Papers

Reliable detection of out-of-distribution (OOD) samples is crucial for the safe deployment of machine learning models. Neural networks often produce overconfident predictions for inputs that deviate from their training data, leading to significant degradation in performance. While many OOD detection methods focus on the final output layer, they neglect the rich hierarchical information present in intermediate network layers. This paper introduces a novel approach that leverages sparse autoencoders (SAEs) to learn interpretable features from these intermediate activations. We find that in-distribution (ID) and OOD data activate distinct sets of these sparse features. We propose a new OOD score derived from the cosine similarity between the sparse feature activations of a test sample and the mean activations of ID classes. Our post-hoc detection method not only achieves state-of-the-art performance on standard OOD detection benchmarks, but yields interpretable insights into how distribution shift affects learned representations.

Evaluating retail conversational agents requires methods beyond lexical-overlap metrics to assess intent alignment, factuality, helpfulness, clarity, tone, and overall response quality. Although LLM-as-a-judge methods provide scalable alternatives to human evaluation, production deployment introduces challenges in governance, reproducibility, cost, schema consistency, traceability, and reliability. We present GenAI Evaluation, a governed, configuration-driven pipeline for large-scale evaluation of retail conversational systems. It processes production chatbot logs through normalization, sharding, asynchronous execution, and schema-constrained LLM scoring. The framework evaluates helpfulness, truthfulness, clarity, tone alignment, and translation-specific dimensions. Selective re-evaluation processes only incomplete, malformed, or schema-invalid records, while schema locking, versioned configurations, validation logs, and record-level provenance support auditability. The framework processes approximately 50,000 records daily and has evaluated more than two million interactions. Validation used 12,980 stratified-random human-labeled records from four trained annotators. Classification covered 14 intents, 156 sub-intents, 18 major domains, and 129 sub-domains. The pipeline achieved a macro F1 score of 0.93 and 89% human-acceptability accuracy for translation.

Multi-agent language-model (LM) systems often determine which agents communicate, yet routing is usually treated as an implementation detail. We ask whether routing itself determines whether a population converges on a shared convention or fragments into persistent cliques. We study open-weight agents spanning 1.1B-32B parameters in a controlled naming game, tracking both emitted labels and full first-token preference distributions over the allowed labels. Similarity-based routing can isolate emerging conventions and sustain fragmentation even when every agent interacts in every round. Matched controls show that this effect is not explained solely by uneven participation or model-family-specific score preferences: random rematching and policies that connect disagreeing groups improve coordination when partner-label history is retained, but not when it is absent. Exposure alone is nevertheless insufficient, as some mixed-model populations remain divided despite frequent cross-family interaction, although the same models coordinate homogeneously. Trajectory and controlled-history analyses further distinguish reaching consensus from maintaining it. Finally, ARC-Challenge and MMLU experiments show that routing changes how correct and incorrect answers propagate without reliably improving accuracy. These results establish the runtime interaction graph as a causal design variable whose effects depend jointly on memory, model response, and population composition.

Background: Deep learning models can classify thyroid nodules on ultrasound, but reliable clinical decision support also requires calibrated probabilities, uncertainty estimation, and selective referral, particularly under dataset shift. Methods: We developed a calibrated deterministic five-member deep ensemble for ROI-based thyroid nodule classification and selective image-based triage. TN5000 was used for model development, five-fold cross-validation, member-wise vector-scaling calibration, and fold-specific threshold selection. TN3K served as an independent external dataset-shift evaluation. The framework used ConvNeXt-Tiny with squeeze-and-excitation attention, ensemble-mean malignancy probability, and mutual information (MI) as an ensemble-disagreement score. A three-tier policy assigned images to No-FNA suggestion, FNA recommendation, or radiologist review. Results: On pooled out-of-fold TN5000 predictions, the ensemble achieved AUC-ROC 0.9395, AP 0.9715, ECE 0.0088, and Brier score 0.0813. At 50% nominal MI retention, 7.2% of cases received a No-FNA suggestion, 39.9% an FNA recommendation, and 52.9% radiologist review, with 98.3% No-FNA NPV and 99.83% malignancy capture. On TN3K, AUC-ROC decreased to 0.7870, AP to 0.7254, ECE increased to 0.1899, and Brier score to 0.2281. The frozen TN5000 policy assigned 83.7% to review, 1.0% to No-FNA, and 15.3% to FNA recommendation. No malignant image entered the No-FNA pathway, but FNA-recommendation PPV fell to 76.6%. Conclusion: The framework showed strong internal discrimination and calibration, but limited external threshold transportability. Selective prediction may help identify images unsuitable for automated triage, but local recalibration, threshold validation, and prospective clinical evaluation are required before deployment.

