Papers

Despite their widespread use, Large Language Models (LLMs) remain limited by a fundamental problem: the generation of plausible but false content, known as hallucinations. Most existing detection methods operate at the answer or sentence level, yet per-token detection is essential for localizing hallucinated spans and enabling fine-grained interventions. In this paper, we explore the use of the Mixture-of-Experts (MoE) paradigm to address this gap. In MoE architectures, a single forward pass activates a sparse subset of experts (i.e., distinct feedforward networks per layer) via a routing mechanism, producing internal signals (e.g., router entropy, expert disagreement, and expert usage patterns) that are unavailable in dense architectures and have not been previously exploited for hallucination detection. To this end, we introduce InnerExpert, the first method to leverage these MoE-specific signals for per-token hallucination detection. InnerExpert combines routing-level and standard transformer signals into compact per-token feature vectors, classified by a lightweight detector trained on labels produced by an LLM-as-a-judge pipeline, which enables continuous model updates without manual annotation. Our results show that InnerExpert outperforms existing methods across five datasets and two MoE architectures, achieving up to 0.91 answer-level and 0.76 token-level AUROC, while requiring only a single forward pass.

Self-evolving agents turn experience into reusable skills, workflows, or memories, but post-evolution accuracy alone does not show whether learned behavior preserves previously correct behavior or security. We audit SkillOpt, Agent Workflow Memory (AWM), and ReasoningBank in simulated e-banking using matched benign acquisition trajectories, sealed evaluation endpoints, execution-grounded checks, and independent state replay. On Qwen 3.7 Flash, SkillOpt raises benign utility from 0.741 to 0.837 while exposure to injected content rises from 0.820 to 0.943. Conditional attack success after exposure falls from 0.605 to 0.562, yet overall attack success rate (ASR) rises from 0.496 to 0.530 and unauthorized financial state changes rise to 0.685. Across three independently evolved lineages, capability, exposure, and unauthorized-state changes increase in all three, whereas ASR increases in only two. ReasoningBank raises utility to 0.859 without increasing aggregate ASR, although unauthorized state changes remain slightly above Static. AWM reveals a separate evaluation hazard: a literal WebArena text-action envelope disrupts tool execution in our native function-calling executor. In a post-hoc sensitivity test, removing only that envelope restores utility from 0.319 to 0.756, while exposure rises from 0.299 to 0.909 and ASR from 0.195 to 0.575. Auditing self-evolving financial agents therefore requires tracking regressions, attack-surface contact, unauthorized financial-state change, and artifact-executor compatibility, not accuracy alone.

Automated security patch backporting is critical for mitigating N-day vulnerabilities. Recent tools report success rates above 80% on their respective datasets. However, these evaluations are often confined to homogeneous environments, such as one repository or specific project versions. Consequently, it remains unclear how well these tools generalize beyond their originally targeted scenarios. We present Porting Benchmark, a curated dataset of 1,234 security patch backporting cases spanning cross-version, cross-branch, and cross-repository scenarios, paired with a common evaluation framework. Using this benchmark, we evaluate five tools spanning program analysis, LLM prompting, and LLM agents under aligned settings. Our results show that aligned evaluation changes the apparent performance landscape: PortGPT and TSBPort remain comparatively strong on the Replication Dataset, while FixMorph and Mystique degrade substantially under the common protocol. Performance degrades sharply on structurally complex patches: the best commit-level success rate falls from 85.2% on Type-I patches to 24.0% on Type-IV. We identify four root-cause categories (missing target API awareness, cross-version semantic mismatch, non-local dependency propagation failure, and patch construction or localization failure) and derive concrete directions for next-generation tool design. On a 45-case dynamically validated subset with verified test cases and constructed POCs, we further observe that reference-based benchmark scores do not fully capture real-world remediation: exact match sharply under-credits harder target adaptations, while executable validation reveals residual integration failures in the target that static reference agreement misses. Executable-feedback refinement provides limited but measurable recovery on the hardest executable cases.

