Sim-to-real transfer for contact-rich manipulation remains challenging due to the inherent discrepancy in contact dynamics. While existing methods often rely on costly real-world data or utilize blind compliance through fixed controllers, we propose a framework that leverages expert-designed controller logic for transfer. Inspired by the success of privileged supervision in kinematic tasks, we employ a human-designed finite state machine based position/force controller in simulation to provide privileged guidance. The resulting policy is trained to predict the end-effector pose, contact state, and crucially the desired contact force direction. Unlike force magnitudes, which are highly sensitive to simulation inaccuracies, force directions encode high-level task geometry and remain robust across the sim-to-real gap. At deployment, these predictions configure a force-aware admittance controller. By combining the policy's directional intent with a constant, low-cost manually tuned force magnitude, the system generates adaptive, task-aligned compliance. This tuning is lightweight, typically requiring only a single scalar per contact state. We provide theoretical analysis for stability and robustness to disturbances. Experiments on four real-world tasks, i.e., microwave opening, peg-in-hole, whiteboard wiping, and door opening, demonstrate that our approach significantly outperforms strong baselines in both success rate and robustness. Videos are available at: https://yifei-y.github.io/project-pages/DirectionMatters/.

Detecting prompt injection and jailbreak attacks is critical for deploying LLM-based agents safely. As agents increasingly process untrusted data from emails, documents, tool outputs, and external APIs, robust attack detection becomes essential. Yet current evaluation practices and production systems have fundamental limitations. We present a comprehensive analysis using a diverse benchmark of 18 datasets spanning harmful requests, jailbreaks, indirect prompt injections, and extraction attacks. We propose Leave-One-Dataset-Out (LODO) evaluation to measure true out-of-distribution generalization, revealing that the standard practice of train-test splits from the same dataset sources severely overestimates performance: aggregate metrics show an 8.4 percentage point AUC inflation, but per-dataset gaps range from 1% to 25% accuracy-exposing heterogeneous failure modes. To understand why classifiers fail to generalize, we analyze Sparse Auto-Encoder (SAE) feature coefficients across LODO folds, finding that 28% of top features are dataset-dependent shortcuts whose class signal depends on specific dataset compositions rather than semantic content. We systematically compare production guardrails (PromptGuard 2, LlamaGuard) and LLM-as-judge approaches on our benchmark, finding all three fail on indirect attacks targeting agents (7-37% detection) and that PromptGuard 2 and LlamaGuard cannot evaluate agentic tool injection due to architectural limitations. Finally, we show that LODO-stable SAE features provide more reliable explanations for classifier decisions by filtering dataset artifacts. We release our evaluation framework at https://github.com/maxf-zn/prompt-mining to establish LODO as the appropriate protocol for prompt attack detection research.

Clinical decision-making requires nuanced reasoning over heterogeneous evidence and traceable justifications. While recent LLM multi-agent systems (MAS) show promise, they largely optimise for outcome accuracy while overlooking process-grounded reasoning aligned with clinical standards. One critical real-world case of this is gene-disease validity curation, where experts must determine whether a gene is causally implicated in a disease by synthesising diverse biomedical evidence. We introduce an agent-as-tool reinforcement learning framework for this task with two objectives: (i) process-level supervision to ensure reasoning follows valid clinical pathways, and (ii) efficient coordination via a hierarchical multi-agent system. Our evaluation on the ClinGen dataset shows that with outcome-only rewards, MAS with a GRPO-trained Qwen3-4B supervisor agent substantially improves final outcome accuracy from 0.195 with a base model supervisor to 0.732, but results in poor process alignment (0.392 F1). Conversely, with process + outcome rewards, MAS with GRPO-trained supervisor achieves higher outcome accuracy (0.750) while significantly improving process fidelity to 0.520 F1. Our code is available at https://github.com/chaeeunlee-io/GeneDiseaseCurationAgents.