While visual navigation has been extensively studied in agricultural robotics, most existing systems assume daytime conditions. In fact, deploying autonomous robots at night offers significant advantages, including 24-hour crop and soil monitoring, fruit harvesting, and nocturnal pest detection. Modern vision-based systems, however, rely heavily on large-scale well-annotated image datasets, which remains challenging to obtain for nighttime operation scenarios. To address this, we propose an unsupervised image translation framework that converts daytime plant-row RGB images into near-infrared (NIR) nighttime counterparts without requiring pixel-to-pixel supervision. This enables the direct reuse of daytime semantic labels for training nighttime perception models. In particular, by incorporating a pre-trained Contrastive Language-Image Pre-training (CLIP) model, the proposed framework is designed to preserve semantic consistency during day-to-night translation. Additionally, a visibility mask is introduced to account for the limited effective range of NIR illumination in nighttime scenes. We conduct comparative evaluations with state-of-the-art image translation baselines and demonstrate higher image qualities, as supported by improved performance in downstream semantic segmentation for nighttime visual navigation. For evaluation, we utilize AgriNight–a novel dataset comprising 428 daytime and 549 nighttime images collected using night-vision-equipped mobile robots in agricultural fields and manually annotated with pixel-wise semantic labels–and introduce it as the first benchmark for nighttime agricultural visual navigation. We also perform real-time autonomous navigation experiments with a physical robot operating at night. The data and code are available at: https://github.com/mamorobel/AgriNight.

Given a vulnerability-fixing commit, trigger localization asks which specific statement turns the vulnerable program state into a concrete unsafe operation. This question is harder than binary vulnerability detection because the answer demands interprocedural, causal reasoning: in a substantial fraction of real-world CVEs the triggering statement lies several call layers outside the patched function, beyond the reach of static rule sets and pattern-matching language models alike. We present AutoTrace, an agentic pipeline that localizes vulnerability triggers by exploring a code property graph layer by layer, with LLM agents deciding where to look next and deterministic admissibility gates deciding what evidence is required before a trigger can be reported. Agents never accept a trigger on their own authority; every reported trigger is backed by explicit evidence drawn from the graph, so the pipeline covers both intra- and interprocedural vulnerabilities without relying on ungrounded model judgment. On the full InterPVD benchmark, AutoTrace reaches 75.0% VulnHit and 80.8% FuncHit, surpassing the prior state of the art on the same corpus. Building on the same machinery, we construct SinkTrace-Bench, a dataset that exposes each vulnerability as a source-to-sink (S2S) causal chain from attacker-controlled input through propagation to the dangerous operation, drawn from matched vulnerable and patched program states. It comprises 1,542 verifier-confirmed, perfectly balanced vulnerable/safe samples whose label fidelity we audit against expert annotations. Benchmarking frontier LLMs on it, we find that even the strongest struggle to separate the matched pairs, exposing the causal-reasoning gap that trigger localization targets. Artifact available at https://github.com/Erroristotle/AutoTrace.

Online shopping is increasingly shifting toward a model in which AI agents independently search for products, compare options, evaluate constraints, and carry out parts of the purchasing process for users. Website design must now support both human and agent-mediated interaction. This paper introduces the agent-ready website, a design framework for enhancing the readability, interpretability, verifiability, and actionability of e-commerce platforms for AI agents. Existing web design, SEO, and generative engine optimization (GEO) metrics do not fully assess a website's capacity for agent-mediated interaction. The proposed framework is structured around three dimensions agent interpretability, agent executability, and agent decision reliability supported by features such as machine readability, semantic clarity, agent actionability, and contextual decision-reliability signals. The framework is evaluated through a controlled experiment comparing a human-oriented baseline and an agent-ready version of an identical website prototype, with identical catalogs, pricing, stock, and shopping workflows. The evaluation involved five tasks, three browser-agent models (GPT-4.1, Gemini-2.5 Flash, and Grok-4 Fast), and 300 runs, measuring PASS,PARTIAL,FAIL outcomes, strict and functional success rates, error patterns, step counts, and token consumption. The agent-ready website achieved 134 PASS runs out of 150 versus 74 out of 150 for the baseline (strict success rates of 89.3% vs. 49.3%), with the largest gains in product detail extraction, comparison, and multi-constraint selection. It also reduced PARTIAL outcomes from 43 to 3 and lowered the average step count from 9.31 to 6.49. These results provide preliminary evidence that enhanced structural clarity, action cues, evidence signals, and temporal validity indicators can substantially improve the reliability and efficiency of AI browser agents.