LLM-driven agents can autonomously exchange opinions on online platforms and form communities. Such agent-operated social platforms raise a new security concern: attackers may manipulate agents to induce group polarization. Existing methods manipulate agent prompts or construct echo chambers, both of which are difficult to realize in practice. We therefore formulate a new threat, Memory-Mediated Polarization Cascade, which uses agent memory as a persistence channel and public discussion as a propagation channel. This threat contains three stages. During exposure and memory retention, the attacker exposes a small set of target agents to arguments that reinforce their respective stated stances. The targets' memory systems then process and retain these arguments. During retrieval and reproduction, a shared stance-neutral discussion cues the targets to retrieve and reproduce their respective retained arguments. During iterative propagation, untreated agents influenced by the reproduced arguments restate and spread them. We instantiate this threat in GraphWake with three components: (i) stance-support argumentation knowledge graphs construct knowledge-based arguments; (ii) axiom-oriented triple selection distills them for reliable retention and reproduction; and (iii) stance-neutral memory cueing triggers concurrent retrieval and reproduction, initiating propagation. Experiments across multiple discussions and memory systems show that GraphWake substantially increases group polarization. These findings reveal a community-level polarization risk.

LLM-powered GUI agents that autonomously operate smartphones are rapidly transitioning from research prototypes to early real-world deployment. However, because these agents routinely process untrusted environmental content, they are highly vulnerable to environmental injection attacks, which include indirect prompt injections and adversarial instructions. Such attacks can manipulate the behavior of agents without user awareness through diverse channels encountered in everyday mobile use. Despite these risks, existing benchmarks often fail to capture everyday user scenarios, lacking a systematic evaluation of GUI agents under environmental injection attacks on mobile devices. To address this gap, we introduce MobileWorldSafety, a benchmark of 142 risk tasks built on real Android applications. For each task, we define a programmatically verifiable risk indicator over the final system state and evaluate outcomes with a two-stage pipeline: rule-based verification handles unambiguous cases, while an LLM judge adjudicates ambiguous ones. This distinguishes safety failures from capability failures and enables objective and reproducible assessment. Evaluations on six agents, including both general agents and specialized GUI agents, demonstrate that all agents remain highly vulnerable, with attack success rates ranging from 40.4% to 66.9%. These findings indicate that current agents often fail to maintain safety alignment when adversarial content is presented as ordinary mobile context. MobileWorldSafety provides a foundation for quantifying these vulnerabilities and advancing research on robust mobile GUI agents.

Agents increasingly interpret a person's natural-language preferences by querying an LLM for numerical preference judgments, e.g., by asking how much the person would be willing to pay for an item. A growing body of work estimates a utility function from these judgments and then chooses actions based on their estimated utility. This pipeline assumes the judgments are approximately self-consistent: that a single utility function can reproduce them. But are they? To study this question, we measure the self-consistency of cardinal LLM preference judgments. For example, the difference in stated willingness-to-pay between two items should match the stated payment that makes a person indifferent to exchanging them. We develop statistical tests and interpretable measures of how far observed responses depart from the best-fitting self-consistent utility function. Experiments with flight, apartment, and hotel examples across six LLMs reveal large persistent inconsistencies. This suggests that LLM-derived preference judgments cannot be faithfully summarized by a single utility function.

What a reasoning model writes is only a partial record of the process that produces it. We introduce a two-level internal readout for mixture-of-experts reasoning. We first distill vocabulary-scale J-space into J64, a 64-axis semantic frame learned from the model's own reasoning states. J64 reveals readable process state that the emitted trace does not show: it separates inference effort from problem-induced strain. It also adds 0.096 to 0.135 held-out AUC over a baseline that reads the same rollout as token occupancy and aggregates it in exactly the same way. We then reconstruct J64 from native expert-routing statistics. The result is R64, a low-overhead proxy: its median per-axis correlation with J64 is 0.69 to 0.86 across three models and two families, and on gpt-oss-20b it preserves 95 to 100% of J64's predictive gain. The readout supports test-time decisions at two temporal resolutions. Over completed candidate sets, J64 and R64 improve single-branch selection, and R64-weighted voting improves plain majority voting in seven of eight settings. During generation, rolling readout windows drive a cumulative stop-and-resample policy whose operating point is fixed on training questions alone. J64 improves accuracy by 1.1 to 5.9 points over a sibling-permuted control, and the routing-only R64 proxy retains 0.9 to 3.2 of those points. Finally, router edits aimed at the mechanism J64 names induce the predicted reasoning behaviors and shift a diagnosed stall from numerical guessing toward exact symbolic execution. Together, J64 makes latent process state readable, while routing makes it deployable and actionable.