Sparse Mixture-of-Experts (MoE) models scale Transformers efficiently but suffer from expert overlap -- redundant representations across experts and routing ambiguity, resulting in severely underutilized model capacity. While architectural solutions like DeepSeekMoE promote specialization, they require substantial structural modifications and rely solely on intra-layer signals. In this paper, we propose two plug-and-play regularization losses that enhance MoE specialization and routing efficiency without modifying router or model architectures. First, an intra-layer specialization loss penalizes cosine similarity between experts' SwiGLU activations on identical tokens, encouraging experts to specialize in complementary knowledge. Second, a cross-layer coupling loss maximizes joint Top-$k$ routing probabilities across adjacent layers, establishing coherent expert pathways through network depth while reinforcing intra-layer expert specialization. Both losses are orthogonal to the standard load-balancing loss and compatible with both the shared-expert architecture in DeepSeekMoE and vanilla top-$k$ MoE architectures. We implement both losses as a drop-in Megatron-LM module. Extensive experiments across pre-training, fine-tuning, and zero-shot benchmarks demonstrate consistent task gains, higher expert specialization, and lower-entropy routing; together, these improvements translate into faster inference via more stable expert pathways.

Large language models (LLMs) show promise for healthcare question answering, but clinical use is limited by weak verification, insufficient evidence grounding, and unreliable confidence signalling. We propose a multi-agent medical QA framework that combines complementary LLMs with evidence retrieval, uncertainty estimation, and bias checks to improve answer reliability. Our approach has two phases. First, we fine-tune three representative LLM families (GPT, LLaMA, and DeepSeek R1) on MedQuAD-derived medical QA data (20k+ question-answer pairs across multiple NIH domains) and benchmark generation quality. DeepSeek R1 achieves the strongest scores (ROUGE-1 0.536 +- 0.04; ROUGE-2 0.226 +-0.03; BLEU 0.098 -+ 0.018) and substantially outperforms the specialised biomedical baseline BioGPT in zero-shot evaluation. Second, we implement a modular multi-agent pipeline in which a Clinical Reasoning agent (fine-tuned LLaMA) produces structured explanations, an Evidence Retrieval agent queries PubMed to ground responses in recent literature, and a Refinement agent (DeepSeek R1) improves clarity and factual consistency; an optional human validation path is triggered for high-risk or high-uncertainty cases. Safety mechanisms include Monte Carlo dropout and perplexity-based uncertainty scoring, plus lexical and sentiment-based bias detection supported by LIME/SHAP-based analyses. In evaluation, the full system achieves 87% accuracy with relevance around 0.80, and evidence augmentation reduces uncertainty (perplexity 4.13) compared to base responses, with mean end-to-end latency of 36.5 seconds under the reported configuration. Overall, the results indicate that agent specialisation and verification layers can mitigate key single-model limitations and provide a practical, extensible design for evidence-based and bias-aware medical AI.

Purpose: Precise port placement is a critical step in robot-assisted surgery, where port configuration influences both visual access to the operative field and instrument maneuverability. To bridge the gap between preoperative planning and intraoperative execution, we present ARport, an augmented reality (AR) system that automatically maps pre-planned trocar layouts onto the patient's body surface, providing intuitive spatial guidance during surgical preparation. Methods: ARport, implemented on an optical see-through head-mounted display (OST-HMD), operates without any external sensors or markers, simplifying setup and enhancing workflow integration. It reconstructs the operative scene from RGB, depth, and pose data captured by the OST-HMD, extracts the patient's body surface using a foundation model, and performs surface-based markerless registration to align preoperative anatomical models to the extracted patient's body surface, enabling in-situ visualization of planned trocar layouts. A demonstration video illustrating the overall workflow is available online. Results: In full-scale human-phantom experiments, ARport accurately overlaid pre-planned trocar sites onto the physical phantom, achieving consistent spatial correspondence between virtual plans and real anatomy. Conclusion: ARport provides a fully marker-free and hardware-minimal solution for visualizing preoperative trocar plans directly on the patient's body surface. The system facilitates efficient intraoperative setup and demonstrates potential for seamless integration into routine clinical workflows.