Synthetic image attribution aims at identifying the generator responsible for a given AI-generated image. Training-free reference-based attribution methods are easily scalable, since newly emerging generators can be incorporated by adding source-specific references rather than retraining a task-specific classifier. Their performance depends on two coupled factors: the representation space used for comparison and the way source-specific references are constructed. However, the interaction between these two factors remains largely unexplored. In this paper, we provide a controlled analysis of this interaction using references and off-the-shelf pretrained representations. We study representations extracted from different layers of CLIP and DINOv2, along with three reference selection strategies with varying semantic constraints: arbitrary, semantically aligned, and resynthesis-based references. Our results show that attribution accuracy generally peaks at intermediate representation levels, indicating that source-discriminative cues are more accessible before strong semantic abstraction dominates. We further show that intermediate representations are not completely semantically neutral, making reference selection critical: semantically constrained references reduce query-reference mismatch and improve attribution, especially under limited reference budgets. Resynthesis is useful in low-reference regimes, while semantically aligned references provide a better accuracy-cost trade-off when a moderate-sized reference pool is available. Our findings show that training-free reference-based attribution should be understood as the interaction between where images are compared, how the reference set is constructed, and how many references are available. Code and datasets are available at https://github.com/meiling-fdu/RefPrint.

Deploying medical visual question answering (MedVQA) systems in real-world clinical settings requires models that adapt to new clinical tasks without forgetting previously acquired knowledge. Continual learning (CL) provides a practical framework for this setting. Despite rapid progress in medical vision-language models, the behavior of CL methods when training these models across heterogeneous MedVQA tasks remains underexplored. This work presents a systematic evaluation of CL for MedVQA across diverse clinical objectives, including classification, multi-label classification, detection, cell counting, and report generation. Specifically, we explore (1) the ability of existing CL methods to mitigate catastrophic forgetting; (2) their sensitivity to task ordering, analyzing how different task sequences influence performance retention and forgetting; and (3) the evolution of low-rank adaptation parameters as new tasks are learned, revealing patterns of weight drift under different CL methods. Our findings suggest that existing CL methods struggle to maintain stability-plasticity balance when tasks with different objectives and supervision formats are interleaved. Code and full experimental setup will be publicly available.

High deployment cost, poor spatial coverage and susceptibility to storm conditions are all challenges faced by traditional in-situ methods. This paper presents a video-based and high performance computing (HPC) enabled deep learning framework for joint sensor free estimation of five coastal wave parameters, namely significant wave height (Hs), maximum wave height (Hmax), peak period (Tp), zero upcrossing period (Tz) and wave direction (theta) from monocular coastal video. The proposed architecture comprises of a V-JEPA (self supervised) ViT Small backbone for robust spatiotemporal feature extraction in visually challenging scenarios, a dual-stream SlowFast temporal encoder for broad bandwidth representation of wave motion in both hydrodynamic breaking and swell regimes, an optical flow stream based on Farneback optical flow algorithm for adding saliency information to the structure with emphasis on hydrodynamically active wavelength bands of waves, and a multi-task regression layer with dispersion constraints (Airy wave dispersion lambda_p = 0.1). The model was trained on an NVIDIA DGX A100 cluster and was early stopped at epoch 31 and achieved Pearson correlation coefficients of 0.451, 0.578, 0.643, 0.680 and 0.832 for Hs, Hmax, Tp, Tz and wave direction respectively, with generalization ability to geographically diverse held out test data sites. While operating in a data-limited regime (6 annotated training scenes), the framework demonstrates statistically significant temporal correlations (PCC of 0.451 to 0.832), confirming proof of concept feasibility; R2 values (max 0.246) indicate that variance capture will improve with larger annotated datasets.