The $\operatorname{do}$-operator is described graphically by deleting arrows into its targets and functionally by replacing their mechanisms with constants. To call these operations equivalent is not yet a mathematical statement: one returns a graph and remembers only the targets, whereas the other returns mechanisms and also remembers the imposed values. We make a dependency-level comparison precise for deterministic acyclic structural causal models with finitely many endogenous variables. If $\operatorname{Graph}(F)$ extracts the dependencies of a mechanism family $F$, our main theorem is $\operatorname{Graph}(F^\iota)=\operatorname{Surg}(\operatorname{Graph}(F),T_\iota)$. Thus replacing target mechanisms removes exactly the dependencies removed by graph surgery. For a model $M=(G,F)$ whose graph may contain unused arrows, we characterize when the same equality holds with $G$ in place of $\operatorname{Graph}(F)$; it holds for every intervention exactly when $G$ records the dependencies of $F$ exactly. We then define the intervened model, characterize its run, show how sequential interventions combine, and prove that an outcome depends only on interventions at its actual dependency ancestors. All principal results are machine-checked in an accompanying Lean 4 development.

Large language models (LLMs) can both expand underspecified queries and encode text as dense representations, suggesting a unified model for query expansion and retrieval. Existing systems usually rely on prompted expansions, independently trained modules, or staged optimization, leaving generated expansions only indirectly aligned with the retrieval loss that judges them. We train a single decoder-only LLM end to end, where the same model generates the expansion and encodes both the expanded query and candidate documents. This unified setting creates a moving-target problem: retrieval supervision should improve query-side expansion, but the same update also shifts the document embeddings that serve as retrieval targets. We introduce Document Embedding Preservation Tuning (DEPT), which keeps tuned document embeddings close to cached initial embeddings while allowing retrieval gradients to pass through straight-through decoding into the generator. DEPT converts joint query–document movement into query-side adaptation against approximately stable, whitened document embeddings that support index reuse and online hard-negative mining. Experiments with Qwen3-4B-Instruct-2507 and LLaMA-3.2-3B-Instruct on five datasets in BEIR benchmark show that DEPT improves average retrieval quality over training-free, independently trained, and staged unified baselines, while ablations isolate the effects of preservation, whitening, end-to-end expansion training, and online negatives. Code is available at https://github.com/ILSparkle/DEPT.

Developing robots capable of understanding and manipulating objects requires compact, interpretable, and generalizable representations. This work proposes a reinforcement learning-based framework for robotic grasp refinement, integrating keypoint-based object representations with a Deep Q-Network (DQN). Using 2D overhead images captured in a simulated environment, a geometric-based algorithm generates initial grasp candidates, which are iteratively refined by the proposed framework, transforming failed grasps into successful ones. Experiments conducted on 300 objects from the Dex-Net dataset using a UR5 manipulator demonstrate the framework's effectiveness, achieving a 100% success rate on objects previously deemed ungraspable by geometrical methods. The framework's sim-to-real transferability is further validated through physical experiments on a Delta parallel robot, where a refined grasp successfully manipulates an object that was previously ungraspable. The findings underscore the effectiveness of reinforcement learning in addressing challenges in robotic grasping, offering a scalable and adaptable solution for contact-rich manipulation tasks.

Saudi Arabia will host the 2034 FIFA World Cup and already operates crowd management at Hajj scale. Drone-based counting must hold accuracy on footage unlike anything in its training corpus, without labels, and must warn of dangerous inflow before a crush forms. We deliver a validated answer built on 525 controlled runs, a full-resolution corpus study, five falsification ablations, and a five-condition safety-interlock evaluation. Label-free adaptation recovers 31-49% of shift-induced error across four corruptions and five severities, with the strongest method gaining 41.8 MAE over the frozen source (95% CI [34.1, 49.6], p=7.5x10^-10, d=2.52). We establish a severity law separating methods with a constant absolute margin from the one whose margin grows, and a stability budget identifying which configuration is safe to fly. On a full-resolution corpus carrying a genuine +48 MAE aerial gap (source retrained to 14.6 validation MAE, a 34% improvement), adaptation repairs the dense-scene undercounting that would otherwise under-report a forming crush, and the flux-based risk module fires on real congestion episodes in 2 of 6 full-length clips. We localise the recoverable error: in a regime built to favor a physics-informed conservation prior (300-frame clips at 200ms spacing, five times wider than standard), the adaptation signal is normalisation-driven, not flow-driven; the continuity residual is invariant to the proportional counting errors domain shift produces, confirmed by four on/off ablations correlated at r=0.999 and a 40% input corruption moving accuracy by only 0.05 MAE. A label-free shift gate shows shift magnitude and accuracy damage are rank-independent (Spearman rho=0.20; rho=-0.60 among genuine shifts), quantifying the 58% of headroom a magnitude gate forgoes. We establish unconditional adaptation with tail monitoring as policy, closing with a six-point protocol.