Diffusion language models (dLLMs) recently emerged as a promising alternative to auto-regressive LLMs. The latest works further extended it to multimodal understanding and generation tasks. In this work, we propose LaViDa-R1, a multimodal, general-purpose reasoning dLLM. Unlike existing works that build reasoning dLLMs through task-specific reinforcement learning, LaViDa-R1 incorporates diverse multimodal understanding and generation tasks in a unified manner. In particular, LaViDa-R1 is built with a novel unified post-training framework that seamlessly integrates supervised finetuning (SFT) and multi-task reinforcement learning (RL). It employs several novel training techniques, including answer-forcing, tree search, and complementary likelihood estimation, to enhance effectiveness and scalability. Extensive experiments demonstrate LaViDa-R1's strong performance on a wide range of multimodal tasks, including visual math reasoning, reason-intensive grounding, and image editing.

Activation steering provides parameter-efficient control over large language models (LLMs) at inference time, but many methods rely on off-distribution supervision and discrete masking, leading to brittle interventions. We propose ROAST (Rollout-based On-distribution Activation Steering Technique), which estimates steering directions from the model's own on-distribution rollouts via ROC and avoids hard sparsification via Continuous Soft Scaling (CSS) and Grouped Mean Normalization. Our empirical analysis reveals that while activation magnitude correlates moderately with directional consistency, the variance in magnitude is significant and often disproportionate to semantic quality. This suggests that high-magnitude activations risk dominating the global steering direction if not properly normalized. To address this, ROAST employs grouped normalization to balance contributions across samples, ensuring a more robust estimation of the consensus steering direction. Across models (0.6B to 32B), ROAST consistently improves performance on diverse tasks (e.g., +9.7% on GSM8K for Qwen3-0.6B and +12.1% on TruthfulQA for GLM4-32B), and analyses show that CSS better preserves activation energy.

Traditional mechanized chestnut harvesting is too costly for small producers, non-selective, and prone to damaging nuts. Accurate, reliable detection of chestnuts on the orchard floor is crucial for developing low-cost, vision-guided automated harvesting technology. However, developing a reliable chestnut detection system faces challenges in complex environments with shading, varying natural light conditions, and interference from weeds, fallen leaves, stones, and other foreign on-ground objects, which have remained unaddressed. This study collected 319 images of chestnuts on the orchard floor, containing 6524 annotated chestnuts. A comprehensive set of 29 state-of-the-art real-time object detectors, including 14 in the YOLO (v11-13) and 15 in the RT-DETR (v1-v4) families at varied model scales, was systematically evaluated through replicated modeling experiments for chestnut detection. Experimental results show that the YOLOv12m model achieves the best mAP@0.5 of 95.1% among all the evaluated models, while the RT-DETRv2-R101 was the most accurate variant among RT-DETR models, with mAP@0.5 of 91.1%. In terms of mAP@[0.5:0.95], the YOLOv11x model achieved the best accuracy of 80.1%. All models demonstrate significant potential for real-time chestnut detection, and YOLO models outperformed RT-DETR models in terms of both detection accuracy and inference, making them better suited for on-board deployment. Both the dataset and software programs in this study have been made publicly available at https://github.com/AgFood-Sensing-and-Intelligence-Lab/ChestnutDetection.

Large language models (LLMs) have achieved remarkable success in generating fluent and contextually appropriate text; however, their capacity to produce genuinely creative outputs remains limited. This paper posits that this limitation arises from a structural property of contemporary LLMs: when provided with rich context, the space of future generations becomes strongly constrained, and the generation process is effectively governed by near-deterministic dynamics. Recent approaches such as test-time scaling and context adaptation improve performance but do not fundamentally alter this constraint. To address this issue, we propose Algebraic Quantum Intelligence (AQI) as a computational framework that enables systematic expansion of semantic space. AQI is formulated as a noncommutative algebraic structure inspired by quantum theory, allowing properties such as order dependence, interference, and uncertainty to be implemented in a controlled and designable manner. Semantic states are represented as vectors in a Hilbert space, and their evolution is governed by C-values computed from noncommutative operators, thereby ensuring the coexistence and expansion of multiple future semantic possibilities. In this study, we implement AQI by extending a transformer-based LLM with more than 600 specialized operators. We evaluate the resulting system on creative reasoning benchmarks spanning ten domains under an LLM-as-a-judge protocol. The results show that AQI consistently outperforms strong baseline models, yielding statistically significant improvements and reduced cross-domain variance. These findings demonstrate that noncommutative algebraic dynamics can serve as a practical and reproducible foundation for machine creativity. Notably, this architecture has already been deployed in real-world enterprise environments.

