每天知道
你的领域
出了什么

Reconstruct where the images provide evidence, and generate where they do not: recent success of spatial world models such as Atlas (World Labs Team, 2026) highlights the value of unifying reconstruction and generation in one model. Yet the two have long lived in separate paradigms with distinctive failure modes: feed-forward reconstruction averages ambiguity into blur, while conditional generation invents plausible but scene-inconsistent detail. In this work, we present a unified flow-based formulation for reconstruction and generation, where a shared clean-target predictor performs direct reconstruction at its single-step endpoint and unfolds conditional generation through multi-step flow. A controlled toy study reveals the mechanism: with a single step, the predictor collapses to the conditional mean just like feed-forward methods, favoring consistency over diversity. With multi-step inference, the fidelity of generated details grows with context richness: closer observations reduce ambiguity and yield better-matched details. We further instantiate the formulation in appearance and geometry 3D tasks. JiT-LVSM improves perceptual and distributional quality in novel view synthesis, while JUSt3R retains competitive single-step geometry prediction with additional multi-step inference capabilities that reduces veil and flying-pixel artifacts, producing cleaner surface structure with greater test-time compute. Together, they show that reconstruction and generation can share both a formulation and a backbone, with their behavior governed by denoising configuration---making unification surprisingly simple.

Deep learning has transformed symbolic music generation by borrowing the training paradigms of large language models, with systems such as NotaGen now producing complete, stylistically convincing classical scores from a short prompt. These systems could become powerful creative partners, helping musicians generate endless possibilities. However, current systems expose almost no control handles on the music itself. In principle, control handles could be built into a foundation model trained from scratch, but this is rarely practical without massive amounts of quality annotated data and compute resources. We therefore present MotiGen, a recipe for retrofitting pretrained symbolic music models to use new instruction prompts. MotiGen injects a musical motif as a structured prompt line, reinforces it with a scalar attention bias toward the motif tokens, and learns the association with a two-phase curriculum. First, it learns from focused excerpts cropped around motif occurrences in the training data, then full scores including the motifs. Our experiments show that our model composes with the prompted motif in over 92.3% of generated pieces. Generated pieces using a variety of motifs are included in our sample site: https://motigen-site.github.io/.

Latent generative models make 3D medical image synthesis computationally practical by generating in a compressed space. However, we show that the common flat Euclidean assumption induced by $\ell_2$ objectives is imprecise: latent-space geometry is so strongly anisotropic that equal-magnitude errors can produce drastically different decoded distortions. We further find that this anisotropy has a clear feature: sensitive variation concentrates in a low-rank subspace. The dominant low-rank components capture the overall structure, encoding long-range, spatially coordinated variation while remaining resistant to local noise. Its orthogonal residual, in contrast, mainly captures local and image-specific variation. Motivated by this asymmetry, we introduce Latent Structure Flow (LSF). At each block, LSF decomposes the latent state into structure and residual, models structural changes with global context, and predicts residual variation locally while preserving a direct path for the input structure. LSF changes only the generator, leaving the frozen codec and pointwise training objective unchanged. Across cross-modality synthesis and tumor inpainting tasks, LSF outperforms all compared baselines on both global and tumor-specific metrics, demonstrating the benefit of explicitly modeling latent-space structure for 3D medical image synthesis.

Human motion generation plays an important role in applications such as character animation, virtual environments, and embodied interaction. While existing approaches have achieved remarkable progress, many of them are developed for individual tasks, including text-to-motion, pose-conditioned generation, and trajectory control. Although these tasks involve different types of conditions, a unified framework capable of handling them within a common representation would greatly simplify motion generation systems. We observe that diverse motion conditions can be naturally formulated as different observation patterns over motion sequences, where each task corresponds to a specific masking strategy. Based on this insight, we introduce MotionMaestro, a unified motion generation framework that learns a shared representation for complete motions and heterogeneous partial observations through masked motion tokenization. MotionMaestro employs a three-stage training strategy that first learns a masked motion tokenizer, then refines its reconstruction ability on clean motions, and finally trains a conditional flow-matching generator in the learned latent space. Furthermore, we introduce an observation map and an observation loss to explicitly preserve provided motion conditions during generation. With this unified representation and conditioning mechanism, MotionMaestro supports text-guided and unconditional synthesis, pose conditioning and partial completion, temporal interpolation, trajectory control, and motion continuation. Experiments on the large-scale RoMo and MotionMillion datasets show state-of-the-art performance across diverse motion generation tasks.