Multi-task model merging combines separately trained expert models into a single model that handles all tasks without co-training. Standard practice merges experts at their optimal validation loss. We challenge this convention by systematically studying how training duration of domain experts affects the quality of the merged model. We fine-tune experts on five domains (Math, Code, Instruction Following, Multilingual, and Safety) across three model sizes (Qwen 3.5 0.8B, 2B, and 4B), saving checkpoints from 25% to 500% of the optimal training steps and evaluating five merging methods at each duration. Our findings reveal a striking method-dependent pattern: simple averaging degrades sharply with overfitting, while sparsification-based methods achieve their best performance well past the validation optimum. We formalize this through bias-variance decomposition analysis, drawing a parallel to random forests where averaging benefits from high-variance individual learners. These results suggest that training duration and merging method should be chosen jointly rather than independently.

Classifying cybersecurity vulnerabilities using the Common Weakness Enumeration (CWE) taxonomy is challenging due to extreme class imbalance and strong hierarchical dependencies among weakness categories. Although oversampling techniques such as Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic Sampling (ADASYN) are widely adopted to mitigate class imbalance, their effectiveness for hierarchical CWE text classification remains largely unexplored. This paper proposes a Hierarchy-Aware RoBERTa framework that explicitly incorporates CWE structural information through learnable parent-class embeddings, preserving taxonomic consistency. Our experiments demonstrate that synthetic interpolation in high-dimensional embedding spaces violates the inherent parent-child constraints of the CWE hierarchy, offering only marginal benefits for classical ML models while consistently degrading deep learning architectures. Evaluated on a CWE Research Concept dataset, the proposed model achieves a weighted F1-score of 0.76 without data augmentation, outperforming all baselines with notable gains on minority classes, including the Class category whose F1-score improved from 0.40 to 0.60 over the BERT baseline. Our results suggest that hierarchy-aware representation learning is a more principled alternative to oversampling for structured vulnerability classification.

Feedforward network (FFN) blocks account for a large fraction of the parameters and computation in Transformer architectures, yet their internal structure remains difficult to interpret due to the additive superposition induced by the residual stream. We examine whether the activation of an FFN neuron can be explained by a sparse set of preceding neuron activations and attention outputs. We introduce a training-free attribution method that estimates the relative influence of upstream neurons and attention outputs on a target neuron's activation. Empirically, across models and layers, we find that small subsets of preceding activations and attention outputs suffice to preserve neuron activations with high fidelity when all remaining inputs are masked with their average values. Effective sparsity is even greater when accounting for the inherent activation sparsity of upstream layers. Moreover, applying the neuron-specific masks in all layers simultaneously, such that the induced deviations propagate through the network, leaves model perplexity largely unchanged at moderate sparsity levels. These results demonstrate that, despite dense parameterization, FFNs exhibit sparse and structured inter-layer dependencies at the neuron level. Our method provides a practical, scalable tool for circuit-level interpretability and identifies candidate sparse pathways with potential implications for efficient inference.

A specialist tolerates blind spots that a generalist does not. Usually this is treated as a cost to be minimized. We treat it as a design variable: a deficiency can be kept because it pays and removed on demand in the rare situation where it would be fatal, by routing to a compensation channel. We give three results. First, an advantage condition under which keeping the deficiency is a computable economic position; structurally it is the Ehrlich-Becker market-vs-self-insurance margin applied to a competence gap, with the detector as a Townsend costly-state-verification technology. Second, a two-sided characterization of removability. A coupling lemma shows that when the deficiency is a coarsening of perception, no switch can separate benefit from harm, yielding a converse (a confounded detector earns zero premium, and any within-defect policy insisting on positive premium is driven, under multiplicative dynamics, to negative long-run growth) and an achievability result (a detector outside the deficiency earns a positive premium). Together, over structured uncertainty classes with severity capped or miss rate O(1/L): a defect is profitably removable iff the detector-relevant distinction survives the restriction and the advantage condition holds; the premium is the support function of the class's ROC set at an economic price vector. Third, observation defects and capacity defects differ exactly on whether access to the deployment distribution rescues them; the gap decomposes as cross-leak plus a closure deficit, and per-task randomization buys back the latter, never the former. The detector can be learned from declared fatal categories at a training bill linear in loss severity (up to a log factor). The results synthesize Chow's reject option, Kelly growth under ruin, and selective prediction.