Multimodal foundation models have demonstrated impressive generalization capabilities, yet efficiently adapting them to new tasks in a few-shot setting remains a critical challenge. In this work, we investigate the few-shot adaptation of Large Audio-Language Models (ALMs) through both training-based and training-free approaches. We introduce MUKA, a multi-kernel adaptation framework that combines the fine-grained, context-dependent representations of instruction-tuning based models like Pengi with the global semantic representations of contrastive pretraining methods like CLAP. By constructing a product kernel that aligns local similarity with global semantics, MUKA enhances representational power while preserving the theoretical guarantees of kernel methods and avoiding additional training. Extensive experiments across 11 diverse audio datasets demonstrate that MUKA achieves state-of-the-art performance among training-free methods and even surpasses training-based adapters in several scenarios, offering a compelling balance between adaptability and efficiency.

Single-image 3D reconstruction with large reconstruction models (LRMs) has advanced rapidly, yet reconstructions often exhibit geometric inconsistencies and misaligned details that limit fidelity. We introduce GeoFusionLRM, a geometry-aware self-correction framework that leverages the model's own normal and depth predictions to refine structural accuracy. Unlike prior approaches that rely solely on features extracted from the input image, GeoFusionLRM feeds back geometric cues through a dedicated transformer and fusion module, enabling the model to correct errors and enforce consistency with the conditioning image. This design improves the alignment between the reconstructed mesh and the input views without additional supervision or external signals. Extensive experiments demonstrate that GeoFusionLRM achieves sharper geometry, more consistent normals, and higher fidelity than state-of-the-art LRM baselines.

Sparse Autoencoders (SAEs) have emerged as a promising tool for interpreting neural networks by decomposing their activations into sparse sets of human-interpretable features. Recent work has introduced multiple SAE variants and successfully scaled them to frontier models. Despite much excitement, a growing number of negative results in downstream tasks casts doubt on whether SAEs recover meaningful features. To directly investigate this, we perform two complementary evaluations. On a synthetic setup with known ground-truth features, we demonstrate that SAEs recover only $9\%$ of true features despite achieving $71\%$ explained variance, showing that they fail at their core task even when reconstruction is strong. To evaluate SAEs on real activations, we introduce three baselines that constrain SAE feature directions or their activation patterns to random values. Through extensive experiments across multiple SAE architectures, we show that our baselines match fully-trained SAEs in interpretability (0.87 vs 0.90), sparse probing (0.69 vs 0.72), and causal editing (0.73 vs 0.72). Together, these results suggest that SAEs in their current state do not reliably decompose models' internal mechanisms.

Predicting flows that occur both through and around porous bodies is challenging due to coupled physics across fluid and porous regions and the need to generalize across diverse geometries and boundary conditions. We address this problem using two Physics Informed learning approaches: Physics Informed PointNets (PIPN) and Physics Informed Geometry Aware Neural Operator (P-IGANO). We enforce the incompressible Navier Stokes equations in the free-flow region and a Darcy Forchheimer extension in the porous region within a unified loss and condition the networks on geometry and material parameters. Datasets are generated with OpenFOAM on 2D ducts containing porous obstacles and on 3D windbreak scenarios with tree canopies and buildings. We first verify the pipeline via the method of manufactured solutions, then assess generalization to unseen shapes, and for PI-GANO, to variable boundary conditions and parameter settings. The results show consistently low velocity and pressure errors in both seen and unseen cases, with accurate reproduction of the wake structures. Performance degrades primarily near sharp interfaces and in regions with large gradients. Overall, the study provides a first systematic evaluation of PIPN/PI-GANO for simultaneous through-and-around porous flows and shows their potential to accelerate design studies without retraining per geometry.

