In vision-language models (VLMs), a few sink tokens attract a disproportionate share of attention. While previous studies have examined why these tokens attract attention, we investigate how the sink-key direction affects attention to image and question tokens. Across four probed VLMs, boosting the magnitude of positive sink-aligned components increases attention on average, whereas boosting negative components decreases it; these responses persist through later layers. This identifies the native sink key as a reference for controlling attention beyond its own position. Motivated by this finding, we propose SHRINK, a training-free method that partially reduces sink-aligned components across visible keys during generation. On nine benchmarks under sampled reasoning, it improves mean accuracy by 5.1 and 2.7 percentage points on LLaVA-OneVision-7B and Qwen2.5-VL-3B, with smaller gains on the other two main models, Qwen2.5-VL-7B and Gemma-3-4B. Beyond SHRINK, rescaling sink-aligned key components can replace the attention operations of two existing attention-steering methods, GIFT and ACG, with comparable accuracy on hallucination benchmarks.
Automated agent design searches for executable architectures that improve LLM-based agent performance, with evolutionary methods using language-model crossover to recombine promising mechanisms for task decomposition, model coordination, and verification. This search, however, is typically guided by aggregate fitness alone, which hides the instance-level successes and failures behind each score. We introduce Execution Guided Genetic Recombination for Agent Design (GRAFT), which uses aggregate fitness for selection and survival and parent-specific execution evidence to guide semantic crossover with integrated mutation. Its feedback harness reuses bounded samples of successful and failed runs from fitness evaluation, withholding reference answers and requiring no additional task executions. We evaluate GRAFT across six reasoning and programming settings using open-weight optimizer and executor models. It achieves a mean test accuracy of 82.0%, exceeding the strongest baseline by 9.8% points. Controlled ablations show that adding execution evidence improves test accuracy with initialization, search rules, repair policies, and evaluation budgets held fixed. Our results support the value of parent-specific execution evidence for evolutionary agent design.
Large vision-language models (LVLMs) and language models (LLMs) often exhibit attention sinks, whereby particular positions, such as the beginning-of-sequence (BOS) token, receive a disproportionate share of attention. Under standard causal masking, however, the BOS query attends only to itself, so the representation that later tokens repeatedly access carries no information about the input. We introduce BOS Summary, a training-free inference-time method that repurposes this persistent pathway to carry input context. At a single transformer layer during prefill, BOS Summary relaxes the causal constraint for the BOS query row alone, letting BOS aggregate all visual and textual input tokens into a self-derived context representation while every other token retains standard causal attention. The enriched representation then remains accessible to the model throughout decoding, with no additional parameters, auxiliary modules, or decoding-time overhead. Across multiple VLM and LLM families and eight multimodal and text-only benchmarks, BOS Summary improves over direct and Chain-of-Thought prompting and outperforms recent sink-intervention methods on multi-step reasoning tasks. A mechanistic analysis shows that making BOS input-specific also weakens its sink dominance and redistributes attention toward content-bearing prompt tokens, indicating that changing what a sink carries reshapes how later layers allocate attention.
Simulating how physical systems respond to changing conditions is central to control, design, and scenario studies. To our knowledge, prior PDE foundation models have not been trained or evaluated on changes within trajectories. Our counterfactual test measures how far predictions move toward a changed solver branch from the same state. Models trained on fixed conditions follow changes only partly. A PDE foundation model given a condition input after pretraining barely responds. PhysDial adapts a pretrained video diffusion transformer to eight two-dimensional PDE families. Coefficients enter with the noise level, condition fields as reference frames, and equation text as a prompt. In Kolmogorov flow, switches within trajectories account for about 5% of the training samples. At matched compute, they halve the error after a change in forcing wavenumber or amplitude. A few trajectories at new amplitudes more than halve extrapolation error. Without physics pretraining, PhysDial matches or exceeds PDE foundation models under training conditions on two turbulent flows. At step 64 on the Kolmogorov test split, PhysDial has 1.07 times the solver's small-scale energy and Walrus 7.1 times. Our experiments suggest that a learned simulator uses a condition only as far as its training data make that condition necessary for prediction. Learning from trajectories that diverge under intervention can help scientific simulators predict what would happen if a system's forcing or parameters changed.
