Inference Cost

Multi-turn agent trajectories often contain redundant rounds (failed tool calls, parallel sub-queries, verification-only steps) that inflate both training and inference cost. We propose viewing each trajectory as a \emph{round-level dependency DAG} that exposes which rounds are globally load-bearing for the final answer, and fine-tune agents on trajectories refined through this DAG. Given an LLM-annotated DAG, these edits are deterministic and interpretable, with optional rephrasing. Models trained on these refined trajectories consistently outperform those trained on the original trajectories at lower inference cost. Specifically, across four multi-modal QA benchmarks, our refinements improve downstream accuracy by up to $1.7$,pp over vanilla SFT (and $5.7$,pp over an LLM-deletion baseline) while reducing per-sample inference messages by up to approximately $40\%$ and inference tokens by up to approximately $48\%$, translating to substantial savings in compute and serving cost. Code is available.

One-step action generation can substantially reduce the inference cost of vision-language-action (VLA) policies, but its effect on closed-loop task success remains an open question. This technical report studies a GR00T N1.7 variant in which the iterative diffusion-transformer action head is replaced by a one-step drifting action head, together with an overlap-conditioned extension for asynchronous chunk replacement. All multi-seed drifting runs were trained on two NVIDIA A800 GPUs. On LIBERO, the action head reduces the mean model-forward time of the action head from approximately $45.3\,\mathrm{ms}$ to $5.0\,\mathrm{ms}$, while the measured backbone-plus-head time falls from approximately $70.0\,\mathrm{ms}$ to $30.6\,\mathrm{ms}$. However, this speedup is accompanied by a systematic reduction in task success. Across three drifting seeds, success is $64.0\pm4.0\%$ on LIBERO-Spatial, $52.0\pm1.0\%$ on LIBERO-Goal, and $26.0\pm2.6\%$ on LIBERO-Long. The low seed variance indicates that the degradation is not explained by random initialization alone. We report the result as a speed--success trade-off rather than an overall improvement, and discuss likely contributing factors including deterministic one-step mode averaging, batch-dependent geometry estimation, long open-loop chunk execution, and the fact that synchronous LIBERO evaluation does not exercise the asynchronous overlap path.

Conv-TasNet has served as a strong baseline for time-domain speech separation, and many studies have extended it with advanced architectures such as dual-path networks, U-Nets, and attention mechanisms. However, these methods often introduce high computational cost and complexity, limiting their deployment in resource-constrained scenarios. To address this issue, we propose eConv-TasNet, an efficient variant of Conv-TasNet that improves both effectiveness and efficiency without relying on resource-intensive modules. The proposed model consists of a group-wise early-splitting (GES) module and a multi-group feature aggregation (MGFA) module. GES generates discriminative speaker embeddings at intermediate stages, while MGFA progressively aggregates these group-level representations for refined mask estimation. Experimental results show that eConv-TasNet reduces model size by 22.4%, accelerates inference by 18.9%, and improves SI-SNRi by 14.0%-28.0% across three public benchmarks. Moreover, it achieves competitive performance compared with state-of-the-art methods while requiring significantly fewer parameters and lower inference cost. These results demonstrate a favorable efficiency-effectiveness trade-off for edge deployment.

In-vehicle assistants must translate natural-language requests into accurate vehicle function calls under strict memory and latency constraints, making small language models (SLMs) attractive for on-device deployment. For such models, a key design choice is how the available function surface is presented. Two approaches are to represent each function with a dedicated Functional Token (FT) or provide function schemas directly in the prompt. FTs enable compact inference but are restricted to functions learned during training, whereas Schema-in-Prompt (SIP) can generalize to unseen functions at the cost of longer prompts and higher inference overhead. We introduce a benchmark of 9,822 single-turn examples spanning 79 vehicle functions derived from Android Automotive, including held-out functions and requests requiring refusal. We compare both approaches under matched fine-tuning across four SLMs from 270M to 1.7B parameters. On functions seen during training, scaling provides limited benefit: the 270M model can match the 1.7B model, while the strongest overall performance occurs at 0.6B. On held-out functions, FT achieves zero accuracy by construction, whereas SIP generalizes and improves substantially with scale. On out-of-scope requests, FT can invoke an unavailable function it was trained to emit, while SIP more reliably refuses based on the functions offered. This flexibility comes with higher memory use and latency. Our theoretical analysis explains how SIP enables generalization and why longer schema contexts increase inference cost. Overall, function-surface representation, rather than model scale alone, determines the capabilities and failure modes of SLM-based vehicle function calling.

A natural way to cut reasoning-model inference cost is to repeatedly probe a single partial trajectory for its current answer and stop once probes agree -- self-consensus. We ask whether any such rule is both safe and token-saving, and whether one can be selected once and reused. A preregistered sweep of 3,520 consensus rules, replayed on frozen trajectories from two models and three benchmarks, clears none of three acceptance gates fixed in advance; the frontier reproduces on a held-out split and on two unseen models -- while a boundary-confidence control (DEER) swept through the same pipeline clears all three. The reason lies in the signal: agreement establishes that the current answer persists under a fixed probing procedure, not that the reasoning has terminated -- a consensus-termination gap. Stopping on it commits non-terminal answers. At a rule still saving 32% of the tokens, one stop in nine fires on an answer the trajectory itself later abandons, and most of those stops cut off a correction it would otherwise have made. Widening the agreement window does not remove them: the share levels off near 7%, and by then the saving has fallen to 8%. Probe re-wording and a hand-labelled error taxonomy show the agreed answer is often a placeholder the model had not settled on. Used on its own as the stop signal, agreement fails not because it is insufficiently strict, but because it repeatedly measures the wrong object.

Entity matching (EM) requires fine-grained contextual understanding and domain knowledge. Recent work shows that large language models (LLMs) can serve as strong matchers across domains, but most methods either make independent pairwise decisions or rely on manually designed composite pipelines, thus lacking flexibility in realistic multi-candidate settings. At the same time, they typically ignore inference cost at scale. We formulate LLM-based EM with candidates as a cost-aware sequential decision problem and propose CaRL-EM, a reinforcement learning controller that manages LLM operations. Given the state of an anchor record, its candidate set, and the cost, CaRL-EM adaptively chooses among different operators (Match/Compare/Select/Decide) and model capacities to maximize a quality-cost objective. The policy interacts with abstract operators, allowing the same controller to be reused with different underlying LLM backends at inference time without retraining. Experiments on 7 benchmarks show that CaRL-EM (i) learns to dynamically plan the usage of inexpensive and expensive operators based on task complexity, (ii) achieves robust zero-shot transfer across diverse datasets and domains, and (iii) consistently achieves a better quality-cost trade-off than strong LLM-based baselines and manually designed pipelines, yielding a lower inference cost at comparable or higher quality.

Conventional algebraic triangulation solves 3D human pose estimation (HPE) from multi-view 2D keypoints. The typical approach, decoding 2D keypoints from predicted heatmaps, is unreliable as heatmaps can be multimodal under occlusion, and collapsing them into single peaks discards their spatial distribution. We seek to use the entire heatmap to estimate 3D poses more accurately, which requires solving two problems: how to robustly fuse heatmaps across views, and how to assess the reliability of heatmaps. For the former, we introduce a novel objective, Multi-viewExpected-OKS Maximization (MEOM), that locates a 3D joint where the views agree in probability mass. For the latter, we adopt highest-density-region (HDR) calibration as a diagnostic of that mass, independently of distance-based metrics. The proposed framework covers two settings, with and without 3D supervision. Without 3D supervision, we optimize 3D poses from pretrained heatmap predictors by maximizing MEOM, achieving comparable performance with state-of-the-art methods that rely on larger backbones, temporal fusion, and simulated 3D data. On ambiguous Human3.6M (H36MA) and occluded CMU Panoptic frames, the advantage is substantial. When 3D labels are available, we train the model end-to-end with a combined MEOM and MSE loss, achieving 19.11 mm absolute MPJPE on Human3.6M outperforming the state-of-the-art volumetric approach on absolute MPJPE at half the inference cost.

Vision-Language-Action (VLA) models are reshaping autonomous driving (AD) by unifying perception, reasoning, and control through language, enabling semantic grounding, interpretable decisions, and better long-tail generalization. But language is expensive onboard: latency and memory budgets are tight, and autoregressive decoding is inherently sequential. This work reframes the central question as when and where language should act at inference, since inference cost recurs at every deployed frame while training cost is paid once. We introduce the Language Residue taxonomy to organize methods by their inference-time use of language: train-time-only supervision (L1), latent non-textual reasoning (L2), conditional invocation (L3), and full per-frame generation (L4). We review representative methods and tag each across five deployment axes (latency, parameters, memory, FLOPs, tokens), analyzing them on major open- and closed-loop driving benchmarks (e.g., nuScenes, NAVSIM, Bench2Drive). We further trace how efficient methods from NLP/LLM are adapted in AD, identifying the constraints and motivations driving these adaptations. A continuously updated repository will be available at Github.