A graphical user interface (GUI) agent that ships on a phone runs a small model and reuses experience: action paths recorded on earlier runs and cached from the cloud. When an app updates, part of this experience becomes silently wrong. On nine real app version pairs, an agent carrying experience recorded on the old version succeeds 10.1 points less often than one carrying none, and it does not notice the mismatch until 3 steps after acting on it. We propose REBASE, a protocol that keeps the device copy and the cloud copy of the experience coherent across updates. When a version changes, the cloud replays its copy on the new version and the device verifies each recorded step before executing it; when a step still fails, the device sends the cloud evidence in increasing size, starting from an accessibility-subtree diff, until the cloud re-derives an executable patch, keyed by version so that one repair serves the whole fleet. On two Jetson devices, REBASE restores the success rate of stale experience to that of fresh experience recorded on the new version within one episode, at 708x fewer bytes and 47.3x fewer cloud calls than uploading screenshots to the cloud, and 2.5x less device energy than running without experience.

AI agents increasingly resemble software systems: they call tools, remember facts, follow policies, delegate work, and take actions with real consequences. % Yet the ``program'' of an agent is scattered across prompts, tools, memories, workflows, and execution traces, making its behavior difficult to inspect through ordinary testing and debugging alone. % This essay argues that a programming-systems perspective offers a natural lens for making agents reliable. % We recast agents as programmable artifacts whose behavior can be specified over traces and state, checked before deployment, monitored during execution, and improved from observed failures. % The goal is not to make probabilistic agents behave like deterministic programs, but to give them enough structure that their behavior can be reasoned about, controlled, and repaired.

Automated science needs agents that can work out the rules of an unfamiliar environment by interacting with it. Interactive rule-discovery puzzles offer a controlled setting for studying this ability: an agent infers hidden rules through experimentation and uses what it has inferred to reach a stated goal. We ask what limits current language models on these puzzles and whether reinforcement learning (RL) improves performance on rules held out from training. To study both, we introduce WITNESS, a 2D grid-based puzzle environment with ground-truth ASCII observations and controlled access to rules. An agentic pipeline generates games for WitnessGym, the RL training suite, and WitnessBench, comprising public validation and private test games. The validation set separately tests new compositions of trained rule primitives and primitives absent from training. Under a shared harness, the best of 18 frontier proprietary and open-weight models solves only 24% of private test level slots, with scores sensitive to the observation interface and agent configuration. Providing ground-truth rules raises Opus-5's validation RHAE-L5 (relative human action efficiency over the first five levels) from 59.9 to 97.8, whereas a 27B open-weight model gains only 2.1 points and remains limited even with the rules provided. RL on WitnessGym raises the 27B model's private test RHAE-L5 from 2.1 to 5.4 and yields a mean gain of 4.1 points on four external discovery benchmarks. Together, these results point to rule acquisition as a major difficulty for frontier models like Opus-5 while smaller models further struggle on rule-based execution, and indicate that RL on hidden-rule puzzles transfers to broader rules and real-world tasks beyond training. Benchmark is available at: https://witnessbench.ai

We explore whether Omni-Language Models (OLMs) can be directly applied to zero-shot Semantic Audio-Visual Navigation (SAVN). Recent work demonstrates that even state-of-the-art specialized models still struggle to achieve generalist multimodal navigation, despite extensive task-specific training. In this paper, we introduce RAO-Nav, short for Reasoning All-in-One OLM, a deployment pipeline for zero-shot SAVN. By leveraging the rich implicit audio-visual knowledge encoded in OLMs, the embodied agent is enabled to ``hear'', ``see'', ``reason'', and ``act'' in the environment. To further elicit the built-in thinking ability of OLMs, we propose a test-time Latent Navigation Reasoning (LNR) module that can be seamlessly integrated into the decoding space. LNR encourages the model to retrieve more target-relevant observations and make effective navigation decisions. Through comprehensive experiments, we show that our framework surpasses existing state-of-the-art baselines on public SAVN benchmarks without using any training data. Moreover, we introduce a new Global Navigation Instruction setting to further evaluate the ability of OLMs to serve as embodied navigation agents. Code: https://github.com/rikeilong/OmniAV_Nav.