Agentic multimodal large language models (MLLMs) that think with images interleave reasoning with zoom-in tool calls. However, this sequential loop is slow, and its answers carry no reliability score. Prior speculation of tokens, reasoning steps, or answers keeps the loop, and answer speculation also accepts confident errors. We find that the loop mostly adds one zoom-in step, with which a small model is ahead of the agent on most splits. Our key insight is to speculate the agent's next observation, which, unlike a drafted answer, can be verified, corrected, and used to test the answer. We propose SpecLook, a training-free speculate-and-verify framework with three levels. A small model drafts a glimpse and re-answers from it, and the question ends if the answer is unchanged and confident. Otherwise, the agent's large model verifies and corrects the glimpse without tools, and their answers are fused. Only questions that still lack evidence reach the agent, in one forward pass that SpecLook-Fast skips, so no question runs the loop. The trace scores every answer's reliability without labels. Extensive experiments on V*, HR-Bench, and POPE show that both are more accurate than every baseline on every split. SpecLook reaches 87.2% and 86.2% mean accuracy with 7.7x and 4.7x speedups over the DeepEyes-7B and Thyme-RL agents, and SpecLook-Fast reaches 9.9x and 6.3x.
Purpose. To evaluate the diagnostic accuracy and image navigation behavior of vision-language model (VLM) agents under varying exploration strategies. Materials and Methods. In this retrospective study, six VLMs read 100 emergency CT and MRI examinations published in a public archive (March 2025-August 2026) under four conditions (free navigation, navigation from a key slice, and all slices or key slices in a single request), twice each (4800 runs). Nine radiologists read 30 of those cases. A large language model judge validated against two radiologists graded the answers. Conditions were compared with generalized estimating equations clustered by case. Results. Pooled accuracy was lowest under free navigation (32.6%; 391/1200), followed by navigation from a key slice (37.9%; 455/1200; P = .003), and highest with all slices (42.2%; 506/1200) or key slices (42.5%; 510/1200) in one request (both P < .001). On the 30 cases read by radiologists, freely navigating models were correct in 30.0% of runs (108/360) versus 77.0% for radiologists (208/270; P < .001). Models formed a first hypothesis after a median of 10.8% of the slices and kept it in 63.1% of runs (755/1196); final answers were correct in 12.4% of runs after a wrong first hypothesis (110/885) versus 90.4% after a right one (281/311; P < .001). Median stated confidence in the final answers was 90%, whereas 38.9% of them were correct (1861/4778). Conclusion. Across all exploration strategies, VLM agents remained far less accurate than radiologists. They adopted an early hypothesis and usually kept it, seldom correcting a wrong one.
Autonomous AI agents pair a frozen base model with an editable harness of prompts, skills, tool rules, and memory. Recursive self-improvement (RSI) adapts this harness from experience without weight updates, yet self-editing without external graders often backfires: reflection and scalar scores conflate what an edit newly broke with what remains unsolved, causing local regressions, skill forgetting, and memory bloat. We introduce GUARD (Grader-free Unsupervised Environment Design [UED] with Adaptive Rule Distillation), a closed-loop framework for reference-free continual harness evolution. Our key insight is to decompose each task into pre-execution self-checks and guide harness updates via directional regret. Each round grades trajectories against these checks, pinning mastered rules while prioritizing newly broken behaviors, and selects among parallel candidate edits by critic-free group-relative fitness. A Soft-HURT gate then promotes an edit only when gains outweigh regressions, scoping specialized fixes to prevent over-triggering and guaranteeing non-regression on its gated buffer. At test time, one frozen harness runs in a single pass. Across benchmarks on memory, skill evolution, continual task streams, and domain reasoning, GUARD consistently improves performance without catastrophic forgetting or memory pollution while substantially cutting deployment token usage.