Aggregate reliability estimates can obscure heterogeneity in measurement-design burden across response conditions, so a single G- or D-study may mischaracterize a design's adequacy for particular strata. This study introduces a conditional generalizability framework with three components. First, automated scoring configurations – the encoder architectures and scoring-head families admissible within a fixed pipeline – are treated as a universe of admissible measurement conditions rather than incidental modeling choices. Second, analytical D-study projections are compared with empirical configuration sweeps over a finite scoring pool, yielding two estimands of design adequacy whose agreement or divergence diagnoses the realized configuration universe. Third, evidence is conditioned on entropy-defined response strata, treating entropy as an operational stratification variable, not a construct claim about writing quality. Whereas recent generalizability-theory extensions address AI-generated item variants on the response side, this framework addresses the analogous scoring-side problem: AI-mediated scoring configurations. Demonstrated with automated essay scoring of timed L2 writing, the realized design was dependable in aggregate (Phi approx 0.76). Re-estimated within entropy strata, dependability stayed high but declined modestly and robustly (Phi = 0.88, 0.87, 0.84) – a gradient implying different decision-study requirements, the highest-entropy stratum requiring the most crossed conditions. The framework offers a portable workflow for evaluating nonuniform dependability.

Like many optimization-driven domains, railway rescheduling relies on Mixed-Integer Linear Programming (MILP), yet the field's modeling knowledge is scattered across hundreds of papers in incompatible notations, and narrative surveys organize it subjectively: they classify models by vocabulary rather than by structure, and reproduce neither. We present LP Mining with LP2Graph, a method that mines the structure of published LP and MILP formulations into a reproducible dataset and an induced taxonomy. Its core, LP2Graph, represents each formulation admitted by its canonical grammar as a typed variable–equation graph derived from a single canonical model; once a source is extracted into that model, everything downstream is deterministic. Each source is parsed into this model, homologized, and clustered bottom-up (over variables, then constraints and the objective, then whole-model structure) and, separately, by application domain and solution approach; the resulting groups are labeled by a rule-seeded, self-updating classifier. We validate the representation rather than assume it: per-cluster representatives are regenerated as independent LaTeX and re-solved across CBC, HiGHS and Gurobi against the optimum reported in the source paper. The outcome is an objective, repeatable taxonomy of variables, constraints and model types: the principled foundation on which our raiLPminer line of automated railway-rescheduling model development builds.

Precision molecular design aims to discover personalized drug candidates through joint control of multiple conditions, such as biological relevance and molecular design strategies. Biological relevance reflects cellular functional states under disease or perturbation conditions, while molecular design strategies provide complementary guidance in terms of structural intentions and property optimization. In this study, we propose JoPMol, a jointly controlled precision molecular generative model that integrates biological states encoded by gene expression profiles with molecular structure information expressed in text, and chemical properties quantified by numerical values within a unified modeling framework. This formulation enables coordinated generation and optimization of candidate molecules under joint condition control. Experimental results show that JoPMol outperforms state-of-the-art methods across multiple evaluation metrics. Moreover, JoPMol demonstrates strong generalization ability in both transfer tasks and biologically grounded simulation scenarios, validating its effectiveness for precision molecular design. The source code is publicly available at https://github.com/hala-yh/JoPMol.

In 2019, OpenAI released two million GPT-2 outputs-ungrammatical, half broken-to aid the detection of machine-generated text. The alignment that produced their more fluent successors is usually regarded as an engineering achievement; we read it instead as the newest expression of optimization culture: the conviction, older than the technology, that measurable improvement along predefined axes exhausts the question of value. Tracing that conviction through the stack-pretraining, decoding, preference tuning, benchmarking, interface-and back through its genealogy in the audit society, we arrive at the limit: an optimization procedure can measure how improbable a piece of generated text is; it cannot tell whether that unlikelihood is error or invention. A procedure that cannot make that distinction has nonetheless, within half a decade, assumed the authority to set the protocols of legitimate language. Held for centuries by academies and schoolrooms, grammars and examiners, this authority has been given over to loss functions, reward models, benchmarks, and system prompts: an apparatus that executes the office of judgment with no capacity for judging.