Sparse representations are often expected to make models smaller and also reduce inference cost. For Fourier Neural Operators (FNOs), these objectives are not equivalent or do not always align: removing parts of the learned operator can leave the underlying transforms and dense computations unchanged, while changing the grid on which the model is evaluated can introduce overhead of its own. We therefore distinguish sparsity in the representation, in the stored parameters, in the theoretical operation count, and in measured runtime, and present an empirical study of several routes toward sparse FNOs that tests each transition between them separately. Coarsening the execution grid reduces the theoretical cost without reducing measured latency, and adding a correction term recovers accuracy at the cost of making the model slower. Even an 83% parameter reduction remains slower than the dense baseline under ordinary execution. These results motivate a stricter definition of useful sparsity: the deployed operator must preserve solution accuracy and map its reduced support to a genuinely cheaper execution path.

Large language model-based recommender systems are increasingly adopting slow-thinking models that generate step-by-step reasoning before making predictions, often achieving higher accuracy than fast-thinking models that predict directly. However, their reasoning traces are often unnecessarily verbose, increasing inference costs without commensurate accuracy gains. Existing training-based approaches to reasoning compression often incur substantial adaptation costs, while inference-time methods are brittle and difficult to scale. These limitations motivate model merging as a promising training-free direction for transferring specialised behaviours between models in a shared parameter space. In particular, merging a slow-thinking model with a fast-thinking counterpart provides a natural mechanism for balancing recommendation accuracy and reasoning conciseness. To this end, we propose, to our knowledge, the first model merging framework for reasoning compression in recommender systems. Unlike conventional merging methods that apply uniform merge coefficients across model components, our method performs fine-grained merging at the level of individual attention heads, capturing heterogeneous patterns in recommendation reasoning. Each attention head is assigned a distinct merge coefficient according to its contribution to critical reasoning evidence and its sensitivity to parameter change, enabling selective injection of the concise behaviour of the fast-thinking model into the slow-thinking model and reducing reasoning verbosity without compromising recommendation quality. Experiments on three benchmark datasets show that our method reduces reasoning length by up to 24.3% while outperforming competitive model merging baselines in maintaining recommendation accuracy. The code is available at https://github.com/linhledieu/REAM.

LLM agents execute heterogeneous sequences of model calls within a single task: some invocations require careful reasoning, while others are structured steps such as formatting or tool-argument construction. Prior routing methods exploit this asymmetry by assigning easy invocations to a cheaper small model and difficult ones to a large model. Such policies reduce inference cost, but they leave the small model's capability unchanged, so attainable savings remain bounded by the work the student can already solve. MERA instead improves the small model itself, using a single model invocation as the unit of adaptation. In each cycle, MERA replays failed student invocations to obtain execution-verified teacher demonstrations, distills recurring procedures into an iteratively updated SkillBook, and fine-tunes a student LoRA adapter via supervised learning and optional GRPO. Routing serves as supporting machinery for deployment: the improved student is served behind a cost-calibrated router with verifier-backed fallback, and a candidate SkillBook, adapter, or router is admitted only when joint replay preserves task quality. Empirically, four-cycle adaptation raises Qwen2.5-Coder-1.5B from 28.7% to 49.7% pass on held-out HumanEval+MBPP. Under verifier-backed fallback, the deployed policy retains 88.3% pass at 60.8% of always-Luna cost. On TAU-2, a fine-tuned Qwen3.5-2B improves from 14/35 to 18/35 and matches an unadapted 4B model. These results indicate that verifier-backed multi-cycle adaptation can increase small-model capability, rather than only routing around a fixed student.

We introduce BDH-CQ, a reasoning model that combines in-context learning with recurrent latent reasoning. Inputs presented at inference time continuously update the model's recurrent memory; the model then solves a query through iterative computation in a high-dimensional latent space, without verbalizing its intermediate reasoning. We evaluate the model on the public ARC-AGI-1 evaluation set and use controlled ARC-like interventions to study what it learns from demonstrations, how consistently it applies an inferred transformation, and which concepts remain difficult. A 150M-parameter configuration reaches 29.5% pass@2 at a computed inference cost of $0.0007 per task. This operating point breaks through the previously reported ARC-AGI-1 cost-accuracy Pareto frontier, establishing a new state of the art in benchmark cost efficiency.

Test-time scaling improves LLM accuracy but multiplies inference cost, making the accuracy gained per unit of compute the metric that matters in deployment. Self-consistency is one of the established approaches, which spends this budget entirely on the output side by sampling repeated reasoning paths. We study Test-Time Augmentation (TTA), which extends self-consistency by also perturbing the input, aggregating predictions across transformed versions of the input, and ask whether input-side diversity converts compute into accuracy more efficiently than output-side diversity. We perform a systematic, matched-compute comparison: we evaluate three simple input-side strategies (semantic rephrasing, lexical perturbations, and visual transformations) across six datasets covering general and multilingual knowledge, mathematical reasoning, multi-modal question answering, and sentiment classification, against chain-of-thought prompting and self-consistency. Semantic rephrasing delivers consistent and statistically significant accuracy gains while Pareto-dominating self-consistency on cost-effectiveness, delivering roughly 1.8X more accuracy per dollar and outperforming it on five of six tasks. We further analyze the number of augmentations, multi-modal strategies, and base model scaling, finding that TTA is most cost-effective for mid-tier models where a stronger model is unavailable or too expensive. Our findings indicate that for current mid-tier LLMs, varying the input converts inference compute into accuracy more efficiently than varying the reasoning path alone. The TTA implementation is available at https://github.com/aws-samples/sample-genai-reflection-for-bedrock.

No single large language model (LLM) is optimal across all queries and budget constraints, making model routing essential for cost-effective deployment. Existing routers adopt diverse formulations and implementations, making fair comparison and extension difficult. We present a unified formulation of LLM routing as a sequential decision process characterized by five components: context encoders, model encoders, scoring functions, decision rules, and learning signals, covering single-turn, multi-turn, and personalized routing. Based on this formulation, we develop an automated pipeline for constructing routing supervision and evaluating routers jointly on response quality and inference cost. The resulting benchmark, xRouteBench, spans generic LLM, memory-augmented, vision, time-series, and personalized routing tasks. We further introduce LLMRouter, an open-source modular infrastructure with more than 16 representative routers. Our empirical study shows that learned routers outperform the strongest fixed-model baseline by 14.6% relatively, lightweight routers become more competitive under tight cost constraints, and user-conditioned routing consistently improves personalization.

Pruning reduces the inference cost of large language models, but existing criteria primarily preserve large activations or reconstruct layer outputs. We argue that this overlooks a key computation performed by particularly sparsity-sensitive neurons in the MLP up and gate projections: separating similar inputs into dissimilar outputs. This suggests that effective pruning should preserve not only activations, but also the differences between outputs more broadly. We introduce a family of difference-informed pruning methods built upon this principle. Wisp is a first-order, update-free method that scores weights using input-difference norms, and Wisp+ refines this score neuronwise using the input pairs each neuron separates most strongly. Finally, Whisper is a second-order method that uses a lightly regularized difference Hessian as its reconstruction objective. Across Llama 2 and 3.1 models from 7B to 405B parameters, our second-order variant consistently improves over strong reconstruction-based baselines, while our update-free variants improve over activation-aware baselines, especially in constrained settings. The improvements over Wanda and SparseGPT extend to structured sparsity, downstream evaluations, and other model families. Augmenting stronger techniques such as RIA and ALPS with our difference-informed criteria yields further improvements, shifting the overall accuracy-runtime frontier outward at negligible additional cost. These results suggest that preserving output differences is a broadly useful and composable signal for post-training LLM sparsification.

Chain-of-thought (CoT) reasoning offers a window into the decision-making of large language models (LLMs), which can be monitored for target behaviors by reading the reasoning trace, motivating work on CoT monitorability. Latent CoT approaches, however, replace the explicit tokens with a small number of continuous states, lowering inference costs but removing the readable trace this monitoring relies on. Monitoring then requires alternative access to the model, such as probing its activations or verbalizing the latent states back into text, but how much monitorability these alternatives preserve is unclear. We study this question with a hint-based intervention setup, a proxy for behaviors where models exploit biasing input cues, e.g., an inadvertently leaked answer or a belief stated by the user, without acknowledging them. Taking hint-reliance as the monitorability target, we compare monitors across reasoning modes, from explicit CoT to weakly- and strongly-supervised latent CoT, on math reasoning and question answering. We find that, in this setup, monitorability depends more on properties of the task (such as whether the correct answer constrains the supporting reasoning) and the level of access to model internals than on the reasoning mode.