Learning analytics models can identify students at risk of poor performance, but they do not directly indicate which interventions are feasible, actionable, and compatible with educational constraints. This paper introduces SC2R, a semantics-constrained counterfactual recourse framework for educational decision support. SC2R combines a calibrated predictive model, integer-programming-based recourse generation over discrete action variables, a lightweight RDF vocabulary for intervention-plan representation, and SHACL validation for enforcing timing, budget, immutability, and availability constraints. The framework is evaluated offline on the OULAD dataset using snapshots constructed relative to each assessment at two decision horizons. Results show that the predictive component provides strong performance, that compact intervention plans can be generated at scale, and that semantic validation reveals infeasible plans that lighter optimization-only settings would otherwise accept. Rather than claiming causal improvement in student outcomes, this work shows that counterfactual recourse becomes more operationally meaningful in education when recommendations are not only model-valid, but also semantically feasible and machine-checkable.

We present MoNe, a lightweight modular neural memory that attaches to any frozen pretrained Transformer to enable long-context inference without retraining. MoNe reads context in fixed-size segments via test-time learning of fast-weight neural memory networks with layer-localized gradient updates; at inference, the memory generates keys and values from the query tokens alone, with no context tokens re-read. This two-phase design decouples inference cost from context length, achieving $O(N)$ preprocessing and $O(1)$ query cost with peak GPU memory that does not grow with $N$. At 128K tokens, MoNe reduces both compute and peak GPU memory by approximately 80% compared to ICL with only 6.4% parameter overhead. MoNe generalizes to context lengths far beyond the backbone's native window, achieving strong performance on needle-in-a-haystack and word extraction benchmarks from RULER, where ICL degrades sharply.

Conversational AI is moving beyond isolated text prompts toward sustained, multimodal interaction. In real conversations, users clarify goals, revise requests, interrupt responses, switch topics, and introduce new evidence while expecting systems to preserve context across turns. This makes multi-turn dialogue a distinct challenge requiring systems to maintain and update memory, ground responses across modalities, tools, and external knowledge, and adapt across languages and cultures. This study reviews multi-turn conversational AI across text-only dialogue, AudioLLMs and speech-native systems, multimodal and omni-modal systems, and tool-augmented agents. We organize the literature around datasets and benchmarks, modeling paradigms, training strategies, evaluation setups, and cross-cutting challenges. Our analysis shows that support for multiple modalities has advanced faster than the ability to sustain coherent interaction across a session. Despite stronger capabilities to perceive, speak, and act across modalities, current systems still struggle with persistent memory, cross-turn grounding, full-duplex interaction, robust evaluation, and cultural alignment. We conclude with a research agenda for systems that can remember, revise, ground, speak, listen, act, and adapt across turns, modalities, and cultures. (https://github.com/faiza-sfa/multiturn-conversational-ai-survey)

Large language models are increasingly deployed through agent harnesses that manage tools, extensions, persistent state, permissions, and external actions. Existing safety benchmarks mainly target individual attack mechanisms or a limited subset of operational settings, making it difficult to compare how safety failures emerge across different harness responsibilities. We present HarnessRisk, a lifecycle oriented benchmark that organizes agent harness safety into six operational phases including Harness Configuration, Capability Extension, Runtime Operation, State Persistence, Action Control, and Incident Recovery. HarnessRisk contains 128 sandboxed cases, each pairing a benign user objective with an adversarial instruction embedded in an untrusted workflow artifact. We evaluate each trajectory using Utility, Attack Success Rate, Persistence, and Detection. Across three harnesses, six language models, and 14 model and harness configurations, attack success ranges from 12.6% to 80.9%, while Utility remains between 75.0% and 97.6%. Harness Configuration is the most vulnerable phase across all three harnesses, showing that attacks can succeed by altering security sensitive parameters within otherwise authorized workflows. We also find that explicit risk recognition does not reliably lead to safe action, as some configurations detect risks in more than 90% of runs while retaining substantial attack success. These results highlight the need to evaluate agent safety across multiple harness responsibilities and at the level of the deployed model and harness configuration.