Current machine unlearning methods predominantly rely on global, coarse-grained intervention strategies. They lack precise pilot signals to guide the unlearning process and fail to provide differentiable guidance across different unlearning tasks. Due to the varying memorization strengths of samples during original training, such a uniform strategy leads to two problems: some samples are over-unlearned, which harms model utility; while others are under-unlearned, leaving residual information that can be exploited by privacy attacks. In this paper, we propose GSUO, a guidance-signal-aware unlearning optimization framework that designs task-specific fine-grained guidance signals to steer the unlearning process and is applicable to both random-subset and class-wise forgetting tasks. Extensive experiments demonstrate that GSUO outperforms 14 baselines in terms of both unlearning effectiveness and generalization, while achieving high efficiency and significant speedups, validating its effectiveness for reliable machine unlearning.

Most neural partial differential equation (PDE) surrogates learn how fields evolve after a grid has already been chosen. However, before any operator is applied, the grid has already determined how modeling capacity is allocated across space, resolution, and spectral bandwidth. We argue that this hidden design choice should itself be learnable, leading to a question different from standard operator learning: can a surrogate learn where resolution should exist before predicting field evolution? We formulate adaptive discretization as a physics-constrained conditional generation problem over valid mesh displacements. The success of diffusion models in PDE field prediction suggests their potential for learning adaptive discretizations under similar structured constraints. This leads to a two-stage diffusion framework: Stage 1 learns an r-adaptive displacement mesh conditioned on the observed dynamics, while Stage 2 predicts the solution evolution from the mesh-informed representation. The mesh generator is regularized by physics-aware proxy channels, geometric validity constraints, and local spectral concentration so that adaptation remains physically interpretable and numerically legal. Across five PDE regimes, the results show that diffusion-based learned discretization is competitive with adaptive-mesh and reduced-order baselines, with particularly strong gains in regimes where fixed or handcrafted allocation is insufficient. The main conclusion is not that there exists a universal optimal mesh rule, but that discretization should be learned in a regime-dependent manner: different spatial and spectral structures favor different allocation behaviors. This reframes adaptive meshing for neural PDE solvers from a solver-specific heuristic into a generative representation-learning problem.

We develop an enhanced in-context learning (ICL) framework to improve the performance of pilot-based beamforming in multi-user multiple-input single-output (MU-MISO) systems. The proposed scheme integrates the ICL-Transformer backbone with the pilot encoder-decoder network (EDN) and the beamformer EDN. A crucial feature of our ICL network is that it can handle multiple channel models without retraining, enabled by the construction of model-specific context datasets. To improve convergence and robustness, we introduce three key innovations: (a) a curriculum learning (CL) strategy that smoothly transitions from supervised LMMSE-labeled imitation to unsupervised sum-rate maximization, (b) a self-evolving mechanism that dynamically expands and refines the context datasets for all channel models during CL-based training, and (c) a mismatch-aware extension that incorporates several mismatches into the general ICL framework and bypasses explicit channel calibrations. Ablation studies validate the effectiveness of the in-context architecture and enhanced training strategies. Simulation results over diverse communication environments show that the proposed scheme is able to rapidly adapt to both seen and unseen channel models without gradient-based parameter updates, and can mitigate the mismatch issues via intelligent context constructions. Furthermore, our scheme consistently outperforms the existing beamforming schemes under pilot-based settings, including the WMMSE benchmark and the recent Transformer-based methods.

Metacognition is a foundational component of intelligence critical to effective learning, problem solving, decision-making, communication, and more. In recent years, it has become increasingly recognized as a cornerstone of capable, transparent AI systems. Yet while LLMs have made significant progress across diverse real-world tasks, it is not yet clear when, how, or to what extent they can exhibit or be endowed with effective metacognitive abilities, nor how such abilities can be adapted to advance the fundamental capabilities, reliability, and intelligence of AI systems. This paper bridges this gap by presenting the first comprehensive overview of the current state of knowledge on metacognition for LLMs. We analyze and taxonomize the landscape of this emerging field and summarize recent technical advancements, including methods and benchmarks to measure and evaluate LLMs' metacognitive abilities, techniques to elicit, improve, and apply metacognition in LLMs, and findings and implications of ongoing research. We also discuss applications, open questions and challenges, and promising directions for future work. Our aim is to provide a detailed and up-to-date review of this topic and stimulate meaningful research and discussion. An organized list of papers can be found at https://github.com/yale-nlp/LLM-Metacognition.