Official development assistance fell by 23.1 per cent in 2025, and the decision of who is helped first when there is not enough for everyone is moving into software. A leading published machine-learning targeting system excluded 53 per cent of eligible households as deployed. If the rule of distribution ends up in code, it should carry the judgement of many cultures, and the machine should execute while a person judges. AllotNet is an interpretable mixture-of-experts conditional logit for the allocation judgements elicited by Allot, a pre-registered six-language instrument: eight linear taste experts over seven ethical criteria, weighted by a gate over participant context, with a damped residual, an encoder for household intake, an outcome head and a budget-constrained planner that builds a review queue and never rejects anyone. A model of human judgement has no answer key on real data, so we wrote one first: a planted synthetic population of 5.6 million sessions and 104 million responses (98 million usable choices), generated by the production code path from a ground truth hidden from the loss. All sixteen pre-specified checks passed: criterion ranking $\rho = 1.000$, hidden segments recovered (ARI 0.671), outcome head at 99 per cent of the noise ceiling, planner at 99.4 per cent of oracle. The caveats are reported with them: two checks are weaker than their titles, and a live-generation failure cost 260,712 seeds. The model has 23,550 parameters, is trained quantisation-aware for NVFP4 with its interpretable core in fp32 (welfare-rank parity $\rho = 0.9937$ at W4A16 on the shipped bundle; one parity threshold was relaxed post hoc), and ships as a 125 KB bundle with C11, Java, Rust and Python runtimes. Nothing has been trained on real data yet; at the study's real scale a hierarchical Bayesian mixed logit is the registered comparator.
Preparing, executing, and validating strongly coupled multi-physics simulations require expert time. Agentic workflows use large language models (LLMs) to coordinate them from natural-language descriptions, but reliability requires checks of physical consistency and fidelity to the problem description. Our agentic harness implements explicit task allocation: LLMs interpret user requests and simulator diagnostics, generate input components not covered by predefined construction rules, and propose repairs, while deterministic routines construct supported components and check inputs, edits, and results. We evaluate CO2-storage problems using the multi-physics simulator GEOS, from compositional flow to flow-poromechanics with frictional fault contact. In 80 attempts on two flow problems, grounding raises the proportion of structurally valid inputs from 5/20 to 20/20, but acceptance reaches 20/20 only with complete rule-based input preparation. With identical inputs, local repairs correct three seeded defects while preserving valid inputs. Supporting tests identify simulations passing physics checks despite differing from the problem description or reference results.
Automatic generation of mathematics assessment items is a growing area in educational artificial intelligence, yet classroom deployment remains challenging. Generated questions must satisfy multiple constraints: mathematical correctness, curriculum alignment, appropriate cognitive difficulty, and educator reviewability. While prior work has advanced text generation, reliable assessment authoring requires granular control over localized examination formats and expert educational judgment. To address this challenge, we present VeriMath, a generalized, human-in-the-loop framework for generating rigorous, curriculum-constrained mathematics assessment drafts. Using annotated reference items, the framework anchors generation to the requested examination section and cognitive reasoning level, then applies sequential automated verification and guided repair before submitting drafts for expert review. We evaluate VeriMath through a national curriculum case study using 4,555 mathematics items from the Vietnamese Grade 12 curriculum and 50 distinct generation requirements covering multiple-choice, true/false, and short-answer formats. Across two model backbones, the framework produces verifier-accepted drafts for 94% (47/50) of requirements per run. Blind expert evaluations confirm high mathematical accuracy and curriculum alignment, with strict teacher-readiness assessments accepting up to 56.8% of generated items. Analysis of rejected drafts reveals persistent domain-agnostic challenges involving formatting, statement validity, and solution-step completeness. These findings demonstrate that VeriMath can bridge generative AI and dependable educational applications by producing highly reviewable assessment drafts within tightly constrained national curriculum frameworks.