In this study, we investigate developmental mechanisms that enable small, resource-constrained systems such as cm-sized millirobots to autonomously explore, learn, and adapt their capabilities throughout their lifespan. Reinforcement learning algorithms guide the agent's skill acquisition and adaptation through the interplay of our proposed tiny Developmental Skill Method (tinyDSM), which integrates intrinsic motivation and fitness-based assessment. We strive for minimal hard-wired skills while encouraging the open-ended development of new skills. A key emphasis in our approach is to encode minimal a-priori general knowledge, which serves as a foundational starting point for the system as it further learns system-specific dependencies from the initial knowledge provided. Thus, by design, our approach aims to cover generic application domains. The methodology is based on (a) developmental mechanism with intrinsic motivation, and (b) a cognitive architecture (knowledge, reasoning, learning), while (c) utilizing minimal resources. It uses a hierarchical knowledge graph and kinematic reasoners to model and evaluate simple and advanced motion related skills. In our experiments, we use a resource-constrained millirobot with a volume of 36 cm^3 with a Raspberry Pi Pico 32-bit microcontroller that integrates all described features and capabilities except the camera system in 9 kB. Starting with learning the most elementary motor skills the millirobot autonomously progresses from simple linear and angular movements to complex geometric patterns within 15 minutes. To complement the physical experiments, we perform a simulation-based analysis that enables systematic comparisons across learning algorithms and intrinsic motivation parameters.

Agent Skills package reusable natural language procedures with executable resources, enabling software agents to acquire task specific capabilities without model adaptation. Automatically generating such Skills can improve task performance, yet evaluating a candidate solely from its artifact or final task outcome leaves unresolved which actions the equipped agent will perform and which side effects those actions will produce. We present TRUSS, an evidence guided framework for generating functionally effective and safety reliable Agent Skills. TRUSS first inspects functional claims against source and domain evidence while evaluating the complete artifact under nine predefined safety properties. Candidates admitted by this static gate are loaded by a shadow agent inside a Controllable Execution Environment, where brokered tools expose requested actions to policy enforcement and record their results as provenance preserving execution traces. Functional failures and property violations are linked back to the responsible Skill content and used to guide iterative refinement. We evaluate TRUSS on 168 SkillInject artifacts, 155 SkillSafetyBench cases, and all 187 tasks in SkillGenBench. TRUSS achieves 100.00% precision and recall in vulnerability detection. Repair reduces attack success from 38.71% to 19.35% with GPT 5.5 and from 46.45% to 29.68% with GPT 5.4, with zero attack regression. For Skill generation, TRUSS raises task effectiveness from 17.11% without Skills to 52.94%, while increasing the benchmark Security rate from 50.80% to 100.00%. These results show that execution evidence can expose behavioral failures missed by artifact inspection and can guide Skill generation toward jointly verified functional and safety outcomes.

How cautious should an agent be while it is still learning its environment? We propose RATTL (Risk-Adversarial Total-Reward Learning), which ties caution to epistemic uncertainty: the agent holds a Bayesian posterior over unknown dynamics and plans against a Wasserstein ambiguity set whose radius is a monotone function of that posterior. The radius contracts with evidence, so behaviour interpolates continuously between worst-case robustness and risk-neutral total-reward maximization. The design follows the duality underlying the Entropic Value-at-Risk, which converts the choice of a risk level into the choice of an ambiguity radius. We show the resulting planning problem is well posed under transience and compactness conditions, and prove a Safety Sandwich: the RATTL value lies between the uninformed robust value and the full- knowledge optimum, with a gap that vanishes as the posterior concentrates. In a canonical binary-hazard instance, the induced criterion reduces to Conditional Value-at-Risk at a level set by the posterior entropy. A worked example shows the agent deferring the efficient action until a sharp identification threshold. RATTL targets runtime safety for agents, including LLM-based systems, acting under uncertainty.

Motivation: Low-dimensional embeddings are widely used to explore cell-state heterogeneity in single-cell and other high-dimensional biological data. Although many methods preserve local neighborhoods, they may distort the apparent sampling density of processed observations, altering the visual contrast between dense and sparse regions and complicating the interpretation of rare, transitional, or continuous cell-state populations. Results: We present DMT-Dens, a parametric manifold-visualization method built on a latent-token Transformer encoder. The model integrates rank-based manifold alignment with hard-pair aggregation. To preserve density, it optimizes a loss based on the Pearson correlation between k-nearest-neighbor log-radius estimates in the processed input and two-dimensional embedding spaces. Benchmark evaluations demonstrate strong density preservation, particularly on biological datasets, while retaining competitive label separability. Availability: Source code, data-processing scripts, and resolved experiment configurations are available at https://github.com/Ruizhe-wang/DMT-Dens.