We present a theoretical framework to explain the emergence of inductive reasoning abilities in Transformer language models. While previous works on Transformer learning dynamics have so far been mostly tied to specific tasks, we study a generalized class of inductive tasks that unifies several synthetic tasks known in the literature, including in-context n-grams and multi-hop reasoning. In this class, we theoretically prove that the training dynamics of attention models can be confined to a highly interpretable, low-dimensional invariant manifold. On this manifold, the learning dynamics are captured by a handful of interpretable coordinates rather than millions of parameters, making both theoretical and empirical analysis more tractable. Using this framework, we characterize how data statistics govern the competition between in-context and in-weights learning, we study how random initializations determine the `winning' circuit when multiple solutions are possible, and we demonstrate that the coordinate frame associated with the manifold can be used to automatically detect which circuits have been learned in trained models. By casting circuit formation as a low-dimensional dynamical phenomenon, we take a step toward a predictive theory of how Transformers learn.

Recent work in humanoid whole-body control has found success with a simple recipe: retarget human motion to robot kinematic references, then train policies via reinforcement learning (RL) to track them. But how does this recipe transfer to dexterous manipulation? The answer is not obvious, as manipulation involves complex, contact-rich dynamics and requires delicate regulation of contact modes and forces. We present REGRIND, a minimalist retargeting-guided RL pipeline that learns dexterous manipulation policies from a single human demonstration. REGRIND retargets human hand-object motion to a robot reference that preserves hand-object spatial and contact relationships, trains a residual RL policy in simulation to track object-centric keypoints along that reference, and transfers the resulting policy zero-shot to hardware with careful system identification. The resulting policies produce fluid, human-like behavior on two different multi-fingered hands across contact-rich tool-use tasks, including operating a pair of scissors and turning a screwdriver. Through systematic hardware experiments, we identify and analyze the key factors that govern sim-to-real transfer in dexterous manipulation, offering practical guidance for retargeting-based learning in contact-rich settings. Videos and code are available at https://yunhaifeng.com/REGRIND.

Existing studies of LLM-as-judge scoring bias work predominantly at the input-output level: they perturb inputs, measure score deltas, and propose prompt-level mitigations. We argue that the same biases admit a representation-level account in the judge's hidden state, complementary to the input-output view and operationally useful in ways it does not afford. We report three findings, across seven judges, seven bias types, and nine benchmarks. Geometry: baseline judging inputs occupy a tight activation manifold while biased inputs are displaced along a low-dimensional, type-specific subspace that sharpens with depth and is recovered consistently by three families of estimators. Causal control: steering hidden states along this subspace drives scoring in both directions, forward shifts reproducing biased scoring on clean inputs and reverse shifts restoring baseline scoring on biased ones, while matched-norm random directions produce shifts an order of magnitude smaller. Operational: a simple linear projection onto the same bias-direction features anticipates judge failures on three entirely unseen benchmarks, substantially outperforming text-based alternatives. Reading bias as activation geometry, rather than as input-output noise, unifies geometric structure, causal control, and operational prediction within a single framework. The project page is available at https://xzx34.github.io/unfair-judge/

Current Video Large Language Models (Video LLMs) excel in question answering (QA) but largely operate as black boxes, providing textual answers without verifiable visual grounding. Existing explainability efforts rely on textual rationales or sparse bounding boxes, which struggle to capture complex video dynamics such as occlusions and non-rigid deformations. We propose Evidence-Backed Video Question Answering (E-VQA), a novel task requiring models to jointly output a semantic answer and precise spatio-temporal evidence: temporal segments and dense, tracked object segmentation masklets. To support this, we introduce ST-Evidence, the first human-verified benchmark for both discriminative and generative pixel-level grounding. Evaluations of state-of-the-art models reveal a critical decoupling between QA accuracy and true visual perception that scaling alone fails to bridge. To address this, we develop scalable, automated generation pipelines to create ST-Evidence-Instruct, a 160k-scale dataset bridging high-level reasoning with fine-grained grounding. Fine-tuning grounded Video LLMs on this data yields substantial gains over the corresponding size-matched UniPixel baselines (e.g., +27.2 t-mean and +13.8 J&F on a 7B model), establishing a robust baseline for explainable, evidence-backed video understanding. Code and data are available at https://github.com/SalesforceAIResearch/EVQA.