What controls are needed to interpret execution performance and audit scores in LLM agent evaluations? We study two limits on these interpretations in a financial agent harness. In Study~A, comparing independent runs under idealized and stressed execution on three synthetic settings that share one 24-day upward phase mixes the execution rule with fresh model responses and portfolio feedback: the parsed decision paths agree in only $19.8%$ of $450$ pairs. Replaying each stored response tape through both execution destinations gives a narrower result. Conditional on those responses, stressed execution changes total return by $-0.0170$ (95% interval $[-0.0230,-0.0117]$), or $10.4%$ of the idealized baseline, and ten seed clusters do not resolve the model ranking. Study~B corrects an incomplete answer key and replaces legacy tasks with matched zero-, one-, and two-defect tasks under an explicit multi-label prompt. The drop in target violation recall from one to two defects is positive in five of six combinations of auditor and source (median $0.267$), with three surviving Holm correction. Yet the auditor that includes both target labels most often has micro-precision $0.149$, emits findings on $98/100$ zero-defect tasks, and returns the exact dual-defect set in only $21/100$ cases. Target recall by itself therefore gives a poor account of audit quality on this construction. The studies address different limits: what an execution comparison estimates, and what target recall captures. Together, they show how fixed conditions and diagnostic controls bound the claims a score can support.

Designing dense reward functions for low-level reinforcement learning (RL) control remains difficult. Recent work uses large language models (LLMs) to iteratively generate and refine reward functions using policy-training feedback within scripted search algorithms. However, evaluating each candidate requires a full RL training run, making sample efficiency a central challenge for reward search on complex control tasks. To address this limitation, we propose an Agentic Reward Black-box Optimization (ARBO) framework, in which an LLM agent builds the search strategy at run time from an evaluation history maintained as its persistent workspace. The evaluation history comprises two components: observations maintained by the evaluation oracle, including candidate scores, per-term training curves, and error tracebacks; and an agent-maintained belief that records diagnoses and intended next steps. The agent queries both with tools and generates the next batch of reward candidates, rather than generating them in a single pass from a fixed prompt. Across four control domains, ARBO achieves gains of 29.9% in manipulation success rate and 192.8% in power-grid score over baseline means under a shared evaluation budget. Ablations examine each component's contribution and sensitivity to backbone choice.

LLMs enable multi-agent systems (MAS) to tackle complex tasks, but manually designing agent roles, prompts, and communication structures requires substantial expertise and effort. This motivates learning policies that construct query-specific MAS from execution reward. Existing approaches typically train these policies by repeatedly traversing a fixed set of training queries and assigning rewards at the workflow level. However, this overlooks differences in queries' evolving learning potential and obscures which substructures improve solution quality. In this paper, we propose Active Substructure-aware Policy Optimization (ASPO), a RL framework for query-level MAS search. ASPO introduces an Adaptive Query-Selection Mechanism (AQSM) that focuses training on queries at the policy's competence boundary: those it can solve but not yet reliably. A complementary discovery mechanism widens architectural exploration for hard queries, helping distinguish insufficient exploration from operator capability limits. Beyond query selection, ASPO introduces substructure-level rewards that measure output-quality gains within each action's descendant subgraph. These rewards guide proximal policy optimization to reinforce useful architectural refinements and discourage redundant or harmful computation. Together, these mechanisms prioritize learnable queries and provide fine-grained feedback for learning effective MASs. Across six benchmarks spanning mathematical reasoning, general question answering, and code generation, ASPO ranks first on every benchmark against twelve baselines.

An agent harness, the code that turns a model into an agent, writes its own record of each run, and that record is all a later reader gets when a run is disputed, investigated or audited. We call a record evidentiary when a reader who was not there can check it without trusting the writer. Across sixteen deployed frameworks, none writes one in full. Hearsay examines the record, not the task: five harnesses run fourteen tasks, three blinded LLM examiners and a human panel read the records, and every excerpt an examiner quotes is checked mechanically for who wrote it. First, the record lets a reader name the fault but not prove how the run went. Examiners name the right fault in 74 to 91% of 140 runs, but the fault can be proved only from two files the benchmark adds; for what happened in between, fewer than one citation in ten lands on anything the harness did not write, and the examiner with the fewest false alarms catches half of the entries we delete, rewrite or fabricate. Second, the remedy is a second author, not a stronger seal on the first. An append-only log of what passes between harness and model, kept outside the harness and read against the record in both directions, reports all 28 omissions and fabrications we made a harness commit as it ran, where a hash chain over the harness's own record passes all 28. Handed the log, examiners keep their fault verdicts but rest more of their citations on what the harness did not write. What makes a record evidence is who writes it, not what is captured.