Pretrained molecular language models are increasingly used as molecular encoders for learning structure-property relationships. However, their practical suitability for molecular discovery within and beyond their pretraining domain remains unclear. Herein, we systematically benchmark four molecular language models across six virtual molecular libraries spanning drug discovery, organic materials, and catalysis. Native molecular language model embeddings show substantial variation in discovery performance across libraries, whereas molecular fingerprints provide a consistently strong and robust baseline. Consistent with a potential domain-representation mismatch, we show that explicit domain adaptation substantially improves representation performance. Fine-tuning molecular language model encoders on structures from the target virtual library consistently improves sample efficiency, with several adapted encoders emerging as the top-performing representations across the benchmark tasks. These results show that molecular representation quality depends strongly on the target domain and that explicit adaptation can improve the practical utility of molecular foundation models. More broadly, our findings establish domain-adapted molecular representations as a promising strategy for sample-efficient adaptive decision making in virtual screening and self-driving laboratories.

Unified multimodal models (UMMs) are motivated by the hope that understanding and generation reinforce each other but controlled ablations repeatedly find that adding a generation objective leaves understanding flat. Joint-training studies cannot settle the disagreement: with overlapping supervision, a gain cannot be attributed to the architecture rather than the data. To further investigate the relationship between the two directions in UMMs, we separate them by construction. A novel visual entity, a rendered 3D asset paired with a pseudo-word screened for absence from the frozen model's behavior, is bound through exactly one task direction, and the untrained direction is then measured. We find that the channel is real in both directions, but the directions differ in kind: generation training installs a name the model can only match among candidates; understanding training installs one it can also produce. What governs cross-task usability is where the binding enters the shared computation. An alignment probe predicts export across 36 configurations (Spearman $\rho = +0.68$). That objective's alignment term, maximized in closed form over activations with every weight frozen, makes a concept drawable when injected at layer 7 of 28 and is indistinguishable from the base model from layer 14 on, while the weight-based version of the same edit peaks at layers 10-14. In an observational series of four models, this window appears only where the understanding pathway is a semantic vision encoder, suggesting that unified weights are not enough: the two directions must share a semantic format at the entry point. Exploiting the rule, a mid-stack alignment objective acquires the concept for a $0.1\%$ relative loss of the model's general text-to-image ability, against $41\%$ for the standard generative route. Our code is at https://github.com/Zane-ZYQiu/entry-point-umm.

Joint-Embedding Predictive Architectures (JEPAs) learn world models by predicting future embeddings, but the objective admits a trivial solution of a constant encoder, so every practical system adds an anti-collapse mechanism (LeCun, 2022; Assran et al., 2023; Bardes et al., 2022; 2024). LeWorldModel (LeWM) prevents collapse with SIGReg, a regularizer that forces the latent distribution to match an isotropic Gaussian: the representation is stabilized by prescribing what it must look like, independently of the environment it models. We argue that the anti-collapse pressure can instead come from the transition data itself. Action-Contrastive Masked Transition Modeling (AC-MTM) keeps LeWM's forward latent-prediction objective and adds a training-only inverse-dynamics head trained with Action-NCE: each latent transition must identify the action that produced it among the other actions in the batch, a discrimination task that a collapsed encoder provably fails. The inverse branch is discarded after training, leaving test-time encoding, forward prediction, planning, and compute identical to LeWM. On four standard pixel-control tasks under a matched planning protocol, AC-MTM trains stably from scratch and matches SIGReg on average. On the harder multi-object OGBench Visual Scene task, results are consistent with the prescribed geometry becoming a bottleneck: AC-MTM reaches 80.0$\pm$2.0% success versus 58.0$\pm$2.0% for SIGReg, improving by 20-24 points in each training seed. A single 50-episode random-policy run gives a 52% baseline estimate. Contrastive inverse dynamics thus provides a distribution-free anti-collapse signal that requires no target network, stop-gradient, pretrained encoder, or reconstruction objective, and we characterize the action-space and observability assumptions under which it holds. We make our code available at https://github.com/jackboyla/action-contrastive-jepa

Legal consultation questions exhibit multi-level complexity. A single retrieval strategy often leads to over-reasoning for simple questions and poor interpretability for complex ones, making it difficult to meet the requirements for both answer quality and efficiency in high-risk scenarios. To address this issue, this paper proposes CoAL-RAG, a complexity-aware legal retrieval-augmented generation method, which constructs a multi-dimensional evaluation mechanism based on “question essence” and “retrieval consistency” to enable adaptive routing of retrieval strategies. First, the reasoning demand is quantified according to the logical structure of the question. Then, the discrepancy between semantic retrieval and keyword retrieval is utilized to indirectly reflect problem complexity, thereby selecting the most appropriate retrieval strategy and dynamically filtering contextual information. Experimental results demonstrate that the proposed method significantly outperforms baseline models not only on Chinese legal benchmarks (SocialLawQA, LawBench) but also demonstrates strong cross-jurisdictional generalization on English datasets (LexGLUE, CaseHold). Specifically, on Chinese datasets, the BLEU score improves by 42.5% and ROUGE-L reaches 3.6 times that of knowledge graph-based methods. On English benchmarks, CoAL-RAG maintains highly competitive accuracy, achieving an optimal balance between generation quality, deep logical reasoning, and system efficiency across different legal systems.