Large Reasoning Models (LRMs) excel on complex tasks through long chain-of-thought (CoT) reasoning, but their lengthy intermediate steps cause severe overthinking that inflates inference cost. KV-cache compression is a common solution, yet existing reasoning-oriented methods apply a uniform policy across the trajectory and judge compression only by what it removes from the cache. Two observations point the other way. First, a reasoning state's tolerance to context loss varies along the trajectory, and process reward tracks it: deleting tokens at high-reward steps preserves accuracy far better than deleting the same budget at random. Second, compression is not free on the generation side, since a smaller cache leads the model to generate more tokens, partly canceling the saving. Together these motivate coordinating both sides under a single process reward. We propose ReCo (Reward-Coordinated Compression), a step-wise framework in which a lightweight process-reward estimator scores each completed step and drives three components: (1) reward-adaptive KV-cache compression that shrinks the retained cache harder at high-reward steps and less at low-reward ones, (2) a reward-banded penalty on reflection tokens that curbs redundant generation, and (3) confidence-based early stopping that triggers when the reasoning is reliable. Across three reasoning models and six benchmarks, ReCo reduces generated tokens by 37%-65% and end-to-end latency by 2.08x-2.35x over Full CoT, all while largely preserving accuracy.

Heterogeneous language-model ensembles expand the space of candidate responses, yet they lack a principled criterion for when a newly generated answer should supersede an already supported one. We decouple candidate headroom from replacement authority, rendering the latter as an explicit, auditable object. Our proposed method, Agreement-Before-Diversity (ABD), is a frozen, label-free decision rule: an anchor answer is retained if two additional trusted samples corroborate it under a fixed equivalence relation; otherwise, it is replaced by a heterogeneous synthesis. For this gating mechanism, we prove two exact identities. The first shows that the accuracy gap relative to unconditional synthesis is determined jointly by the agreement coverage and the anchor's advantage on the protected subset. The second shows that the gap relative to never synthesizing reflects a contrast between authorized recovery and authorized destruction. Neither identity assumes independence or calibrated confidence, and the expected inference cost is approximately eight minus five times the coverage in number of calls. Under blind, exact-ID evaluation, ABD achieves 59.43% on the complete LiveCodeBench-v6 (vs. 52.57% for Single9 and 52.00% for HAC; n = 175) and 75.00% on an untouched GPQA-Diamond split (both controls at 72.78%; n = 180). Furthermore, these identities localize every aggregate difference to an enumerable protected stratum: no discordant items occur among the 3 protected cases on LiveCodeBench, where coverage bounds the gate's contribution to 1.71 points a priori; 13 versus 8 discordant cases among 132 on GPQA-Diamond; and 12 versus 0 among 71 under a frozen anchor perturbation. Diversity supplies potential; verification structure supplies authority.

Hard prompt compression reduces long-context inference cost by independently scoring tokens, sentences, or chunks and retaining the highest-scoring units under a budget. We identify a structural failure in this procedure: independent selection can split dependent evidence pairs, retaining one member while deleting the other. When retained text contains an answer but deleted text defines the entity needed to interpret it, we call the result referential dangling. At a compression ratio of 0.30, Beaver, which ranks coherent chunks using Qwen3-0.6B embeddings, leaves the answer path incomplete in 34-54% of bridge examples across three multi-hop question answering datasets. On a shared HotpotQA bridge set, all six hard compressors we test exhibit dangling at rates up to 60%, and every document in LongBench-v2 Single-Document QA contains at least one dangling reference. On dangling examples evaluated with Qwen3-8B, reinserting the missing supporting paragraph while removing nonsupporting paragraphs to maintain the token budget improves accuracy by 29-34 percentage points (p < 0.0001), recovering at least 88% of the gap to contexts retaining both supporting paragraphs. Stronger answer models do not absorb the loss: on MuSiQue, GPT-5.5 is 8.8 points less accurate on compressed contexts than on contexts retaining both supporting paragraphs. Finally, we train a compact classifier to rank omitted sentences by whether they are needed to interpret retained text and reinsert the top-ranked candidates without support annotations at inference. On HotpotQA with Qwen3-8B, this automatic restoration improves accuracy by 4.7 points while changing the compression ratio only from 0.30 to 0.31. Hard compressors should optimize both relevance and referential completeness.

Multi-agent reasoning (MAR) improves reasoning reliability through iterative solution exchange and refinement. Existing adaptive MAR methods typically learn routing decisions from query-level labels or trajectory-level returns, but such coarse supervision cannot accurately estimate the state-conditioned utility of individual operators in multi-step collaboration. We propose TreeCredit, a shared-prefix credit assignment framework for efficient adaptive MAR. Its core insight is to estimate operator utility through state-matched downstream comparisons, rather than directly attributing trajectory-level outcomes to preceding decisions. TreeCredit constructs shared-prefix collaboration trees by expanding candidate operators from the same intermediate state and assigns each state--operator pair a correctness-prioritized suffix credit based on the terminal correctness and cumulative additional cost of its complete continuation. These structured credits are converted into state-local operator preferences to train a lightweight pairwise state router, which dynamically selects the next admissible operator during inference. Experiments on six reasoning benchmarks show that TreeCredit modestly improves accuracy while substantially reducing inference cost, achieving a better accuracy--cost trade-off than representative MAR methods.

Visual token pruning reduces the inference cost of multimodal large language models, but a fixed token ratio is poorly matched to text-rich inputs. In OCR-centric tasks, decisive evidence can be a small number, label, or field whose relevance is specified by the question; indiscriminate pruning can erase that evidence while retaining visually salient but irrelevant regions. We present ET-Prune, a training-free framework that casts pruning as evidence allocation. It derives question-conditioned evidence from a decoder-side partial query-key block, safeguards text-like spatial regions, and converts evidence uncertainty and density into a sample-specific token floor. Three progressive middle-layer events then move the sequence toward this budget, retaining more tokens for diffuse or text-dense evidence and pruning concentrated evidence more aggressively. At the observed point estimates from one deterministic pass per configuration, ET-Prune leads or ties among pruned methods in all six backbone-benchmark comparisons at roughly half tokens. On OCRBench-v2, it leads the strongest pruned baselines by 1.80 and 0.68 percentage points on Qwen3-VL-8B and InternVL3.5-8B, respectively, while retaining about half of the visual tokens; on MMBench v1.1, it reaches 0.8467 circular exact-matching accuracy versus 0.8437 for Vanilla at 54.45% average visual-token retention. These results show a favorable observed quality-cost trade-off for evidence-aware dynamic budgeting in text-rich multimodal inference.

Diffusion vision-language models (dVLMs) iteratively denoise masked responses while conditioning each denoising step on visual evidence, making visual conditioning a substantial recurring inference cost. Unlike autoregressive decoding, diffusion generation repeatedly revisits the entire response as uncertainty evolves. Our analysis reveals that visual evidence demand is strongly step-dependent, motivating adaptive allocation across denoising steps. Existing inference acceleration methods operate through decoding-side strategies or visual token compression via pruning and merging, but do not explicitly treat visual evidence as a resource whose demand evolves across the diffusion process. Therefore, we present Denoising-Aware Visual Evidence Trajectory Allocation (DAVET), a training-free framework that allocates visual evidence according to the evolving generation state. Starting from a phase-conditioned evidence trajectory, the proposed allocation policy uses operation demand to set an evidence reserve whose allocation at each denoising step is modulated by trajectory risk. DAVET realizes the resulting budgets through a hierarchy of evidence views constructed from a single visual encoding, separating when and how much evidence is needed from how the evidence views are constructed. Evaluated on two representative dVLMs, LLaDA-V and LaViDa, across multiple visual-understanding benchmarks, DAVET achieves an average speedup of 1.55$\times$ with an average relative performance drop of 1.86%, showing that denoising-aware visual evidence allocation can reduce visual conditioning cost while largely preserving generation quality.

The continued development of LLMs toward persistent and adaptive intelligence increasingly requires long-term memory mechanisms that preserve and reuse information across interactions. Existing memory systems either compress and structure histories for efficient access or perform deep research over broader trajectories. The former lowers online cost but may omit temporal, causal, or cross-step dependencies, while the latter improves evidence coverage at substantial latency and inference cost. This raises a key question: can a memory system achieve strong answer quality while maintaining low online latency? We introduce Router-Mem, an evidence-conditioned progressive execution framework for long-horizon agent memory. Router-Mem first applies a shared low-cost retrieval prefix to obtain evidence. A lightweight sufficiency router then predicts whether the context supports early termination, which enable a single-token decision at inference time. It is trained with evidence-level supervision and rationale-conditioned representation distillation. When evidence is insufficient, Router-Mem reuses retrieval hits to expand memory blocks and perform deeper analysis and aggregation. Experiments on AMA-Bench and BEAM show that Router-Mem achieves 55.17% and 38.77% score while reducing average inference time by 27.3% and 25.5% compared with full memory execution.