Access to large-scale clinical text collections is crucial for advancing medical information retrieval research, including clinical note search, patient cohort identification, and domain-adapted retrieval models. However, real clinical notes are rarely available due to privacy regulations, which particularly limits progress for low-resource languages such as European Portuguese. We present LusoClin, a publicly released dataset of synthetic clinical notes in European Portuguese, created to support privacy-preserving research on clinical text retrieval. LusoClin was generated by pre-processing real clinical data, fine-tuning the open-source Llama 3.2 (3B) instruction-tuned model using Quantized Low-Rank Adaptation, and applying prompt-based clinical note generation. To facilitate semantic indexing and clinical search, the notes are automatically annotated with clinical entities using a BERT-based model. We evaluate LusoClin in terms of utility, fidelity, diversity, and privacy using automatic metrics and human evaluation. Results indicate that the dataset captures clinically relevant language patterns and can support downstream IR applications such as clinical document retrieval and domain-specific model adaptation, although annotators can often distinguish synthetic from real notes and some medical inconsistencies persist.
Compact convolutional backbones such as EfficientNet achieve strong parameter efficiency through local depthwise separable convolutions, yet two structural questions remain central to hybrid design: where does global self-attention improve recognition over pure convolution, and can repeated blocks share weights without losing representational capacity? We investigate both questions on ImageNet-1k (50,000-image validation) using controlled EfficientNet-B0 experiments. First, we show that wrapping a classification backbone in multi-scale feature pyramids (BiFPN) fails to outperform a plain MBConv trunk (69.99% - 71.81% vs. 72.86% plain B0) until paired with stage attention (73.15%). In contrast, keeping the MBConv trunk intact and inserting lightweight zero-gated residual self-attention adapters at stages P3 (28x28 windowed), P4 (14x14 global), and P5 (7x7 global) reaches 74.02% standalone and outperforms a bit-for-bit matched plain-B0 control by +0.954 percentage points (73.88% vs. 72.92%) across an 88-epoch stateless replay schedule, with learned attention amplitude rising monotonically with semantic depth (27.2% at P3 -> 45.4% at P4 -> 66.3% at P5). Second, we evaluate recurrent weight sharing across repeated MBConv blocks with pass-specific BatchNorm statistics (+1,536 parameters per pass). While post-hoc kernel averaging degrades accuracy to 17.03% and single-template stage tying (-27.6% parameters) plateaus at 66.10%, alternating two shared block templates ([B1, A, B, A]) in Stage 6 cuts total model parameters by 10.9% (582K weights) and recovers by +10.82 points in a single epoch to 70.92% (-1.84 points vs. unshared control).
Deep vision models operating in resource-constrained environments depend heavily on low-bit quantization, yet compressing weights below 4 bits frequently induces catastrophic accuracy collapse. While mixed-precision recovery strategies aim to rescue accuracy by selectively retaining high-precision parameters, the fundamental mechanics of where and why these models collapse remain poorly understood. In this study, we systematically analyze the anatomy of quantization collapse across standard Vision CNNs (including ResNet and EfficientNet backbones) evaluated on the full 50,000-image ImageNet-1k validation set. First, we map the exact progression of degradation, demonstrating that collapse is highly non-linear and concentrated in late-stage pointwise projections and depthwise layers. Second, we introduce a structured, multi-step recovery waterfall that evaluates selective integer precision (INT4/INT8) and selective floating-point scales. We show that keeping less than 5% of critical layer activations in 8-bit precision recovers up to 92% of the accuracy lost during aggressive sub-4-bit compression. Finally, we bridge the hardware reality gap by evaluating deployment performance on standard ARM and x86 CPU architectures. We show that while low-bit quantization reduces memory storage requirements, standard runtime software bit-unpacking introduces a 1.48x to 21.5x execution slowdown depending on batch size, highlighting a critical trade-off between model footprint and actual edge throughput. Our findings provide a practical blueprint for navigating the limits of sub-4-bit deployment without sacrificing hardware performance.