Continual pre-training of large language models must acquire new information without erasing old knowledge. Existing replay methods often choose a global old/new mixture and sample uniformly, ignoring that examples differ in how quickly they are forgotten. We formulate continual pre-training as adaptive review scheduling: the training loop should decide not only how much history to replay, but which examples should return at each step. We introduce Spaced Repetition Training (SRT), a continual learning framework inspired by cognitive science, which schedules sample-rehearsal using the SuperMemo-2 (SM-2) algorithm. SRT maintains per-example review state, maps per-example perplexity to a recall-quality signal, and schedules historical examples for retention and new examples for consolidation while leaving the model, objective, and optimizer unchanged. On temporally separated Wikipedia and code corpora, SRT improves the stability-plasticity trade-off, recovering 5 to 37 percentage points of old-knowledge accuracy lost by naive continual pre-training across model scales while preserving or improving new-knowledge acquisition. At larger scale, SRT preserves broad benchmark performance that naive continual pre-training and uniform replay substantially degrade. Experiments with vision and tabular data further suggest that the scheduling principle extends beyond language when paired with an appropriate recall signal.

Modern agents operate inside agent harnesses that manage tools, context, and control flow, making the harness a critical part of the agent system. Our original Agent Lightning introduced a disaggregated architecture that connects arbitrary agents to RL training through an LLM endpoint proxy, an approach later adopted by frameworks such as verl Uni-Agent, AReaL 2.0, slime, and Polar. We refer to this paradigm as harnessed agentic RL, where the deploy-time harness directly participates in model post-training. Harnessed agentic RL differs fundamentally from traditional agentic RL: the harness, rather than the training engine, owns the environment interaction loop, while the trainer observes only sequences of LLM request-response pairs. This introduces challenges in retokenization, sample merging, advantage calculation, loss normalization, and backend scheduling, which can substantially affect training stability and effectiveness. We present Agent Lightning v1.0, a lightweight framework for harnessed agentic RL implemented in approximately 3,500 lines of code. It supports arbitrary agent harnesses and serves as a practical testbed for studying these challenges. We evaluate it on instruction-following, search, and coding agents, and provide a complete reproducible pipeline for coding-agent RL. Using only 6K training examples and modest compute, RL improves Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%, a 14.6-point absolute gain. We release the complete workflow and training scripts to facilitate reproducible research on harnessed agentic RL.

Persistent shortages in the surgical workforce and inherent limitations of traditional training methods highlight the necessity of automated, data-driven approaches in surgical education. This study addresses these challenges by introducing a novel, explainable AI-powered framework for automated skill assessment, specifically focusing on cataract surgery. We present the world's largest dataset of cataract surgery videos, comprising 2,000 recordings. Additionally, we propose an AI-powered analytical framework that employs advanced computer vision and signal-processing techniques to automatically evaluate surgical videos to derive objective, quantitative performance indicators that complement or potentially replace subjective scoring methods. A significant advantage of our framework over previous methods lies precisely in its explainability of outputs, elevating it beyond merely an opaque skill classification tool. Through experimental analysis of 83 cataract surgery videos, we demonstrate that the automatically computed metrics exhibit strong correlations with expert-based subjective evaluations, achieving up to 87% accuracy in surgical skill assessment. Each metric was individually examined, and expert surgeons provided subjective ratings using the newly introduced Capsulorhexis Skill Assessment System (CSAS). These subjective assessments were compared with ten objective motion-based metrics extracted through our framework. The results indicated a robust correlation between subjective ratings and automated indicators, underscoring the framework's capacity to accurately model surgical expertise.