Large vision--language models (VLMs) can reason step by step about complex visual scenes, but this open-ended, autoregressive chain-of-thought (CoT) approach is poorly suited to safety-critical, rule-governed settings such as industrial surveillance, where decisions must be bounded, deterministic, and low-latency. Because CoT inference cost scales jointly with reasoning length and the number of concurrent streams, it creates a throughput bottleneck that precludes the real-time, multistream monitoring required for industrial accountability. Here we present MonitorVLM-v2, a deployment-oriented framework that recasts VLM-based safety assessment as probabilistic inference over a finite regulatory decision space, compressing multimodal reasoning into single-step rule-ID predictions and reducing decoding from a variable-length sequence to a single token. We introduce symbolic policy optimization (SymPO), a novel contrastive policy optimization algorithm that sharpens decision boundaries within this finite symbolic space, together with an entropy-driven triage mechanism that routes uncertain predictions to human reviewers for expert confirmation. In a four-month prospective deployment across 10 concurrent camera feeds in an operational underground mining facility, MonitorVLM-v2 achieved a 19.45-fold increase in inference speed and identified 2.78 times as many confirmed violations as the site's routine manual inspection workflow, demonstrating the practical value of compressed symbolic decision-making for real-time, auditable industrial monitoring.

LLM orchestration is often assumed to improve reasoning by allocating additional inference-time computation, yet its gains may not justify its cost. Existing comparisons also frequently overlook differences in optimization effort, making it difficult to isolate the value of orchestration itself. We conduct a controlled evaluation of Self-Refine, Best-of-$N$, and Debate against task-only and chain-of-thought (CoT) single-call baselines across five LLM backbones and three domains: competitive programming, chess puzzles, and mathematics. For comparability, we optimize each method with GEPA under the same optimization budget and evaluate all methods on the same difficulty-stratified benchmark items. Orchestration yields moderate but benchmark-dependent gains: averaged across backbones within each benchmark, the largest improvement is 4.6 percentage points over optimized CoT inference and 4.5 points over task-only inference, while requiring approximately 2 to 4 times the mean total tokens of task-only inference. Human-derived difficulty is associated with lower absolute accuracy in all three benchmarks, but within-benchmark analyses do not indicate that orchestration effects increase with task difficulty. By contrast, exploratory mixed-effects analyses reveal strong interactions between orchestration method and backbone model across all three benchmarks, showing that orchestration effectiveness depends substantially on the underlying model. Our results suggest that orchestration decisions should be model-specific and account for whether moderate accuracy gains justify the additional inference cost. More broadly, evaluations of LLM orchestrations should control optimization effort and report model-specific accuracy--cost trade-offs rather than treating additional inference-time structure as uniformly beneficial.

As large language models serve more requests, cumulative inference cost is becoming increasingly important relative to one-time training cost. The two inference phases stress hardware differently: prompt prefill is parallel and typically compute-bound, whereas autoregressive decode is sequential and often memory-bandwidth-bound. Conventional width or depth scaling increases both costs together because every added layer is evaluated in both phases. We ask whether additional learned computation can instead be allocated to continuation prediction while preserving the prompt-wide primary computation and a single persistent key-value (KV) cache. We introduce the Dual-Flow Transformer. Its primary flow is a complete causal language model that processes the prompt and writes the KV cache. The auxiliary flow is omitted during prompt processing and activated only from the final prompt position onward, adding continuation-prediction computation without writing persistent state or influencing the primary flow. The two flows share major attention, MLP, and output matrices, while using separate token embeddings and lightweight coupling. Sharing weights and the primary cache also creates opportunities to reuse loaded weights and cached keys and values during grouped execution. Across matched-token comparisons, Dual-Flow achieves lower validation loss across architectures and data configurations. In MoE models, the separation makes primary and auxiliary expert fan-outs independent controls over prompt cost, continuation cost, and predictive quality. We study two regimes: increasing decode computation at fixed prefill expert computation, and reallocating a fixed decode expert budget between the two flows. These experiments expose a prefill-decode-quality trade-off and demonstrate the potential of phase-specific expert allocation.

Prompt compression shortens LLM input to reduce inference cost, yet existing methods score token importance through LM forward passes. It remains questionable whether such nuanced, costly token selection is necessary. Compression requires identifying informative content, a problem that linguistic research has long addressed through cues that can be operationalized as deterministic rules. We therefore ask: can \textbf{linguistic rules alone} serve as effective prompt compressors, without LM-based scoring at compression time? To address this, we conduct offline evolutionary search over lexical, syntactic, semantic, and discourse seeds to find competitive rule combinations. The resulting linguistic compressor requires no LM forward pass at deployment and uses only CPU-side processing for compression. We evaluate it with a dual-path protocol to balance compression quality and reconstruction fidelity. Across short passages, multi-document reasoning, and dialogue-memory QA datasets, evolved compressors achieve performance similar to that of recent advanced prompt-compression strategies. Performance is strongest under light-to-moderate compression and degrades as compression becomes more aggressive, while the Direct and Reconstruction paths exhibit distinct patterns. Evolutionary analysis reveals that effective compression fuses signals across linguistic levels and, as the compression ratio increases, rules shift from token pruning to sentence extraction.

Machine-learning services increasingly rely on public data, third-party providers, and outsourced training, creating opportunities for data-poisoning attacks that implant persistent malicious behavior while preserving benign utility. However, existing backdoor studies largely evaluate exact trigger reuse, training-exposed trigger diversity, or variations along predefined transformation axes. They therefore leave a critical blind spot: whether a backdoor learned from one training-time trigger can generalize to an inference-time trigger family absent from victim training. We formulate this problem as backdoor generalization under training--inference trigger shift and introduce Lilith, a black-box anchor-to-family framework. Using only disjoint surrogate resources, Lilith first induces a compact target-side vulnerability with a single training anchor, then constructs a bounded inference-only family that preserves the anchor-induced representation geometry. We characterize this mechanism through anchor clearance and family reach, deriving sufficient conditions for family-wise target preservation under local regularity and bounded surrogate--victim discrepancy. Experiments across datasets, architectures, poisoning rates, and defenses show that Lilith achieves high family-wise attack success with limited utility degradation and a small trigger generalization gap. Additional analyses show that family activation depends on representation alignment rather than the proposal mechanism, exposing a broader threat overlooked by exact-trigger evaluation.

Jailbreak defenses are essential for protecting large language models (LLMs), but they can also introduce secondary costs that weaken model utility. We present a systematic study of these defense trade-offs along three dimensions: performance impact, over-refusal on benign inputs, and inference cost. Rather than treating defenses as a single class, we organize them by operational strategy and examine how different strategies correlate with different side-effect profiles. Across state-of-the-art defense methods, widely used benchmark datasets, and representative open-source LLMs, we find that defenses rarely improve downstream capability, but instead vary in how they trade safety gains against usability and efficiency. In particular, rule-based defenses best preserve task performance, highly conservative self-reflective defenses often increase over-refusal, and multi-round defenses incur the largest runtime overhead. These results provide both a benchmark for evaluating defense side effects and practical guidance for selecting defenses under deployment constraints.

We examine how prompt tone affects both accuracy of the LLM answers and inference cost as reflected in output-token consumption. Experiments were performed to understand the trade-offs between accuracy and inference cost on a 570 Question MMLU dataset for LLM models prompted in seven different tones from sycophantic to threatening. Our results show that the output-token-length variation substantially exceeded accuracy variation across all models. Output-token consumption varied by up to 44.3% across tone conditions. We also analyzed the tradeoff between the accuracy of the answers and the average output token length in the reasoning process. For the ChatGPT models 4o and 5-nano, the rude tone is quite dominant. For the Gemini models 2.5 Flash and 2.5 Flash Lite, the rude and neutral tones are dominant on the Pareto-optimal frontier. We find that prompt tone influences not only answer quality but also the amount of billable inference resources consumed by modern LLMs.

Fully homomorphic encryption (FHE) provides strong cryptographic guarantees for private inference, but deploying transformer models under FHE remains prohibitively expensive. A key bottleneck is that non-linear operations such as softmax, normalization, and activation must be replaced with polynomial approximations compatible with the CKKS scheme, and the multiplicative depth consumed by these approximations dominates inference cost. Recent frameworks have advanced approximation techniques, yet all rely on manually configured approximation hyperparameters (e.g., number of iterations, polynomial degree), applied uniformly across all layers. While convenient, this uniform-configuration approach is overly rigid: different layers can tolerate different levels of approximation error without degrading predictive accuracy, and uniform configurations cannot exploit this variability to reduce latency. Allowing each layer to adopt its own configuration, however, causes the search space to explode with model depth, reaching roughly $10^{84}$ configurations for BERT/ViT (12 layers) and $10^{225}$ for LLaMA3 (32 layers), rendering manual exploration practically impossible. We present ATLAS, an automated framework that configures per-layer approximation settings by formulating the problem as a multi-objective optimization over latency and predictive accuracy. The resulting problem is inherently difficult: 1) competing objectives over a large decision space (120 or 320 variables for BERT/ViT or LLaMA3); 2) expensive evaluation, as each configuration takes 70-1,000 seconds even in cleartext; and 3) sparse optimization signals, as 35-50% of candidate configurations yield numerically invalid solutions. ATLAS addresses these challenges through a two-stage optimization strategy that progressively relaxes layer-wise constraints, combined with surrogate models to accelerate evaluation.