While multimodal language models can interpret diagrams and generate structured outputs, their capacity to connect passive radio-frequency (RF) design intent, layout geometry, and electromagnetic behavior remains largely untested. To address this gap, we introduce RFBench, a simulator-grounded dataset and benchmark for passive RF layout design and understanding. The dataset contains 10,000 two-port planar layouts simulated using ADS Momentum, with each layout represented by editable geometric primitives, a raster image, a 16x16 occupancy grid, complex S-parameters at 201 frequency points, a structural category, and an RF response label. Building on these aligned representations, we define seven tasks spanning specification-conditioned layout generation, representation translation, response prediction, simulator-based validation, design diagnosis and improvement, structural understanding, and response classification. Extensive baselines across seven multimodal language models reveal systematic weaknesses in generating structurally valid layouts and predicting electromagnetically faithful responses. We show that RF-specific fine-tuning increases T1a structural validity from 0.0% to 83.2% and the T3 primitive-input within-2-dB rate from 17.7% to 62.8%, demonstrating the value of domain adaptation while leaving considerable room for further progress. RFBench establishes a common testbed for evaluating multimodal RF capabilities and guiding the development of models that connect design intent and layout representations with physical behavior.
Large language model (LLM) agents achieve state-of-the-art performance across domains through chain-of-thought (CoT) reasoning. Despite their ubiquitous use, little is understood about the relationship between the externalized trace and the model's actual computation. We address this by investigating how LLM agents make decisions in iterative games with incomplete information. Using a series of mechanistic interpretability techniques, we find that the model forms and iteratively revises an internal representation of the game throughout the CoT. Once the model prematurely commits to an incorrect belief, it rarely recovers. This failure is manifested in both the literal trace and the layer activations. Ultimately, as multi-agent systems are more prevalent, our work opens the door to a more focused diagnosis of their strategic reasoning errors. For illustrative purposes, we offer an interactive visualizer of our methodology on selected sample traces at https://interp-self.github.io/gametrace/
The rapid scale-up of AI infrastructure has made datacenter reliability a cornerstone of the cloud business economy. Since hardware failures in hyper-scale AI clusters can stall multi-million dollar training workloads, the downtime of individual machines can cause outsized economic losses. Minimizing machine downtime, therefore, is more important than ever. To address this, we present Hardware Diagnoser (HardwareDx), an AI-driven hardware diagnosis system that applies a graph foundation model approach to automated datacenter maintenance. It allows a single model to diagnose many platforms and accelerate onboarding of new hardware products. Deployed at production scale at Google, HardwareDx yields substantial efficiency gains across diverse compute architectures. Real-world evaluations demonstrate that it reduces production downtime by 18.3% on GPU platforms and 16.1% on TPUs. Furthermore, it significantly accelerates hardware deployment, cutting quality assurance (QA) time by 28.3% for GPUs and 25.1% for TPUs. These results establish HardwareDx as a field-tested solution for maximizing the reliability and cost efficiency of next generation AI fleets.
On-policy distillation (OPD) trains a student language model to match a teacher model based on its own generations. When using OPD, one must choose whether to apply it at the sequence or token level and which divergence to minimize. However, despite the wide adoption of OPD in practice, the formal understanding of which OPD objective is appropriate for a given setting is limited. Toward addressing this gap, we theoretically analyze how the choice of OPD objective governs which tokens are prioritized, in what order, and how efficiently. Our analysis reveals that sequence-level OPD exhibits a future agreement bias, which myopic token-level objectives avoid. This bias induces a reverse learning order, in which sequence-level OPD first matches the teacher at later token positions, while at earlier positions it can favor tokens that are improbable under the teacher, and thus often undesirable. As for the choice of divergence, we prove that it strongly influences the amenability of OPD to gradient-based optimization. In particular, while the forward KL divergence enjoys a well-behaved objective landscape, other $f$-divergences, such as reverse KL, can suffer from vanishing gradients even when the student and teacher are far apart. Across language model post-training settings, we demonstrate the implications of these theoretical findings for which OPD objective is preferable, highlighting the importance of accounting for the interplay between the student, teacher, and task at hand.