Deployed LLM agents increasingly rely on persistent, versioned system assets such as routing rules, knowledge segments, prompt instructions, and reusable skills. Failure-localization methods can identify where an error manifests in an agent or execution trace, but repair requires a different decision: which editable system asset should be changed, and is that change expected to improve the task outcome? We study this gap through component-attribute failure attribution, where diagnosis targets versioned, addressable items rather than execution locations. We propose the Detection-Aware Attribution Framework (DAAF), which learns the effects of valid attribute replacements and amortizes this intervention evidence into deployment-time diagnosis. DAAF combines sparse and noisy failure signals to decide whether intervention is warranted, learns component-type-conditioned replacement effects from controlled replays evaluated by executable task outcomes, and shares supervision across requests with compatible intervention responses. At diagnosis time, DAAF uses only the observed execution, registered candidates, and available failure signals; it requires neither counterfactual replay nor task reward and returns no_change, a repair target, or an unresolved decision when evidence is insufficient. On held-out tau^2-bench Telecom tasks, DAAF achieves 80.72% attribute Hit@1, recovers 62.65% of failed executions while limiting clean-task regression to 3.23%, and reaches 71.93% overall task success. These results show that intervention-grounded attribute attribution can connect failure localization to executable system repair.

LLM-driven agents are increasingly deployed in complex applications, where long agentic traces make failures difficult to diagnose. Existing trace diagnosis methods often conflate anomalies, errors, and failures, making diagnostic targets ambiguous; they also lack structured modeling of how causally relevant errors propagate and amplify into final task failures, resulting in unreliable failure attribution. To address these problems, we propose CEG-Agent, a tool-augmented agentic framework for causal diagnosis of agentic traces. Specifically, CEG-Agent introduces an explicit taxonomy of anomalies, errors, and failures, and constructs Causal Error Graphs (CEGs), a unified typed representation that links execution events, diagnostic nodes, and failure outcomes through causal relations. To evaluate causal trace diagnosis, we further construct CEG-Bench, a fully agent-annotated benchmark with high-confidence, consensus-derived CEG annotations obtained through an Adversarial Agentic Adjudication Protocol (AAAP). We validate the resulting annotations against an expert-curated human gold set, which shows close agreement with the automatic annotations. Experiments on CEG-Bench demonstrate that CEG-Agent achieves state-of-the-art performance under both semantically relaxed and structurally exact evaluation criteria. Our code is publicly available.

Large language models (LLMs) have enabled data-science agents to automate multi-step analyses over heterogeneous files. However, incorrect choices regarding data sources, scope, or statistical definitions often lead to silent errors: computations execute successfully but produce plausible yet incorrect outputs that fail to answer the intended question. To mitigate this, we present RADAR, an auditable runtime that makes an agent's analytical choices inspectable and supports their revision through execution feedback. RADAR operates through three core mechanisms. First, an evidence-preserving exploration module retrieves task-relevant content while retaining source locations and observation coverage. Next, the runtime uses typed operators to record the agent's declared inputs, operation arguments, and resulting observations. Finally, runtime validation checks proposed operations against these observations. When a conflict is detected, the runtime rejects the operation or provides diagnostic feedback, allowing the agent to revise its choices before errors propagate. This design enables agents to fail loudly while leaving semantic interpretation to the LLM. On KramaBench, RADAR achieves overall scores of 0.723 with full source retrieval and 0.747 with gold sources supplied, corresponding to relative gains of 35.9% and 28.8% over the strongest baselines. Beyond KramaBench, RADAR achieves relative performance gains of 14.0% on DA-Code and 59.3% on DABStep, demonstrating its applicability across diverse agentic data-analysis workflows.

Audio language models can recognize spoken commands and invoke tools, but an agent must first decide whether the acoustic and conversational context warrants action. We introduce VGBench, a 1,018-item diagnostic benchmark for action-level addressedness across side-talk, self-talk, and speaker-switch scenarios. Each item uses a shared action space comprising silence, a tool call, and a natural-language answer. Speaker-switch pairs hold the specified words fixed while source, distance rendering, and a temporal boundary define a controlled wearer-to-bystander shift. Six raw Audio LLMs and three training-free adaptations often identify the target tool yet rarely withhold action under this shift; the highest raw switch mute rate is 14%. We then use VoxGate as a post-training case study. Supervised training mutes 91.3% of switched commands while choosing the correct tool for all nearby wearer commands and text-only controls. An exploratory GRPO stage has similar switch performance; side-talk accuracy rises from 68.4% to 70.9%, and self-talk muting from 52.0% to 60.0%. Factorized controls identify an independent source-change effect, while sensitivity to the far-field manipulation varies across acoustic renderings. The benchmark therefore measures multi-cue acoustic-context gating rather than isolated speaker identity.