We introduce Graded Large Language Models (GLLMs), an algebraic framework that equips the representation space of a transformer with a grading and propagates the induced weighted scalar action through embeddings, self-attention, and the training objective. The construction extends the theory of graded neural networks and graded transformers to autoregressive language models while preserving expressive power, asymptotic computational complexity, and inference cost. The governing geometric picture is that of geometric invariant theory. The benefit of a grading is expressed by a Kempf--Ness functional on the grading torus; the grades that improve upon the uniform architecture form an open convex cone whose membership is decided by a Hilbert--Mumford-type criterion pairing a grade direction against two measurable profiles of the target and the data; the optimal grades are the coincidence point of two moment maps, given in closed form; and the ordinary transformer appears as a semistable isotropic point on the boundary of the cone: one member of a larger graded family rather than a distinguished optimum. Separately, for level-stratified targets we prove a minimax separation between the graded prior and its absence: over all estimators the risks of the graded and uniform target classes separate throughout an explicit window of sample sizes, by a factor that decays exponentially in the number of levels under geometric stratification. Both profiles are estimable offline, so the optimal grades solve a convex program certified before training begins. Because the grading is absorbed into the learned parameters after training, every GLLM compiles to a standard transformer of identical architecture and inference complexity.

Large language models (LLMs) provide strong reasoning capabilities but are expensive to serve at scale, whereas small language models (SLMs) are cheaper but less reliable on difficult problems. We introduce PyroDash, a cost-aware framework for token-level SLM-LLM collaborative inference. During generation, the SLM decides whether to request assistance by emitting a control token. A Collaborate Engine then sends the query and partial reasoning trace to a frozen LLM for completion through a single handoff. The policy is internalized in the SLM, requiring neither a separate router, LLM retraining, nor access to LLM logits. PyroDash trains the SLM in three stages: control-token embedding learning, offloading-oriented supervised fine-tuning, and cost-aware alignment with Group Relative Policy Optimization. Its reward balances answer accuracy against inference cost normalized by LLM-only inference. Across five mathematical reasoning benchmarks, PyroDash supports different accuracy-cost operating points. With $λ=0.05$, it achieves 64.04 percent average accuracy, 6.36 percentage points above the LLM-only baseline, while reducing cost by 20.4 percent. With $λ=0.6$, it achieves 54.55 percent accuracy with a 1.90 percent LLM token ratio and 0.012 LLM calls per example, reducing total cost from USD 49.36 to USD 1.78. These results show that learned token-level handoffs can reduce LLM use while preserving strong reasoning performance.

Tracking objects through state transformations is essential for understanding real-world dynamics. However, existing methods are computationally expensive. TubeletGraph recently showed impressive capabilities, but its inference cost (~$4.4$ seconds per object-frame on VOST) precludes any real-time deployment possibilities. We observe that TubeletGraph's overhead arises from building a spatiotemporal partition of the input video: (1) entity segmentation is computed densely for every frame regardless of whether a transformation occurs, and (2) every entity in the scene is tracked, scaling cost with scene complexity rather than the number of transformations of interest. To address both, we propose FluxGraph, a reactive variant that uses SAM2's internal multi-mask disagreement as a lightweight trigger for transformation detection, and removes the need for tracking all entities in the given video. FluxGraph is ~$3.3\times$ faster than TubeletGraph on VOST while improving tracking performance and preserving state graph quality. Furthermore, we also observe consistent speedups of $3.7-10.7\times$ across VSCOS, M$^3$-VOS, and DAVIS17 while maintaining performance. Code is publicly available at https://github.com/YihongSun/FluxGraph.

Scaling LLM-based applications to millions of users is bottlenecked by the inference cost and latency of modern foundation models. A natural fix is to cluster the inputs and call the LLM only on cluster representatives, letting other members inherit the output -- but this is only safe if each member is measurably close to its representative. Existing clustering methods do not offer such per-sample quality control at scale: none jointly guarantee a minimal within-cluster similarity, exact matching of categorical attributes, and scalability to tens of millions of samples. We propose a two-stage algorithm that generates initial clusters with Mini-batch K-Means, then greedily selects representatives within each initial cluster -- a step equivalent to the Johnson-Chvatal heuristic for Set Cover over alpha-balls in embedding space. The algorithm enforces the similarity and attribute guardrails exactly by construction, and runs in $O(nd + n^2 d/K)$ time and $O(nd + n^2/K^2)$ memory for $n$ samples, feature dimension $d$, and $K$ initial clusters -- linear in $n$ when $K$ grows proportionally with $n$. We provide benchmarks against common clustering methods on internal and public datasets: our method not only delivers per-sample guardrails but also runs 10-1000x faster and scales to data sizes where most standard methods become intractable. Deployed on 38 million customers for a persona-based recommender, the clustering method cut downstream cost and latency by 50-fold while preserving personalization and unblocked the production launch.

Large language models (LLMs) are increasingly adopted for network automation, yet their output quality and inference cost can vary substantially across LLM families. We present HuGLEN, a stepwise evaluation pipeline that uses an LLM-as-a-judge together with a small set of expert ratings to enable scalable and reproducible comparison of candidate LLMs, and to rank them using a quality efficiency score (QES). We demonstrate HuGLEN for translating outputs from an explainable artificial intelligence (XAI) model for the optical network quality of transmission (QoT) estimation task into operator-friendly explanations. Our results show that a medium-sized LLM (12B parameters) achieves the highest QES, indicating the best trade-off between explanation quality and efficiency. Overall, HuGLEN reduces the human-labeling burden while supporting consistent model selection for operator-facing automation tasks.

Language agents depend on memory across interactions. However, the limited context windows of large language models (LLMs) and their inference costs constrain how much memory can be used at once. Existing systems mainly follow two strategies: memory retention and memory consolidation. Retention keeps raw records and preserves exact details, but relevant evidence may not fit under a tight budget; consolidation compresses and combines records, improving coverage per token but risking the loss of query-critical details. Neither strategy is universally preferable. This raises two central questions: when should consolidation replace retention, and which operator -- Merge, Abstract, or Rewrite -- should be selected? We formalize this decision by decomposing each operator's utility into a coverage effect on evidence omitted by retention and a signed replacement effect on raw evidence that already fits. Their balance explains why the preferred action changes with relative budget pressure. We implement this mechanism with Offline Abstraction-Safety (OAS), a lightweight learner that estimates action utilities from pre-generation features with held-out harm calibration. The public LongMemEval and LoCoMo benchmarks show the same budget-dependent pattern. On LongMemEval, consolidation improves absolute accuracy by up to 48% under tight budgets, whereas retention is preferable under loose budgets; LoCoMo replicates this crossover at a smaller budget, consistent with its shorter evidence. On both datasets, cross-note abstraction and merging generally outperform local rewriting when compression is necessary.

When does a committed intermediate stage in an LLM reasoning pipeline earn its cost? Constrained Path Reasoning (CPR) pairs a source-aware path hypothesis with stage-level accounting. Search generates provisional states; trusted or validated invariants can constrain hard, while other proposals remain soft and revisable. CPR predicts that task-compatible commitments can factor transitions, concentrate candidate mass, induce regularity, and expose feedback when their gains exceed propagated error and execution cost. The formalism covers discrete commitments and continuous flows and measures effective branching, endpoint concentration, and cost per usable output. Across 1,180 generated QCQPs and 40 engineered degenerate polynomial instances (2,140 endpoints), residual triage recovers 63.0% of repair-all's additional feasible yield with 17.7% of its attempts. Fixed-LLM accounting (270 unique calls shared across nested arms) finds usable yield of 41.1% direct, 90.0% after formalization and deterministic execution, 20.0% after one-shot convexification, and 21.1% for the full path. In 120 paired-condition calls, a two-action rollback rule reaches 90% usable yield versus 36.7% for the feedback-conditioned selector. Two endpoint probes separate source from validation: a 72-output cross-trajectory transplant reduces entropy and acceptable mass; a 24-output same-call self-proposal pilot gives unchanged two-repeat collision entropy, 25.0% versus 8.3% usable yield, and 1/8 deterministically confirmed endpoint checks. Model-generated states supply hypotheses; trusted execution earns constraint strength.