Precise in-hand manipulation of force-sensitive objects typically requires judicious coordinated force planning as well as accurate contact force feedback and control. Unlike multi-arm platforms with gripper end effectors, multi-fingered hands rely solely on fingertip point contacts and are not able to apply pull forces, therefore poses a more challenging problem. Furthermore, calibrated torque sensors are lacking in most commercial dexterous hands, adding to the difficulty. To address these challenges, we propose a dual-layer framework for multi-finger coordination, enabling high-precision manipulation of force-sensitive objects through joint control without tactile feedback. This approach solves coordinated contact force planning by incorporating graph rigidity and force closure constraints. By employing a force-to-position mapping, the planned force trajectory is converted to a joint trajectory. We validate the framework on a custom dexterous hand, demonstrating the capability to manipulate fragile objects-including a soft yarn, a plastic cup, and a raw egg-with high precision and safety.

Whether neural networks can serve as cognitive models of morphological learning remains an open question. Recent work has shown that encoder-decoder models can acquire irregular patterns, but evidence that they generalize these patterns like humans is mixed. We investigate this using the Spanish \emph{L-shaped morphome}, where only the first-person singular indicative (e.g., \textit{pongo} `I put') shares its stem with all subjunctive forms (e.g., \textit{ponga, pongas}) despite lacking apparent phonological, semantic, or syntactic motivation. We compare five encoder-decoder transformers varying along two dimensions: sequential vs. position-invariant positional encoding, and atomic vs. decomposed tag representations. Positional encoding proves decisive: position-invariant models recover the correct L-shaped paradigm clustering even when L-shaped verbs are scarce in training, whereas sequential positional encoding models only partially capture the pattern. Yet none of the models productively generalize this pattern to novel forms. Position-invariant models generalize the L-shaped stem across subjunctive cells but fail to extend it to the first-person singular indicative, producing a mood-based generalization rather than the L-shaped morphomic pattern. Humans do the opposite, generalizing preferentially to the first-person singular indicative over subjunctive forms. None of the models reproduce the human pattern, highlighting the gap between statistical pattern reproduction and morphological abstraction.

As robots become increasingly integrated into everyday tasks, their ability to perceive both the shape and properties of objects during in-hand manipulation becomes critical for adaptive and intelligent behavior. We present SemanticFeels, an extension of the NeuralFeels framework that integrates semantic labeling with neural implicit shape representation, from vision and touch. To illustrate its application, we focus on material classification: high-resolution Digit tactile readings are processed by a fine-tuned EfficientNet-B0 convolutional neural network (CNN) to generate local material predictions, which are then embedded into an augmented signed distance field (SDF) network that jointly predicts geometry and continuous material regions. Experimental results show that the system achieves a high correspondence between predicted and actual materials on both single- and multi-material objects, with an average matching accuracy of 79.87% across multiple manipulation trials on a multi-material object.

Existing Multimodal Large Language Models (MLLMs) for image forgery detection and localization predominantly operate under a text-centric Chain-of-Thought (CoT) paradigm. However, forcing these models to textually characterize imperceptible low-level tampering traces inevitably leads to hallucinations, as linguistic modalities are insufficient to capture such fine-grained pixel-level inconsistencies. To overcome this, we propose ForgeryVCR, a framework that incorporates a forensic toolbox to materialize imperceptible traces into explicit visual intermediates via Visual-Centric Reasoning. To enable efficient tool utilization, we introduce a Strategic Tool Learning post-training paradigm, encompassing gain-driven trajectory construction for Supervised Fine-Tuning (SFT) and subsequent Reinforcement Learning (RL) optimization guided by a tool utility reward. This paradigm empowers the MLLM to act as a proactive decision-maker, learning to spontaneously invoke multi-view reasoning paths including local zoom-in for fine-grained inspection and the analysis of invisible inconsistencies in compression history, noise residuals, and frequency domains. Extensive experiments reveal that ForgeryVCR achieves state-of-the-art (SOTA) performance in both detection and localization tasks, demonstrating superior generalization and robustness with minimal tool redundancy. The project page is available at https://youqiwong.github.io/projects/ForgeryVCR/.