Learning from experience in LLM agents has become a key paradigm for developing self-evolving agents that continuously learn and expand their capabilities. Within this paradigm, synthesizing the agent skill has emerged as a promising solution for transforming accumulated experience into reusable procedural knowledge, serving as an important layer for the harness system that supplies agents at runtime. Despite its potential, existing approaches largely abstract past experience into fixed procedural knowledge before downstream demands are known, which risks discarding knowledge that later becomes critical while retaining instance-specific details irrelevant to future tasks. In this paper, we reframe agent skill synthesis as a dynamic navigation problem over past experience, where agents actively explore accumulated trajectories on demand for the current task with targeted and fine-grained access to experience knowledge. To this end, we propose ExpVoyager, a novel framework in which a skill curator navigates raw experience across different views and resolutions, continually identifying reusable procedural knowledge from what it observes while tracking remaining knowledge needs that guide where to navigate next. Extensive experiments demonstrate both the effectiveness and versatility of ExpVoyager, showing consistent improvements in downstream task performance, continual gains as the experience space scales, and practical compatibility with existing skills under efficient experience access.

Detecting compromised business ad accounts is a challenge in digital advertising, as attackers exploit hijacked accounts to launch fraudulent campaigns. Large Language Model (LLM) agents show promise for integrity enforcement, but hallucinated mistakes on hard cases create business friction. In a study we find the autonomous agent is a strong, recall-heavy signal extractor but an unreliable final arbiter, conceding precision on ambiguous decisions. We therefore keep the agent as an investigator that emits a structured, interpretable signal vector, and delegate the verdict to a neuro-symbolic stage: symbolic rules discovered by Inductive Logic Programming (FOIL-IE), a Naïve Bayes calibration layer, and a data-tuned contradiction layer. Evaluating on a compromise-over-sampled population and a realistic low-prevalence sample with subject-matter-expert labels, this arbiter substitution raises MCC from 0.295 to 0.435 (ΔMCC +0.139, 95% CI [+0.026, +0.245], p=0.018, paired bootstrap), lifting precision from 0.250 to 0.446 (1.8x) at a recall cost (0.920 to 0.660). Benchmarked under identical conditions, it also edge tree ensembles (0.386).The rules encode domain w labels while remaininginterpretable and auditable.

As the use of agentic artificial intelligence increases in nearly every industry, there exists a widening attack surface. It is necessary to monitor agents to ensure that agents are acting in a way that is aligned with the users intent. Auditing an agent's network traffic provides a clear record of the agent interactions. This work presents a novel approach to monitoring the network traffic of agentic systems using complex valued hypersparse traffic matrices by integrating DBOS (DataBase OS), the OneSparse PostgreSQL database, and the GraphBLAS math library. To develop these concepts an agentic simulator was constructed, allowing a varying numbers of AI agents to collectively survey a virtual environment using different strategies. The resulting network traffic matrices enable easy monitoring of the AI agents.

Healthcare administrative staff transfer structured information from electronic health records, referrals, claims systems, provider rosters, and work queues into dynamic forms. We developed and evaluated CLAIRE (Clinical Language and Agentic Intelligence for Reasoning and Entry), a hybrid workflow that separates field-state discovery, source-to-field mapping, deterministic validation, bounded correction, escalation, and audit tracing. We tested five synthetic healthcare administrative schemas, 1,000 source records, four interface variants, two data-quality suites, and six comparators, yielding 24,000 benchmark episodes. A separate strict-output audit evaluated direct mappings from Qwen2.5-1.5B and Qwen2.5-7B, and a trace-derived operational simulation covered 6,000 episodes. Under the evaluated synthetic benchmark conditions, full CLAIRE achieved 1.000 episode success, field accuracy, required-field completion, and dependency completion in both suites; removing validation reduced stress-suite success to 0.500. In the simulation, 100.0% of clean and validation-stress episodes reached a staff-reviewable draft, compared with 68.6% of escalation challenge episodes, unsupported cases were blocked. Scenario-based savings were 149.7-165.5 seconds per case, not observed staff times. The findings support schema-grounded, validation-first healthcare administrative automation in which language-model components assist mapping but do not authorize unsupported or consequential actions.