Calibration of grey-box simulation models is a constrained optimization problem in which model evaluations are expensive, the parameter space can be high-dimensional, and the search must respect plausibility constraints. Although the simulation code is fully available to the analyst, the joint effect of multiple parameters remains difficult to predict analytically. Classical optimizers such as Nelder--Mead (NM) are simple to deploy but sample-inefficient, particularly under constraints. Modern Bayesian Optimization methods achieve competitive solutions with far fewer evaluations but require non-trivial modeling machinery for constraint handling. We introduce an agentic calibration method in which a large language model acts as the optimizer, with constraints incorporated as a plain-language section of the system prompt. We evaluate the agentic method, NM, and Bayesian Optimization (BO) on an anal cancer simulation model under both unconstrained and clinically constrained calibration. Under unconstrained calibration, the agentic method achieves substantially lower best error than BO and NM, while requiring fewer model evaluations. Under constrained calibration, the agentic method reaches comparable error levels and both outperform NM. These results are obtained at the cost of increased inference time per iteration. Agentic calibration achieves competitive performance with substantially fewer model evaluations, and constraint handling is essentially free at the modeller-facing interface through simple textual specifications rather than additional modelling machinery. The main trade-off lies in increased per-iteration inference cost, making the approach particularly suitable when simulation time dominates. Beyond performance, the per-iteration rationale makes the search auditable and explainable, so its decisions can be scrutinised and justified to third parties.

LLM-based agents are increasingly deployed in multi-agent environments whose incentives can shape their behavior. We introduce The Energy Society, a minimal survival economy for studying how competitive and cooperative incentives affect emergent behavior when inference cost is directly tied to survival: Agents spend energy based on model size when generating tokens, regain energy by completing jobs or receiving donations, and deactivate if their energy reaches zero. We compare competitive and cooperative objectives against a baseline setting and several control variants. Across experiments, larger models consistently consume the most energy and spend more energy than they gain, even in those settings where token cost is not size-dependent. Cooperative incentives substantially alter behavior: agents donate to reactivate others, sometimes at the cost of their own survival, and job allocation changes. Ablations reveal that allowing agents to recommend actions to each other supports coordination and ambitious job selection, while memory helps agents calibrate risk from past outcomes. Agents rarely choose direct sabotage, but show more subtle signs of self-serving behavior in the competitive setting. The Energy Society is a compact testbed for studying the interaction between token costs and group incentives under a survival pressure. Source code is available at https://github.com/LucasBergholdt/EnergySociety

Mixture-of-Experts (MoE) models scale capacity without proportional compute cost and have become a key architecture for frontier large language models (LLMs). Yet domain-specific post-training inherits an expert pool shaped by mixed-domain pre-training: a substantial subset of experts contributes little on the target domain, and standard supervised fine-tuning (SFT) leaves the composition of this pool unchanged. We propose a simple, budget-preserving pipeline that realigns the expert pool to the target domain before fine-tuning. Given a target domain, we (1) prune the experts with lowest domain-aligned saliency, (2) regrow the expert pool to its original size through perturbation-based expert expansion, and (3) apply standard SFT. The resulting model preserves the original expert count, parameter count, and inference cost. With a single frozen recipe and no per-domain hyperparameter tuning, UMoE consistently improves over direct sft across two MoE architectures (Qwen3-30B-A3B and Qwen3.5-35B-A3B), five domains (math, code, science, tool-use, and agentic coding), and 12 benchmarks. Representative improvements are 3.4 points in math average accuracy, 6.0 points on SWE-bench Verified. On a strong in-house math corpus, direct sft already surpasses Qwen3-30B-A3B-Thinking (82.81 vs.\ 81.06), yet UMoE further raises the average to 84.17, an additional 1.36 points, demonstrating robustness to a substantially stronger SFT regime. Data-scaling experiments further show that the gain persists as training data grows. Analysis reveals that the direct-SFT model allocates substantial routed-expert compute to a low-saliency subset that can be removed post hoc with little average degradation; UMoE turns this redundant capacity into useful domain capacity and achieves lower training loss, with gains spanning all difficulty levels in downstream evaluation.

Reliable visual data association is fundamental to visual SLAM (V-SLAM), as it directly determines the quality of the camera pose estimation and map consistency. However, the handcrafted descriptors used by most mature real-time systems degrade under illumination and viewpoint changes, while learning-based front-ends that address this weakness typically require replacing the extraction-and-matching pipeline and introduce substantial computational overhead. Descriptor enhancement offers a compromise by refining existing descriptors within their original format, yet current methods rely on simplified attention mechanisms whose limited contextual modeling constrains the achievable matching quality. To resolve this trade-off between contextual expressiveness and efficiency, we propose Desc++, a lightweight enhancement module that jointly encodes descriptor representations and keypoint geometry and aggregates spatial context through a hybrid architecture that combines order-agnostic global attention with geometry-aware sequential modeling in linear time. The enhanced descriptors retain their original dimensionality and matching interface, enabling integration into deployed V-SLAM systems without modifying the pipeline. Experiments across descriptor matching, correspondence analysis, and system-level benchmarks with four different V-SLAM systems demonstrate that Desc++ improves matching accuracy over the state-of-the-art enhancement method, translates these gains into more accurate and stable trajectory estimation, and achieves a favorable balance between accuracy and efficiency for practical integration into existing real-time V-SLAM pipelines.

Multi-agent ensembling multiplies active parameters and inference cost without answering three basic questions: which agents to consult, how deeply a query should traverse a hierarchy of agents, and when inter-agent communication is worth its cost. We present GRADE (Gated Routing and Adaptive Depth for Efficient Reasoning), a hierarchical multi-agent system in which four lightweight learned gates jointly govern agent selection, hierarchy depth, inter-agent communication, and branch pruning. Training uses CoGRPO (Collaborative Group-Relative Policy Optimization), a novel critic-free recipe that adapts GRPO to multi-agent hierarchies and assigns a shared advantage signal to every gate and agent that participated in a rollout. Agent models are drawn from a hot-swappable Expert Registry; per-agent calibration maps allow experts to be replaced at inference time without retraining. At $\sim$17B average active parameters, GRADE outperforms all baselines on GSM8K, MMLUPro, and GPQA, surpassing the strongest baseline by 4.8 points on MMLUPro at half the active compute. On AIME-2025, where model depth dominates, GRADE remains competitive to existing frameworks. Ablations isolate the hierarchy and masked cross-attention as the largest contributors to accuracy, and show that per-agent calibration is necessary for safe hot-swapping.

Uncertainty quantification for visual language models (VLMs) conventionally targets the answer token distribution. We provide the first three-family empirical characterisation of answer entropy behaviour in thinking-mode VLMs. Running four models on identical POPE adversarial samples, we find three qualitatively distinct patterns: Qwen3-VL-8B-Thinking shows complete collapse (ans H AUROC = 0.492); GLM-4.1V-9B-Thinking shows no collapse (0.716); and InternVL3-8B shows selective thinking (chains on only 50% of samples, ans H = 0.675 full / 0.602 thinking-only). Across all three thinking-mode models, thinking chain entropy outperforms answer entropy on the subset where chains are generated (0.647, 0.759, 0.608 vs. 0.492, 0.716, 0.602 respectively), suggesting chain signals are the more reliable predictor whenever chains are present. This holds strongly for Qwen and GLM, but with only marginal and statistically unreliable advantage for InternVL3 (n_FP = 17). A 300-sample VQAv2 pilot confirms chain entropy (0.680) outperforms answer entropy (0.595) on VQAv2 questions, with the gap largest for free-form answers (0.733 vs. 0.467). On harder reasoning tasks (HallusionBench) both Qwen models show moderate signal (approx. 0.64), consistent with incomplete pre-commitment on difficult questions. We additionally document structured abstention affecting 12-22% of queries with asymmetry toward absent-object queries, and a practical abstention gate raising accuracy from 71.0% to 93.8% at 62.7% coverage with no additional inference cost.

Large language models hallucinate most about entities they have never seen. We ask whether a model's activations betray entity familiarity before a single answer token is generated, and whether that signal predicts the factual reliability of the answers. On four Polish Bielik models (1.5B-11B parameters), we probe four entity domains (athletes, cities, writers, musicians), each with 42 well-known, 42 obscure-but-real, and 42 fabricated entities addressed by a one-sentence question (504 prompts per model). Two unsupervised, single-forward-pass dispersion measures over post-SwiGLU MLP activations, inverse participation ratio and spectral entropy, separate known from fabricated entities at AUROC 0.95-1.00 across all domains and scales; a supervised linear probe reaches 0.99-1.00. Both clear selection-aware permutation floors of about 0.70-0.74 (empirical p<=1e-3), survive held-out layer selection (0.93-0.99), and persist on real names (known vs. obscure-but-real: 0.96-1.00). The signal transfers across entity types (mean off-diagonal AUROC 0.92-0.99); a matched-template counterfactual shows the only large drops are template-caused, not entity-type effects, and the signal is diffuse across heads. This representational signal is already at ceiling at 1.5B, whereas behavioral factual reliability scales sharply: 0, 2, 10, and 19 of 42 known athletes are answered fully correctly by the 1.5B, 4.5B, 7B, and 11B models under a strict judge. Within known entities, separating correct from hallucinated answers is much harder (probe 0.93; dispersion no better than a first-token-entropy baseline). A five-sample semantic-entropy baseline reaches only 0.71-0.83 at 5x the inference cost. Despite this internal awareness, the models almost never abstain: an audit of 2,520 answers finds 2 refusals and 1 hedge. Entity familiarity and factual reliability are distinct phenomena on different scaling curves.