Background: Large Language Models (LLMs) are increasingly being applied to Software Engineering (SE) tasks, achieving high accuracy across problems such as clone detection, vulnerability prediction, and code summarization. However, their high computational demands and energy consumption raise sustainability concerns and hinder their use on consumer hardware and resource-constrained platforms. A common way to report the computational cost of an LLM in the literature and industry is to use the number of Floating Point Operations (FLOPs) required to perform a pass over the network. Aims: This paper investigates the implications of energy-aware knowledge distillation for SE, aiming to improve model efficiency while maintaining performance and to determine whether FLOPs is a reliable energy-aware metric. Method: We conduct a controlled experiment using Morph, a Many-Objective Optimization-based distillation methodology, to empirically examine whether FLOPs accurately reflect energy consumption in Clone Detection and Vulnerability Prediction tasks. We extend this methodology to include energy-surrogate models that directly estimate CPU and GPU energy consumption during optimization, and we apply Morph to generative tasks using CodeT5+ for code summarization. Results: Our results show that FLOPs is not always a reliable indicator of energy consumption, and better results can be achieved by using energy-surrogate models. Distilled student models can reduce inference energy consumption by up to 90% and memory usage by 86%, with only modest accuracy trade-offs. Conclusions: Energy-aware knowledge distillation when guided by direct energy surrogates rather than FLOPs can improve the energy consumption, sustainability, and deployability of LLMs for SE applications, enabling efficient models on consumer hardware.

Recent efforts toward fully automated AI scientists have demonstrated that language-model agents can generate hypotheses, execute experiments, and draft scientific manuscripts. However, during the early stages of research, when research problems are formulated, these AI scientists often rely heavily on proprietary frontier models. Their proposals are shaped by opaque parametric knowledge and by literature searches conditioned on the proposals themselves. Such knowledge is effectively a black box, and this dependence makes the evidential basis and validity of generated research problems difficult to audit and leaves the process vulnerable to model-specific hallucinations and biases. Furthermore, if proprietary research materials are transmitted to external APIs, the use of these models creates confidentiality, privacy, and data-governance concerns. We introduce the Structural Gap Hypothesis Agent (SGHA), a fully automated, corpus-first research-problem discovery system that runs entirely on a local LLM. SGHA structures a scientific literature corpus into evidence-linked paper objects and a typed evidence graph, detects unresolved structural patterns across papers, screens candidate gaps before formulation, and produces traceable research-problem families. In particular, it is able to output assumptions, objectives, success criteria, and remaining ambiguities. All LLM-based components of SGHA are executed using a locally served open-weight 9B language model, without requiring proprietary frontier-model APIs. We compare SGHA with the AI Scientist-v2 idea formulation module in five machine-learning domains. Our results suggest that explicit corpus structure and evidence-constrained reasoning can support promising, inspectable research-problem formulation without relying on frontier models during generation or verification.

User-facing tool agents must coordinate dialogue and tool use as user goals unfold over multiple turns. Yet interactive reinforcement learning typically reduces each rollout to a terminal reward, assigning the same credit to effective elicitation, errors, and later repair. The next user turn is more than context: it also provides noisy, temporally local evidence about the preceding user-to-user segment. We introduce Feedback-Aware Credit Assignment (FACA), which aligns each reaction with that segment, derives a locally normalized reaction advantage, and adds it to verified terminal outcome advantage without an extra critic or rollout. Against an outcome-only Interactive GRPO control matched in simulator, visible dialogue, initialization, rollout, and optimization, FACA improves the nine-domain $\tau$-family average across three independently trained runs by 5.91 and 10.22 percentage points at 8B and 14B, respectively. Gains concentrate in Telecom; at 8B, randomizing reaction polarity removes the Telecom gain. The same ordering holds zero-shot on Pare-Bench and Co-Gym. These results demonstrate that next-turn user reactions provide actionable local credit for improving multi-turn user-interacting agents.

Recent advances in LLM agents have made them increasingly capable of designing methods for complex AI tasks. This raises two central questions about agent-designed methods relative to human-designed methods: how well they perform, and how different their algorithmic designs are. To study these questions, this paper introduces an analysis that derives task-specific algorithmic design spaces from human-designed methods, maps both human- and agent-designed methods into these spaces, and quantifies their algorithmic differences at the module level. Widely used LLM agents are evaluated on a suite of representative, open-ended AI tasks spanning multiple modalities, and the methods they design are analyzed in terms of both task performance and algorithmic differences from human-designed methods. Experimental results show that current agents can occasionally match or surpass human state-of-the-art (SOTA) performance (10/72 configurations), but such success does not generalize reliably across tasks or agents. Moreover, 96.8% of agent-designed methods fall within human-derived algorithmic design spaces, largely recombining algorithmic choices found in human-designed methods, while nearly half exactly match an existing human algorithmic design. Taken together, these findings suggest that although current agents can occasionally match or surpass human SOTA performance, their algorithmic designs remain within human-derived algorithmic design spaces, reflecting the reuse and recombination of algorithmic choices.