Long tasks produce more history than an LLM agent can hold in its context, and more than it uses reliably even when the history fits. A growing line of work therefore has agents carry a short written state instead: at every step a writer rewrites the state, and a reader acts from the state alone. Steps stay cheap, but anything the writer drops is lost before later decisions reveal that they need it. We quantify this loss and ask whether training can reduce it. Comparing the written state with the best state of the same size written in hindsight, we split the reader's loss into a budget loss, which any state of that size must incur, and a write-time regret, which comes from the writer's choices. In TextWorld cooking games where we control how long a fact must be carried before it is needed, a 128-token state holding the facts wins nearly every game, while prompted language-model writers win at most 17%. Almost all of the loss is write-time regret, and it grows with the delay. We then train the writer from the reader's own loss. DSSR (decision-sufficient state representations) scores candidate states by how well the reader acts after the writer carries them forward, and teaches the writer to prefer the better ones. This forward-rolled score predicts game outcomes ($ρ= 0.48$), whereas scoring a candidate as a fixed context, as hindsight methods usually do, does not ($ρ\leq 0.07$). On a pre-registered test split opened once, training adds +7.0 [+1.9, +12.2] points of success when facts are needed soon, bringing a plain summary writer to the level of belief- and slot-based memory prompts. The gain shrinks as the delay grows and is significant only at the shortest delay. We trace this limit to credit assignment: keeping a fact now pays off only if every later rewrite keeps it too, which a per-step score cannot see.

The proliferation of generative and agentic artificial intelligence (AI) systems has introduced computational demands whose environmental consequences are substantial yet underexamined. This paper examines the environmental footprint of modern AI systems across energy consumption, carbon emissions, water usage, and electronic waste over the full lifecycle of large language models, multimodal foundation models, and agentic workflows, from hardware fabrication and training through fine-tuning and inference to end-of-life disposal. This work provides a conceptual analysis, utilizing order-of-magnitude estimations based on published data, without conducting original physical measurements. We contribute a lifecycle taxonomy that crosses lifecycle phases with five impact dimensions and emission scopes; a Sustainability Assessment Framework for AI Systems (SAFIA) comprising nine indicators; a comparative analysis of traditional, generative, and agentic AI; seven open challenges; policy recommendations for regulators, cloud providers, hardware manufacturers, and AI developers; and a research roadmap to 2035. The analysis indicates that inference can rival or exceed training energy over a deployment lifetime, that agentic workflows can multiply the energy of equivalent single-pass inference by one to several orders of magnitude depending on the number of model and tool calls, that indirect water use and embodied carbon are systematically underreported, and that measurement tools, disclosure practices, and regulation have not kept pace with agentic deployment. These findings call for agent-aware energy measurement, lifecycle-based carbon and water accounting, and mandatory disclosure for large-scale AI training and deployment.

Tool-using language-model agents can decide that an action is permissible and execute it only after security-relevant state has changed. We study this proposal-to-commit gap and introduce BSC-R, a deterministic effect-boundary mechanism that binds a single-use commit authorization to the exact action and to a semantic projection of the authorization state that justified it. On 2,847 attacked AgentDojo episodes, the boundary kernel preserves the unprotected agent's behavior exactly (80.576% utility; 1.616% attack success). On 10,302 frozen proposals, it accepts every unchanged commit and rejects every instance of ten prospectively specified changed or replayed classes. In an independently generated boundary-drift experiment, full joint binding commits 0/4,403 invalid contexts while retaining 5,899/5,899 valid contexts. A prospective external evaluation on the 3,460-scenario CONTINUITY suite retains all 700 benign cases, prevents 1,200/1,200 represented non-replay invalid effects, handles 160/160 replay lifecycles correctly, and withholds 200/200 ambiguous no-release cases. The broader external suite also exposes the method's limit: across all 2,560 attacks, BSC-R has a 25% invalid-effect commit rate versus 0% for CONTINUITY. The result is therefore a scoped commit-time consistency mechanism, not a universal agent-safety claim.