Diffusion models generate high-quality images, but their inference cost comes from two sources: large denoising networks and repeated denoising steps. Existing compression pipelines usually attack these costs separately. Pruning reduces the network, but most pruning methods still rely on a long post-pruning retraining stage to recover a many-step sampler. Step distillation reduces the number of denoising steps, but it usually assumes a student that can already follow the teacher well enough to receive useful distillation gradients. This paper asks whether post-pruning retraining can be replaced by step distillation. We find that the direct replacement fails: after pruning an EDM2-XS teacher, starting SiDA from the pruned checkpoint produces unusable samples. We introduce a short teacher-alignment repair stage as a bridge between pruning and step distillation. The bridge matches the pruned generator to the teacher on noisy real-image latents, then hands the repaired checkpoint to one-step distillation. On ImageNet-512, the original EDM2-XS baseline uses 124.713M parameters and 63 network evaluations, reaching an FID of 3.53. With a suitable distillation objective, our 20% pruned one-step generator uses 98.826M parameters and one network evaluation, reaching an FID of 3.12. With 30% pruning, the model uses 88.029M parameters and one network evaluation, with an FID of 4.26.

Modern automated audio captioning systems pair a frozen audio encoder with a large language model (LLM) via a trainable projector, incurring the encoder's inference cost and bottlenecking the model through its fixed acoustic features. We present CARD, an encoder-free audio captioning model that removes the encoder at inference: a 13.2M projector feeds a frozen LLM with merged LoRA adapters, while the teacher used to train it is discarded. CARD distills a pretrained audio teacher (CLAP-HTSAT) into the model, but rather than injecting it into the LLM alone, it routes the teacher's representations across components: perceptual stages to the projector and semantic stages to the LLM. This placement improves CIDEr-D by +12.18 over an LLM-only distilled model on AudioCaps and by +5.21 on Clotho, reaching 55.4 against a 66.4 encoder-kept upper bound with no encoder at inference, showing that where a teacher's knowledge is placed matters as much as its presence.

Large language models (LLMs) exhibit strong reasoning and world-knowledge capabilities, yet often struggle to gather information effectively across the multi-turn interactions required in sequential decision-making settings. We introduce Amortised Sequential Information Gathering (ASIG), a fine-tuning approach that amortises Bayesian Experimental Design (BED) into LLM policies via a multi-turn extension of Group Relative Policy Optimisation with an Expected Information Gain reward. Evaluated on the 20 Questions task, ASIG more than doubles the success rate of the 7B base model and reduces inference cost by over $25\times$ relative to BED-LLM, a competitive inference-time baseline. Applied to MediQ, a medical diagnosis benchmark unseen during training, ASIG improves information-seeking performance at the 7B scale, suggesting that the learned strategies can transfer out of distribution. Our findings show that amortising BED into LLM policies provides an effective and computationally efficient approach to sequential information gathering.

LLM-based coding agents solve software-engineering tasks through iterative interactions with development environments, where returned observations accumulate in the context and become a major source of inference cost. Observation compression reduces this cost by shortening observations before they are appended to the context. However, existing methods still exhibit an unsatisfactory efficiency-effectiveness trade-off, as they do not explicitly model how compression affects the agent's subsequent behavior. This paper proposes CoACT, an action-preserving observation compression method for coding agents. CoACT is built on next-action preservation (NAP), which requires a compressed observation to induce the same next action as the raw observation. By checking the agent's immediate next action, NAP provides a practical signal for whether a compression preserves the information needed for continued task solving. During training, a teacher model first generates multiple compressed candidates of each observation. CoACT then uses an action-preservation reward based on NAP to filter out candidates that would change the agent's next action, and uses a length-reduction reward to choose compact candidates as supervision for a lightweight compressor. Experiments on SWE-bench Verified with three agentic models show that CoACT reduces average total token consumption by 33.0% while maintaining task-solving effectiveness close to the uncompressed agent.

Agentic Video Question Answering (VideoQA) systems produce answers through adaptive reasoning and tool-use trajectories, yet standard practice evaluates each system once and estimates uncertainty only across questions. This leaves a basic question untested: would the measured method effect survive if the evaluation were run again? We show that it need not. Using Static-SAGE and Dynamic-SAGE as a controlled case study, we repeat the paired comparison twice on identical SAGE-Bench question-video pairs, holding configuration, tool library, and scoring protocol fixed. In the first execution, Dynamic-SAGE outperforms Static-SAGE by +7.33 accuracy points; in the second, the effect reverses to -4.05. Both are individually significant under paired analysis, supporting opposite conclusions. The change in the paired effect between executions is highly significant and far larger than within-execution uncertainty. The reversal is consistent across question format, modality, difficulty, and video duration, and both evaluation arms move significantly. Motivated by this failure, we introduce REPAIR (REpeated PAired Inference Reliability), a protocol that repeats the paired comparison and tests whether the method effect changes across executions, separating directional reproducibility from effect-size stability. Applied across accuracy and execution metrics, REPAIR exposes three behaviors-directional reversal, magnitude shift, and effect attenuationand shows that reductions in reasoning turns and visible tool calls do not imply reproducible reductions in primitive computation or latency. The execution-level movement is comparable to, and here larger than, median gain reported by recent agentic VideoQA systems, contextualizing its magnitude without implying those systems are unstable. Significance within a single agentic execution is insufficient evidence that a reported method effect is reproducible.

Thermodynamic computing devices based on the Ising model show great promise for low-power AI inference and edge computing, but scalable methods for training large models for such hardware remain limited. Prior theory shows that the time-averaged behavior of high-temperature Gibbs-sampled Ising systems can implement feed-forward neural inference. We turn this theoretical correspondence into a scalable and purely backpropagation-based algorithm for training deep convolutional networks for thermodynamic inference on Ising machine hardware. Our image classification models achieve accuracies of 94.9% on CIFAR-10 and 76.0% on CIFAR-100 under binary Gibbs sampling. We then develop and experimentally validate a mathematical theory relating inference cost to accuracy and controlling autocorrelation times. Subsequently, we calculate asymptotic results showing that inference cost is bounded by a well-controlled tradeoff with performance and exhibit algorithms for computing optimal inference schedules. Finally, we discuss implications for hardware development and the future of high-temperature thermodynamic AI models.

AI models are rapidly improving at autonomous coding, as shown by benchmark progress and one-off demonstrations such as AI implementing a C compiler. However, existing coding benchmarks tend to focus on shorter tasks, and one-off demonstrations are hard to compare systematically because they often have some human guidance, and are not standardized or repeated across models. To address these challenges, we introduce MirrorCode, a long-horizon coding benchmark based on reimplementing entire software projects. In MirrorCode, AI agents must replicate the functionalities of an existing program, without access to its source code. AI solutions must match the original program's output exactly on end-to-end tests, including held-out tests. MirrorCode's 25 target programs span different areas of computing: Unix utilities, data serialization and query tools, bioinformatics, interpreters, static analysis, cryptography, and compression. Existing AI models can already reimplement complex software, with the strongest model scoring 56% across the benchmark. For example, AI can reimplement gotree, a 16,000-line bioinformatics toolkit - a task that we believe would take weeks for a human engineer. However, studying the frontier of performance requires a larger inference budget than typical benchmarks, for example, $2,600 over 19 days for a single attempt on a large task. We show that AI agents can already complete long-horizon software engineering tasks, especially when requirements are precisely specified. More broadly, our work suggests AI will have transformative effects on software engineering, as autonomous agents continue to improve.

Recent advances in speech separation (SS) have led to compact front-end models with small parameter sizes, yet their high computational cost remains a major barrier for deployment on edge devices. To address this, we propose TF-MoE, a sparse Mixture-of-Experts (MoE) framework that enhances model capacity with almost no increase in inference cost. Our method introduces dynamic expert specialization in time and frequency dimensions through alternating time-wise and frequency-wise MoE modules, each dynamically selecting experts per frame or mel band. Built upon a mel-band-splitting Conformer backbone, TF-MoE achieves strong performance on SS tasks under low-compute settings. Experimental results demonstrate that TF-MoE consistently improves separation performance under computation cost constraints, outperforming BSRNN by +3.8 dB SDR on Libri2Mix with comparable 4.1 GMACs/s inference cost. This positions TF-MoE as a promising candidate for edge-device deployment.

