, we show that models can be unblocked by switching from instruction-tuning for sequential message formats to instruction-tuning for multiple, parallel streams of computation, splitting each role into a separate stream. Every forward pass of the language model then simultaneously reads from multiple input streams and generates tokens in multiple output streams, all of which causally depend on earlier timesteps. We argue that this data-driven change remedies a number of usability limitations as outlined above, improves model efficiency through parallelization, improves model security through better separation of concerns and can further improve model monitorability.

Transformer-based large language models face severe scalability challenges in long-context generation due to the computational and memory costs of full-context attention. Under practical computation and memory constraints, many inference-efficient long-context methods improve efficiency by adopting bounded-context or segment-level execution only during inference, while continuing to train models under full-context attention, resulting in a mismatch between training and inference execution and state-transition semantics. Based on this insight, we propose a training-inference consistent segment-level generation framework, in which training and inference follow the same segment-level forward execution semantics. During training, consistency with inference is enforced by restricting gradient propagation to KV states carried over from the immediately preceding segment, while permitting head-specific access to past KV states during the forward pass without involving them in gradient propagation. Across long-context benchmarks, our approach achieves performance comparable to full-context attention, while achieving competitive latency-memory trade-offs against strong inference-efficient baselines, and substantially improving scalability at very long context lengths (e.g., approximately 6x lower peak prefill memory at 128K compared to full-context attention with FlashAttention).

Large reasoning models (LRMs) improve problem solving through extended reasoning, but often misallocate test-time compute. Existing efficiency methods reduce cost by compressing reasoning traces or conditioning budget on perceived difficulty, yet largely overlook solvability. As a result, they may spend large budgets on queries beyond the model's capability while compressing hard-but-solvable queries that require deeper reasoning. In this work, we formulate adaptive reasoning as a computational investment under uncertainty, where budget should follow the expected return of reasoning rather than perceived difficulty alone. To instantiate this principle, we propose Budget-Efficient Thinking (BET), a two-stage framework that combines behavioral cold-start with GRPO under an investment-cost-aware reward. By aligning solve-or-fold decisions with rollout-derived solvability, BET learns three behaviors: (1) short solve, answering easy queries concisely; (2) nice fold, abstaining early when continued reasoning has near-zero expected return; and (3) hero call, preserving sufficient compute for hard-but-solvable queries. Across seven benchmarks and three base models, BET reduces reasoning tokens by ~55% on average while achieving overall performance improvements, and transfers zero-shot from mathematical reasoning to scientific QA and logical reasoning with comparable efficiency gains.

Large Language Models (LLMs) are increasingly deployed in multi-turn dialogue settings where preserving conversational context across turns is essential. A standard serving practice concatenates the full dialogue history at every turn, which reliably maintains coherence but incurs substantial cost in latency, memory, and API expenditure, especially when queries are routed to large proprietary models. Existing approaches often struggle to balance the trade-off between response quality and efficiency. We propose a framework that exploits the early turns of a session to estimate a local response manifold and then adapt a smaller surrogate model to this local region for the remainder of the conversation. Concretely, we learn soft prompts that maximize semantic divergence between the large and surrogate small language models' responses to surface least-aligned local directions, stabilize training with anti-degeneration control, and distill the mined cases into localized LoRA fine-tuning so the surrogate runs without prompts at inference. A simple gate enables a one-time switch with rollback on drift. We further provide a theoretical analysis for key components in SOMA. Extensive experiments show the effectiveness of SOMA. The source code is provided at: https://github.com/LabRAI/SOMA.

Today's inference-time workloads increasingly depend on timely access to a model's internal states. We present DMI-Lib, a high-speed deep model inspector that treats internal observability as a first-class systems primitive, decoupling it from the inference hot path via an asynchronous observability substrate built from Ring^2, a GPU-CPU memory abstraction for capturing and staging tensors, and a policy-controlled host backend that exports them. DMI-Lib enables the placement of observation points across a rich space of internal signals and diverse inference backends while preserving serving optimizations and adhering to tight GPU memory budgets. Our experiments demonstrate that DMI-Lib incurs only 0.4%--6.8% overhead in offline batch inference and an average of 6% in moderate online serving, reducing latency overhead by 2x-15x compared to existing baselines with similar observability features. DMI-Lib is open-sourced at https://github.com/ProjectDMX/DMI.