Monitoring chain-of-thought (CoT) reasoning is a foundational safety technique for large language model (LLM) agents; however, this oversight is compromised if models learn to conceal their reasoning. We explore the potential for steganographic CoT -- where models hide secret reasoning within innocuous text -- to inform risk assessment and deployment policies. We systematically evaluate the limits of steganographic capabilities across 28 models, ranging from past generations to the current frontier. We measure monitor evasion, refusal rates, encoding fidelity, and hidden task accuracy across four datasets, comparing steganographic acrostics against plain reasoning and filler-token baselines. We find that current models cannot yet sustain hidden reasoning for complex math and arithmetic tasks. However, in a simplified counting experiment, Claude Opus 4.5 achieved 92% accuracy on the hidden task, demonstrating nascent capability. Notably, in rare cases (<1%), GPT-5.2 might refuse steganographic instructions while simultaneously complying with them. Our findings underscore the need for continuous evaluation of steganographic risks. This study provides a methodology to preemptively detect and prevent hidden reasoning that might empower misaligned scheming and deceptive behavior.

Post-training GUI agents in interactive environments is critical for developing generalization and long-horizon planning capabilities. However, training on real-world applications is hindered by high latency, poor reproducibility, and unverifiable rewards relying on noisy visual proxies. To address the limitations, we present GUI-GENESIS, the first framework to automatically synthesize efficient GUI training environments with verifiable rewards. GUI-GENESIS reconstructs real-world applications into lightweight web environments using multimodal code models and equips them with code-native rewards, executable assertions that provide deterministic reward signals and eliminate visual estimation noise. Extensive experiments show that GUI-GENESIS reduces environment latency by 10 times and costs by over $28,000 per epoch compared to training on real applications. Notably, agents trained with GUI-GENESIS outperform the base model by 14.54% and even real-world RL baselines by 3.27% on held-out real-world tasks. Finally, we observe that models can synthesize environments they cannot yet solve, highlighting a pathway for self-improving agents.

We study Neural Optimal Transport in infinite-dimensional Hilbert spaces. In non-regular settings, Semi-dual Neural OT often generates spurious solutions that fail to accurately capture target distributions. We analytically characterize this spurious solution problem using the framework of regular measures, which generalize Lebesgue absolute continuity in finite dimensions. To resolve ill-posedness, we extend the semi-dual framework via a Gaussian smoothing strategy based on Brownian motion. Our primary theoretical contribution proves that under a regular source measure, the formulation is well-posed and recovers a unique Monge map. Furthermore, we establish a sharp characterization for the regularity of smoothed measures, proving that the success of smoothing depends strictly on the kernel of the covariance operator. Empirical results on synthetic functional data and time-series datasets demonstrate that our approach effectively suppresses spurious solutions and outperforms existing baselines.

Large Language Models (LLMs) have empowered autonomous agents to handle complex web navigation tasks. While recent studies integrate tree search to enhance long-horizon reasoning, applying these algorithms in web navigation faces two critical challenges: sparse valid paths that lead to inefficient exploration, and a noisy context that dilutes accurate state perception. To address this, we introduce Plan-MCTS, a framework that reformulates web navigation by shifting exploration to a semantic Plan Space. By decoupling strategic planning from execution grounding, it transforms sparse action space into a Dense Plan Tree for efficient exploration, and distills noisy contexts into an Abstracted Semantic History for precise state awareness. To ensure efficiency and robustness, Plan-MCTS incorporates a Dual-Gating Reward to strictly validate both physical executability and strategic alignment and Structural Refinement for on-policy repair of failed subplans. Extensive experiments on WebArena demonstrate that Plan-MCTS achieves state-of-the-art performance, surpassing current approaches with higher task effectiveness and search efficiency.