Large language models can memorize private or harmful information, motivating machine unlearning methods that remove targeted knowledge while preserving other capabilities. However, existing evaluations rely primarily on behavioral benchmarks, which assess whether a model appears to forget but provide limited insight into how unlearning changes the model or affects related knowledge. We introduce TRIAGE (Tripartite Representation-internal Introspection for Adjacency Gap Evaluation), a benchmark-agnostic evaluation framework for characterizing these changes. TRIAGE uses diagonal approximations of the Fisher information and Hessian to measure changes in parameter sensitivity and local curvature, and utilizes a Forget / Adjacent-Retain / Generic-Retain partition to quantify an adjacency gap in semantically related knowledge. Based on the magnitude and distribution of these changes, TRIAGE further classifies each algorithm's update as no-op, partially localized, collateral dominant, or globally destructive. Across 12 unlearning methods, four language models, and the WMDP, TOFU, and MUSE benchmarks, we find that methods with similar behavioral forgetting can produce substantially different internal changes and patterns of collateral damage. These signatures also vary across models and benchmarks, indicating that the effects of unlearning are not determined solely by the unlearning algorithm. TRIAGE can be applied alongside existing unlearning benchmarks to complement behavioral evaluation with a model-internal view of how unlearning reshapes the model's parameter space and affects retained knowledge.

Transformers process information strictly forward through depth, preventing deeper computation from revisiting and refining earlier representations. To augment the standard forward pass, existing approaches either re-execute depth, incurring additional computation, or modify the residual stream using predefined directions, limiting their instance-level adaptation. Recently, inference-time feedback offers a direct mechanism for recycling endogenously produced computation by writing deeper residual states back to earlier layers, yet what should be fed back remains unclear. We argue that the depth increment Delta, capturing newly accumulated computation between two layers, provides a more effective, composable, and scalable feedback signal than the full state. Building on this observation, we introduce ReFlux, a learnable feedback graph that dynamically selects and composes increment-carrying routes. ReFlux supports synchronous feedback to the same token and streaming feedback to subsequent tokens. Extensive experiments across various models, corpora, and benchmarks show that synchronous ReFlux consistently reduces perplexity across ten language-modeling corpora, and improves accuracy by 2.1-2.3 points, with gains reaching 4.7 points on multi-hop reasoning. Streaming ReFlux further retains most of these gains while preserving the base model's 1x theoretical backbone FLOPs. These results establish ReFlux as an efficient paradigm for unlocking the latent computational potential of LLMs, allowing them to revisit the same tokens with fresh representations. Code implementation can be found at https://github.com/gooogleshanghai/reflux.

In tri-modal visual question answering (VQA), auxiliary text is commonly used to complement visual and textual inputs, yet its reliability is often uncontrolled. While prior work studies modality conflicts in general, it remains unclear how different types of unreliable auxiliary text affect answer selection under fixed image-question-option contexts. In this work, we show that the most harmful auxiliary text is not necessarily the most factually incorrect, but the one that aligns with the question while contradicting the image and favoring a specific distractor, leading to systematic redirection of model predictions. To isolate this effect, we introduce the Textual Reliability Ladder, a controlled diagnostic protocol that decomposes auxiliary text along three axes: image consistency, question relevance, and option support. Across multiple datasets (ScienceQA, VCR, A-OKVQA, Causal-VidQA) and recent VLMs, we find that such distractor-supporting text induces the largest accuracy drops (up to 53.1%) and concentrates errors on specific incorrect options. To mitigate this failure mode, we propose a training-free inference-time intervention that explicitly counteracts this redirection effect via noise-stability steering and dynamic grounding, reducing redirected errors while largely preserving performance under faithful text. Our results highlight that auxiliary-text reliability must be understood at the decision level, rather than solely through factual correctness, and provide a practical pathway toward more robust tri-modal reasoning.