As large language models span dense, mixture-of-experts, and state-space architectures and are deployed on heterogeneous accelerators under increasingly diverse multimodal workloads, optimising inference energy has become as critical as optimizing latency and throughput. Existing approaches either treat latency as an energy proxy or rely on data-hungry black-box surrogates. Both fail under varying parallelism strategies: latency and energy optima diverge in over 20% of configurations we tested, and black-box surrogates require hundreds of profiling samples to generalize across model families and hardware. We present EnergyLens, which uses symbolic regression as a structure-discovery tool over profiling data to derive a single twelve-parameter closed-form energy model expressed in terms of system properties such as degree of parallelism, batch size, and sequence length. Unlike black-box surrogates, EnergyLens decouples tensor and pipeline parallelism contributions and separates prefill from decode energy, making its predictions physically interpretable and actionable. Fitted from as few as 50 profiling measurements, EnergyLens achieves 88.2% Top-1 configuration selection accuracy across many evaluation scenarios compared to 60.9% for the closest prior analytical baseline, matches the predictive accuracy of ensemble ML methods with 10x fewer profiling samples, and extrapolates reliably to unseen batch sizes and hardware platforms without structural modification, making it a practical, interpretable tool for energy-optimal LLM deployment.

Tokenizer-free language models eliminate the tokenizer step of the language modeling pipeline by operating directly on bytes; patch-based variants further aggregate contiguous byte spans into patches for efficiency. However, the average patch size chosen at the model design stage governs a tight trade-off: larger patches reduce compute and KV-cache footprint, but degrade modeling quality. We trace this trade-off to patch lag: until a patch is fully observed, byte predictions within it must rely on a stale representation from the previous patch to preserve causality; this lag widens as patches grow larger. We introduce Scratchpad Patching (SP), which inserts transient scratchpads inside each patch to aggregate the bytes seen so far and refresh patch-level context for subsequent predictions. SP triggers scratchpads using next-byte prediction entropy, selectively allocating compute to information-dense regions and enabling post-hoc adjustment of inference-time compute. Across experiments on natural language and code, SP improves model quality at the same patch size; for example, even at $16$ bytes per patch, SP-augmented models match or closely approach the byte-level baseline on downstream evaluations while using a $16\times$ smaller KV cache over patches and $3$-$4\times$ less inference compute.

Enterprise workloads are dominated by deterministic, structured, and knowledge-dependent tasks operating under strict cost, latency, and reliability constraints. While these are often addressed through large language model (LLM) deployment or distillation into smaller models, we argue this is inefficient, unreliable, and misaligned with enterprise task structures. Instead, AI systems should treat language models as interfaces rather than monolithic engines, externalizing knowledge and computation into dedicated components for greater reliability, scalability, and transparency. Our theoretical evidences show that finite-capacity models cannot fully capture the breadth of knowledge required for enterprise tasks, creating inherent limits to efficiency and interpretability. Building on this, we take the position that language models should primarily be used for structured extraction in deterministic enterprise workflows, while computation and storage are delegated to knowledge bases and symbolic procedures. We formally demonstrate that such modular architectures are more reliable and maintainable than monolithic frameworks, offering a sustainable foundation for enterprise tasks.

Text embedding and generative tasks are usually trained separately based on large language models (LLMs) nowadays. This causes a large amount of training cost and deployment effort. Context compression is also a challenging and pressing task, which is vital to reasoning-driven generation, and agentic tasks requiring long context and continual learning. In this paper, we explore how to unify reasoning-driven generation, reasoning-enhanced text representation and context compression tasks in one forward pass for LLMs. Through meta latent tokens and a unified generative, representative and compressive tuning approach, we propose a training framework named GRC that bridges the three tasks. The trained models can accomplish three objectives in a single forward pass while maintaining modular, LEGO-style flexibility during inference. This design greatly reduces the deployment effort for retrieval-augmented generation (RAG) and achieves efficient inference and three times data utilization during training. Furthermore, this framework design enables a new paradigm for text embedding: self-reason-latent embeds, and a new generation paradigm, latent memory-augmented generation, where compressed and internalized KV cache with O(1) length is used as the updatable memory. We also propose hybrid paged attention to speed up the inference of our models. Extensive experiments on reasoning-intensive retrieval benchmarks, generative tasks, document compression, latency evaluation, and RAG settings demonstrate the effectiveness of our method and may shed light on the truly unified model that can handle reasoning-driven generation, embedding and compression tasks seamlessly.