The generation of classical Chinese \textit{Ci} poetry, a form demanding a sophisticated blend of structural rigidity, rhythmic harmony, and artistic quality, poses a significant challenge for large language models (LLMs). To systematically evaluate and advance this capability, we introduce \textbf{C}hinese \textbf{Ci}pai \textbf{V}ariants (\textbf{CCiV}), a benchmark designed to assess LLM-generated \textit{Ci} poetry across these three dimensions: structure, rhythm, and quality. Our evaluation of 17 LLMs on 30 \textit{Cipai} reveals two critical phenomena: models frequently generate valid but unexpected historical variants of a poetic form, and adherence to tonal patterns is substantially harder than structural rules. We further show that form-aware prompting can improve structural and tonal control for stronger models, while potentially degrading weaker ones. Finally, we observe weak and inconsistent alignment between formal correctness and literary quality in our sample. CCiV highlights the need for variant-aware evaluation and more holistic constrained creative generation methods.

Standard factuality evaluations of LLMs treat all errors alike, obscuring whether failures arise from missing knowledge (empty shelves) or from limited access to encoded facts (lost keys). We propose a behavioral framework that profiles factual knowledge at the level of facts rather than questions, characterizing each fact by whether it is encoded, and then by how accessible it is: cannot be recalled, can be directly recalled, or can only be recalled with inference-time computation (thinking). To support such profiling, we introduce WikiProfile, a new benchmark constructed via an automated pipeline with a prompted LLM grounded in web search. Across 4 million responses from 13 LLMs, we find that encoding is nearly saturated in frontier models on our benchmark, with GPT-5 and Gemini-3 encoding 95--98% of facts. However, recall remains a major bottleneck: many errors previously attributed to missing knowledge instead stem from failures to access it. These failures are systematic and disproportionately affect long-tail facts and reverse questions. Finally, we show that thinking improves recall and can recover a substantial fraction of failures, indicating that future gains may rely less on scaling and more on methods that improve how models utilize what they already encode.

Despite their success, large pretrained vision models remain vulnerable to catastrophic forgetting when adapted to new tasks in class-incremental settings. Parameter-efficient fine-tuning (PEFT) alleviates this by restricting trainable parameters, yet most approaches still rely on cross-entropy (CE) loss, a surrogate for the 0-1 loss, to learn from new data. We revisit this choice and revive the true objective (0-1 loss) through a reinforcement learning perspective. By formulating classification as a one-step Markov Decision Process, we derive an Expected Policy Gradient (EPG) method that directly minimizes misclassification error with a low-variance gradient estimation. Our analysis shows that CE can be interpreted as EPG with an additional sample-weighting mechanism: CE encourages exploration by emphasizing low-confidence samples, while EPG prioritizes high-confidence ones. Building on this insight, we propose adaptive entropy annealing (aEPG), a training strategy that transitions from exploratory (CE-like) to exploitative (EPG-like) learning. aEPG-based methods outperform CE-based methods across diverse benchmarks and with various PEFT modules. More broadly, we evaluate various entropy regularization methods and demonstrate that lower entropy of the output prediction distribution enhances adaptation in pretrained vision models.

Large, openly licensed speech datasets are essential for building automatic speech recognition (ASR) systems, yet many widely spoken languages remain underrepresented in public resources. Pashto, spoken by more than 60 million people, has historically lacked large-scale openly licensed speech data suitable for modern ASR development. This paper presents a release-level analysis of the Pashto component of the Mozilla Common Voice corpus, focusing on version 24.0 (December 2025) and contextualizing trends across major releases. We document rapid growth from 1.49 recorded hours in mid-2023 to 2,768.7 total hours in 2025, including 975.89 validated hours available for supervised ASR training. Beyond scale, we analyze validation throughput, contributor participation inequality, demographic metadata completeness, and sentence-level concentration in the validated subset. We find that participation is extremely concentrated (Gini = 0.941), age representation is strongly skewed toward young adults, and 41.97\% of clips lack self-reported gender labels, limiting subgroup auditing based on metadata. At the textual level, prompt reuse is moderate: 35.88\% of unique sentences account for 50\% of validated clips, suggesting that structural concentration is driven primarily by uneven contributor activity rather than dominance of a small prompt set. These results provide a quantitative audit of a rapidly scaling low-resource speech corpus and highlight practical priorities for improving dataset maturity, including expanded validation capacity and broader demographic participation.