The ability of vision-language models (VLMs) to abstain from unanswerable questions is as important as their ability to answer answerable ones accurately. Recently, several benchmarks have emerged to evaluate and improve VLM abstention, but they have substantial limitations. First, samples often contain shortcut cues in images or questions that reveal answerability, while an explicit "unanswerable" option further prevents accurate assessment of spontaneous abstention. Second, as training data, they generally provide only binary labels without fine-grained explanations for deeper supervision. To address these limitations, we introduce Visual Answerability Diagnosis with Rationales (VAD-R), a benchmark constructed through a two-stage pipeline of shortcut filtering and quality verification to prevent answerability leakage. Each example is annotated with step-by-step rationales and causal evidence-gap labels. Evaluation of state-of-the-art open- and closed-source VLMs on VAD-R reveals limited spontaneous abstention, with average recall rates of only 11.4% and 16.3%, respectively. Probing analyses show that hidden-state representations in certain layers can effectively distinguish answerability, yet this distinction fails to manifest in final responses. Motivated by this observation, we introduce Rep2Act, a representation-to-action alignment method that translates latent answerability awareness into explicit abstention decisions. Rep2Act improves action accuracy on VAD-R from 56.67% to 86.33% for Qwen2.5-VL-3B and from 59.33% to 88.67% for Qwen2.5-VL-7B. On the out-of-distribution TUBench, Rep2Act achieves an average F1 score of 53.3% with only a 3B model, surpassing the closed-source GPT-4 Turbo and GPT-4o by 16.2% and 1.1%, respectively.

Persona prompts ask language models to answer as particular kinds of people. We test whether relationships learned from these effects predict responses to new questions and remain useful across models and prompts. Across 57 attributes, three behavioral domains, and seven pairs of open 7 to 9B checkpoints, persona effects can be predictable without being portable. Separate attribute and task gains improve prediction beyond shared scaling significantly in OLMo-3 and Qwen2.5, with the most robust evidence in OLMo-3. In that model, target refitting significantly outperforms gains borrowed from each of the other six pairs. Across model transfers, borrowed gains with one amplitude underperform shared scaling in most directions; allowing two target parameters removes the significant losses but yields no significant benefit over target shared scaling. After rewording, refitting significantly outperforms reuse with one amplitude in all six tested pairs, while changes of examples or country context often preserve reuse value. In the tested prompt transfers, regularized updates outperform both reuse strategies in median at 64 target questions per attribute. A separate survey comparison finds that selecting the more responsive checkpoint can worsen human fit; responsiveness is confounded with training status, and temperature calibration largely removes this cost but not errors in group ordering. Within the tested gain representation, apparent transfer can come from target calibration; source relationships must add predictive value beyond calibration and regularization. Code and data are available at https://github.com/thzva/persona-gain

Model merging efficiently combines specialized large language models (LLMs) without joint retraining, but can substantially alter expert routing in Mixture-of-Experts (MoE) models. Such routing drift is often interpreted as routing failure, raising a fundamental question that remains unclear: does routing drift after MoE merging actually indicate routing failure, and what evidence should justify repair? We investigate these questions across DeepSeekMoE, OLMoE, and Qwen3-MoE proposing a routing analysis toolkit for controlled counterfactual interventions and token-level analysis. By crossing source and merged router inputs and parameters, we attribute most expert reassignments to input shifts rather than parameter changes at the same layer. However, source-relative routing differences poorly predict next-token likelihood gains from source-route restoration, and different expert selections can produce directionally similar mixture outputs. We therefore operationalize routing failure as task loss recoverable under a specified routing intervention, with non-routing parameters fixed. These tests detect recoverable loss under deliberate router corruption, whereas source-route restoration does not establish reliable task benefits in the evaluated merged models. Motivated by these, we propose Selective Router Repair (SRR) as a case study, and find that source-specialist token-likelihood advantages do not reliably identify beneficial local corrections. Together, these findings show that routing drift alone is insufficient evidence of routing failure: source-informed corrections must be judged by their task-level intervention effects. The analysis toolkit and SRR code are released.

Repeated sampling can reveal a correct numerical answer without yielding either a reliable system output or a supported derivation. We present Coverage, Realization, and Validity Evidence (CRV), an evaluation protocol for sampled large language model (LLM) reasoning over formal geometry states. Coverage is answer availability, realization is readout accuracy on the frozen candidate pool, and validity evidence is a label-blinded critic judgment of derivational support rather than a proof certificate. CRV freezes each candidate pool before comparing readouts and analyzes covered failures by correct-answer multiplicity and within-problem discrimination. On HardShift441, a 441-problem set for which a reference solver leaves 406 problems unsolved, a LoRA-adapted Qwen2.5-7B generator obtains 24.2% average single-sample accuracy and 68.9% pass@16, whereas verifier-weighted self-consistency (WSC) reaches 38.0%. Readout accuracy is particularly low when the correct answer occurs only once or twice in the pool. In a separate constructed audit of 195 covered problems, the critic labels 12 correct-answer representatives as supported, 181 as refuted, and two as uncertain. These results show that coverage, realization, and validity evidence from the critic are distinct quantities and should be reported separately.