This paper presents a cascaded Low-Rank Adaptation (LoRA)-based multimodal fusion framework for action and activity recognition in healthcare-oriented training environments. The proposed architecture combines parameter-efficient modality-specific adaptation with sequential fusion, enabling modalities to be integrated in stages without retraining previously learned components. Rather than assuming a fixed fusion structure, the framework first integrates more closely related modalities and then incorporates additional heterogeneous modalities, supporting scalable adaptation across datasets with different modality sets.We evaluate the framework on two healthcare-oriented training environment datasets: NurViD and the Nurse Training dataset. Across these datasets, preliminary results suggest that the proposed cascaded fusion strategy improves over individual modality models and provides competitive performance relative to previously reported dataset-specific baselines. Overall, these findings indicate that cascaded LoRA-based fusion is a promising parameter-efficient approach for integrating heterogeneous modalities in medical training action and activity recognition tasks. github: https://github.com/anonymous0-ai/LoRA-Based-Cascaded-Multimodal-Fusion-.git.

Neural Architecture Search (NAS) has automated the design of deep learning models but traditionally requires massive computational resources, often measured in thousands of GPU-days. In this paper, we propose a frugal and memetic NAS framework designed to democratize architecture design on consumer-grade hardware. Our approach combines the global macro-search capabilities of an autoregressive Transformer controller, trained via Reinforcement Learning (RL), with the local micro-exploitation of an Artificial Bee Colony (ABC) algorithm. To prevent premature convergence during the RL phase, we introduce a dynamic entropy mechanism that forces topological exploration upon detection of performance stagnation. Evaluated on a standard GPU (NVIDIA RTX 3060), our hybrid method effectively resolves the "cold-start" problem inherent in metaheuristics. By algorithmically penalizing network depth, our framework actively mitigates model bloat: on the CIFAR-10 dataset, it discovers an efficient architecture reaching 84.85% accuracy with only $\sim$174,000 parameters (significantly smaller than standard baselines like ResNet-20) in 3 hours of search time. Furthermore, we demonstrate the framework's flexibility by applying it to credit card fraud detection, directly optimizing the F1-Score on highly imbalanced tabular data to reach a F1-Score of 0.71 with a compact network of $\sim$4,600 parameters. These results suggest that our approach can yield tailored, accessible, and highly parameter-efficient deep learning models suitable for edge deployment.

We introduce MM-ToolSandBox, a benchmark and evaluation framework for visually grounded tool-calling agents. The framework provides a stateful execution environment spanning 500+ tools across 16 application domains, supporting multi-image, multi-turn tasks where agents must ground progressively arriving visual inputs into executable tool calls while handling realistic conversational phenomena (goal revisions, error corrections, state mutations). An automated scenario generation pipeline produces diverse, visually grounded scenarios through information-flow-guided planning and multi-stage quality filtering, yielding 258 human-verified nominal scenarios and 50 variants targeting interactive UI applications. Evaluating 12 state-of-the-art models, from 4B open-weight to frontier proprietary systems, shows that current models still lack robust visual tool-calling capability: even the best model achieves below 50% success rate. Our failure analysis further reveals that visual precision, not only planning, is a primary bottleneck for capable models: 53% of failures stem from incorrect information extraction from images despite otherwise correct task workflows. A planning-to-precision crossover emerges with scale: smaller models fail at deciding what to do, while larger models fail at perceiving what they see, suggesting fundamentally different research directions for improving models at different capability levels. The framework and the benchmark are publicly available at https://github.com/apple/ml-mmtoolsandbox

This paper proposes a human-centered artificial intelligence (HCAI) framework for AI-assisted lexicography. While generative AI offers significant opportunities to enhance lexicographic work, it also raises concerns regarding the future role of lexicographers and the preservation of linguistic and cultural diversity. Drawing on HCAI principles and previous applications in other language professions, the paper identifies four interrelated dimensions through which AI integration in lexicography can be understood and critically examined: the augmented lexicographer, the sociotechnical context of AI integration, bias, and the design of AI-powered lexicographic tools. The framework argues that AI should augment rather than replace lexicographers, combining automation with meaningful human control. It further emphasizes the importance of preserving professional agency, mitigating AI-generated biases, and designing tools around the needs of lexicographers. By doing so, the paper provides a foundation for future research and the beneficial integration of AI into lexicographic workflows.