We introduce LM-Lexicon, an innovative definition modeling approach that incorporates data clustering, semantic expert learning, and model merging using a sparse mixture-of-experts architecture. By decomposing the definition modeling task into specialized semantic domains, where small language models are trained as domain experts, LM-Lexicon achieves substantial improvements (+7% BLEU score compared with the prior state-of-the-art model) over existing methods on five widely used benchmarks. Empirically, we demonstrate that 1) the clustering strategy enables fine-grained expert specialization with nearly 10% improvement in definition quality; 2) the semantic-aware domain-level routing mechanism achieves higher expert efficacy (+1%) than conventional token-level routing; and 3) further performance gains can be obtained through test-time compute and semantic expert scaling. Our work advances definition modeling while providing insights into the development of efficient language models for semantic-intensive applications.

Code generation remains a challenging task that requires precise and structured reasoning. Existing Test Time Scaling (TTS) methods, including structured tree search, have made progress in exploring reasoning paths but still face two major challenges: (1) underthinking, where reasoning chains tend to be shallow and fail to capture the full complexity of problems; and (2) overthinking, where overly verbose reasoning leads to inefficiency and increased computational costs. To address these issues, we propose LogitsCoder, a novel framework that enhances chain-of-thought reasoning through lightweight, logit-level control mechanisms for code generation. LogitsCoder iteratively generates and refines reasoning steps by first steering token selection toward statistically preferred patterns via Logits Preference Decoding, then selecting and aggregating diverse reasoning paths using Logits Rank Based Path Selection and Thoughts Aggregation. This results in coherent and effective reasoning chains that balance depth and efficiency. Extensive experiments demonstrate that LogitsCoder produces more efficient and higher-quality reasoning chains, leading to superior code generation performance compared to baseline methods.

Decentralized federated learning (DFL) enables edge devices to collaboratively train models through local training and fully decentralized device-to-device (D2D) model exchanges. However, these energy-intensive operations often rapidly deplete limited device batteries, reducing their operational lifetime and degrading the learning performance. To address this limitation, we apply energy harvesting technique to DFL systems, allowing edge devices to extract ambient energy and operate sustainably. We first derive the convergence bound for wireless DFL with energy harvesting, showing that the convergence is influenced by partial device participation and transmission packet drops, both of which further depend on the available energy supply. To accelerate convergence, we formulate a joint device scheduling and power control problem and model it as a multi-agent Markov decision process (MDP). Traditional MDP algorithms (e.g., value or policy iteration) require a centralized coordinator with access to all device states and exhibit exponential complexity in the number of devices, making them impractical for large-scale decentralized networks. To overcome these challenges, we propose a fully decentralized policy iteration algorithm that leverages only local state information from two-hop neighboring devices, thereby substantially reducing both communication overhead and computational complexity. We further provide a theoretical analysis showing that the proposed decentralized algorithm achieves asymptotic optimality. Finally, comprehensive numerical experiments on real-world datasets are conducted to validate the theoretical results and corroborate the effectiveness of the proposed algorithm.

Length generalization is the ability of language models to maintain performance on inputs longer than those seen during pretraining. In this work, we introduce a simple yet powerful position encoding (PE) strategy, Random Float Sampling (RFS), that generalizes well to lengths unseen during pretraining or fine-tuning. In particular, instead of selecting position indices from a predefined discrete set, RFS uses randomly sampled continuous values, thereby avoiding out-of-distribution (OOD) issues on unseen lengths by exposing the model to diverse indices during training. Since assigning indices to tokens is a common and fundamental procedure in widely used PEs, the advantage of RFS can easily be incorporated into, for instance, the absolute sinusoidal encoding, RoPE, and ALiBi. Experiments corroborate its effectiveness by showing that RFS results in superior performance in length generalization tasks as well as zero-shot commonsense reasoning benchmarks.