Post-training quantization is the standard way to serve large language models at low cost. However, the best quantization recipe differs across models and precisions. As a result, experts must redesign it for every new model, which is time-consuming and often sub-optimal. Fortunately, an LLM agent can take over this design by refining the recipe over rounds of building and evaluating candidates, so that more rounds lead to better recipes. However, evaluating one quantized reasoning model takes tens of GPU-hours, and combining methods from different codebases means porting each to the new architecture by hand. We present QDA, a system in which an agent designs the recipe for each model and precision. To make evaluation fast, QDA scores each candidate in minutes by its output KL divergence from the original model on reasoning traces that the agent never sees, and checks promising ones with long generations to catch failures such as repetition loops. To make building fast, QDA provides a library of public PTQ methods that are compatible with one another and with new model architectures. Across four reasoning models, QDA improves accuracy over the strongest public recipe by 0.9, 1.4 and 2.2 percentage points on average at W4A16, W4A8 and W3A16, using 64 GPU-hours per search. On the 320B-parameter MoE model GLM-5.3-Flash at NVFP4 W4A4, it recovers 98.3% of the FP8 model's accuracy, against 96.9% for the best vendor checkpoint. Every QDA checkpoint also runs on existing serving kernels with no inference overhead. We will release the library, the evaluator and the harness.
Coding agents are billed by the token, yet software engineering says nothing about what a token actually buys. Here we introduce a theory of software thermodynamics. The theory draws analogues from classical and stochastic thermodynamics, extends them, and applies them to modern agentic coding. We count the tokens an agent generates as its cost and divide that cost into work and heat. The test suite acts as a software Maxwell's daemon that judges every coding change. We test the theory with an analogue of the experiments in which a single molecule is pulled open and let fold again: an agent makes a change to a repository (coding) and then undoes it (uncoding), many times over, and we count the tokens of every model call. The cycle returns the requested edits to where they began, with each direction started from a stored reference tree, and the spread of its cost over repeated cycles defines a software temperature, which rose with the sampling temperature of the model. About 80 per cent of the tokens of a successful run are generated in model calls that leave the requested change where it was, and we call these tokens heat. The free energy of a change, which we estimate by analogy with the equality of Jarzynski, is 14 to 22 per cent of the mean cost of a successful coding run. Coding costs 1.1 to 1.5 times as much as uncoding. The daemon is imperfect: the verdict of a test suite varies with its environment, and a suite that accepts a broken change has a cost in tokens. The practical implication of the proposed theory could be new tools for analysing and optimising coding agents through the lens of thermodynamics.
Language models increasingly exhibit behaviors that are difficult to detect in transcripts, such as reward hacking, deception, and sycophancy. Asking the assistant follow-up questions about its actions is unreliable, since models are generally not trained to report their own misbehavior. As a result, approaches to oversight typically involve querying external models. These face limitations: external transcript monitoring operates only on visible model outputs, and activation decoders lack full context and are prone to confabulation. We introduce introspective oversight, which trains a new role in the model to answer arbitrary questions about the assistant, rewarding answers that are verified against the assistant's actual behavior. We implement this by adding a separate, third-person observer role that answers these queries, which we can train solely for accuracy and separately from the assistant. On five held-out oversight evaluations, an observer trained on Qwen3.6-27B obtains a mean AUROC of 0.907, outperforming honesty training (0.864), transcript monitoring (0.828), and activation oracles (0.796). We further find that training an observer makes the assistant itself easier to monitor, with every baseline method improving on average when applied to the fine-tuned assistant. These results suggest that training models to answer verifiable questions about themselves is a tractable approach to overseeing unintended behavior.