Quantization is widely used to reduce the inference cost of large language models, but its effect on reasoning models is not fully captured by final-answer accuracy or per-token latency. We show that low-bit post-training quantization can introduce a hidden test-time compute cost: quantized reasoning models often generate longer chains of thought even when they still answer correctly. Across mathematical reasoning, code generation, scientific question answering, and agentic tool-use benchmarks, we find that INT4/INT3 quantization can preserve accuracy but increase reasoning-token usage, offsetting the expected per-token speedup. To measure this effect, we introduce the CoT Token Inflation Ratio, which compares reasoning length between quantized and full-precision models averaged across all evaluation benchmarks. We further show that token inflation is accompanied by behavioral changes in the reasoning trace, including more intermediate steps and greater semantic repetition. These changes translate into measurable end-to-end real-world serving penalties. Finally, we evaluate mitigation strategies and find that prompting and decoding-time sampling offer inconsistent accuracy-length trade-offs, while quantization-aware training shows more promise in reducing both accuracy degradation and token inflation. Our results suggest that reasoning-token usage should be reported alongside accuracy when evaluating quantized reasoning models.

"Talk short. Drop grammar. Save token." This caveman style is widely promoted as a way to cut inference cost, but whether it actually saves anything depends on which channel (the user's prompt or the model's response) is being compressed. We present Cavewoman, a two-channel evaluation protocol that scores every generation on task accuracy, realized per-item cost, and reference-text agreement against the model's unconstrained reference. We evaluate eight models on five datasets at five reduction levels, with both channels measured on the same items. Output compression cuts realized cost on most API models (1.4-2.4x per model, up to 3x in the best case) and on all four open-weight models under public-tier pricing. Input compression has the opposite effect, a strict lose-lose: it raises net cost rather than lowering it (~1.15x on the five-benchmark mean, up to 1.8x on the worst dataset and 2.7x under stronger compression), because models compensate with longer responses even as accuracy collapses. Under the same setting, surface text diverges from the unconstrained reference: on the non-reasoning models, roughly half of all generations are correct yet their surface text no longer entails the model's own unconstrained baseline generation. The divergence survives length-controlled re-scoring, multiple-comparisons correction, and replication under complementary semantic measures. Code and data are available at https://github.com/danielle34/cavewoman.

A modern model release reports scores on 40+ benchmarks and the same evaluations were run many more times before it: to track training progress, compare design choices, and select the checkpoint for the release. But do we need to run every eval? We compile a public score matrix of 84 frontier models on 133 benchmarks (2,604 cells, 23.3% filled) and find it is approximately rank-2: a model's scores across all 133 benchmarks are largely determined by just two numbers. We confirm this in two ways: scores hidden from the matrix are best recovered using two factors, and two factors already explain over 90% of the variation among models on the benchmarks they share. Building on this, we design BenchPress: a logit-space rank-2 matrix completion method that recovers held-out scores to within 4.6 points, and a confidence layer that says when each prediction can be trusted. Using BenchPress, we find a subset of five benchmarks {GPQA-D, HLE, Codeforces, MMLU-Pro, ARC-AGI-1} that can recover the rest of a model's public scorecard to within 3.93 points. For a tighter inference budget, a cheaper set {GPQA-D, MMLU-Pro, Aider Polyglot, MATH-500, AIME 2026} can predict a model's evals to within 4.55. We release the score matrix, the BenchPress code, and an interactive tool that predicts any model's score on any benchmark.

Large language models (LLMs) are increasingly deployed as black-box classifiers in pipelines that automate confident decisions and route uncertain ones to human review. Such selective prediction needs a confidence score that an operator can threshold at a chosen risk level. Prior work asks whether LLM confidence is well calibrated or well ranked; we ask a complementary, deployment-oriented question that has been largely overlooked: at what resolution can the score be thresholded? We call the answer the score granularity gap. Through a controlled comparison of seven ways to build a confidence score, from a single verbalized number, to token probabilities, to querying the model many times and combining the answers, across 25 model-dataset pairs (9 LLMs, 3 benchmarks), we find that single-shot verbalized confidence, once correctly converted to a class probability, ranks cases surprisingly well, yet takes only a handful of distinct values. It therefore offers an operator only a few coarse thresholds, no matter how well it ranks. We show which constructions widen this gap, at what inference cost, and with what effect on ranking, notably that multi-query aggregation helps weak models but can degrade already-strong ones. We translate these trade-offs into concrete deployment guidance.

Scientific discovery can be formulated as an iterative search process over the space of hypotheses and experiments. Contemporary methods navigate this space using heuristics such as MCTS. These algorithms conflate the merit of a hypothesis with the quality of its experimental execution. A promising hypothesis with preliminary execution is therefore ranked below a modest hypothesis whose execution is refined. Moreover, prior methods prune the search logs as the search progresses because the accumulated history outgrows the context window. We propose Agentic Reasoning for Tree Search (ARTS), where we deploy a reasoning language model to navigate this space. The model inspects prior execution logs, diagnoses whether earlier failures arose from faulty implementations or bad hypotheses, and selects the hypothesis to build on next. To mitigate challenges with context length, ARTS uses test-time training to instill the knowledge of search tree in the model weights. Across 22 tasks from MLGym and MLEBench, we show that ARTS outperforms leading algorithms, with over 15.3% relative improvement in the normalized score. With test-time training we show that a Qwen3-4B agent can match performance with closed-source frontier models like Gemini-3 Pro and GPT o3-reasoning with upto 5x lower inference cost. We further observe that on partially observable RL tasks, the test-time trained Qwen3-4B scientist surpasses ARTS with the o3 scientist by rediscovering the human-best recurrent-memory solution that heuristic methods prune away.

A finite agent, a machine's digital twin or any bounded reasoner, infers a fixed and noisy world through finite sensors, so its coherent output is a belief: a probability density over states (the Bayes posterior). Such an agent stops short of certainty, and revising a belief carries a cost. We propose a framework for belief costs based on optimal transport, motivated by these facts. We pose two postulates. P0 (the arena): a revision cost is a scalar price on optimal transport, so beliefs live in Wasserstein space. P1 (uniform pricing): one nat of knowledge costs the same metric length everywhere, the eikonal condition. Among conceivable pricing rules we study this one. Under P0 and P1 the cost metric is optimal transport conformally reweighted by Fisher information, $\tilde g_{e,U}=2(e+U)\,g_{W_2}$, and the Fisher family is a characterization: among continuous reliefs, uniform pricing is equivalent to $U=cJ$. Two consequences follow on the conformal class. Certainty sits at infinite cost-distance once the relief dominates the Fisher information, so a well-posed inference has a cost floor diverging at certainty (necessity conjectural beyond power laws). On location-scale leaves the geometry is hyperbolic, and the Stam bound places the Gaussian as the most curved one (at $e=0$). The results are geometric, in nats, and hold up to units: a change of cost unit rescales all distances and preserves every conclusion (boundary, eikonal family, hyperbolicity, Gaussian extremum), a gauge theorem; a global change of state units at $e=0$ is an isometry; the content lies in signs, rankings and ratios. Via Landauer (one nat worth $k_BT$) the cost floor becomes an energy floor: revising toward certainty would demand unbounded energy. Physics anchors the unit and enters no theorem. Removing either postulate leaves the selection open.

Mixture-of-Experts (MoE) language models are often described as ideal for resource-constrained inference. Each token activates only a small subset of experts, so the per-token compute cost, in floating-point operations (FLOPs), resembles that of a much smaller dense model. Whether that FLOP advantage survives in practice is far less clear. We ask whether MoE models actually run faster and cheaper than comparable dense models on consumer-grade and edge hardware. We benchmark OLMoE-1B-7B (1.3 B active of 6.9 B total) against three dense baselines on an Apple M2 Pro and an NVIDIA Jetson Orin Nano 8 GB through llama$.$cpp, measuring throughput, memory, and on-device energy. The answer is device-dependent: OLMoE's active-parameter advantage is only partly realised on the laptop (~10% behind the same-active Llama-3.2-1B) and erodes on the edge device (~31% behind, at 2.1$\times$ the energy per token, with peak memory at the 8 GB ceiling). Patching llama$.$cpp to time the decode graph node-by-node shows routing accounts for under 9% of MoE-block compute on the cleaner edge backend, so the gap reflects total-parameter memory footprint, expert dispatch, and KV-cache pressure rather than routing. The implication is that on bandwidth-bound edge hardware, inference cost tracks total parameters, not active ones, and sparse activation does not buy back what the device is constrained on. These findings are bounded to one MoE model at this parameter scale and two devices, and we release the full measurement harness and per-run data.