每天知道
你的领域
出了什么

Building LLMs that behave well socially, not merely correctly, requires Building LLMs that behave well socially, not merely correctly, requires more than producing locally helpful responses. A socially competent agent must infer users' unstated goals, respect their preferences, and adapt as the conversation unfolds. These behaviors are inherently multi-turn and social, making them hard to optimize: real interaction data is scarce, and user preferences are typically latent rather than directly observable. To address these challenges, we build on a persona-driven social simulation environment (consisting of a persona library, LLM-based user simulators, and a user-satisfaction scoring system ranging from [0, 1]), to introduce preference-batched GRPO (PB-GRPO), a post-training algorithm that learns socially adaptive policies from conversation-level feedback. Compared to vanilla GRPO, PB-GRPO computes advantages using a normalization estimated across a bucket of users with similar preferences, stabilizing training across a diverse social population. Empirical evidence shows that PB-GRPO improves models' social behavior over strong reinforcement learning baselines in our simulated environment.

The prevailing approach to computer-use agents couples a model with a domain-specific harness: a browser or desktop environment equipped with human engineered tools that are fixed before task execution. As models' coding capabilities improve, the GUI native and static harness prevents them from direct programmatic operation on system state, as well as flexible construction of tools. To this end, we introduce CUAWright, a minimal terminal harness of roughly 3K lines of code that uses bash commands as its sole action interface, and a file system as its evolvable space for dynamically creating tools and managing the context. We conduct comprehensive experiments across a wide range of digital tasks, and demonstrate that by giving the agent a minimal, programmable interface, it achieves substantially stronger results compared to their GUI or hybrid CLI interface across a wide range of tasks. On OSWorld 2.0, CUAWright delivers a 33.2% relative improvement in partial reward while reducing estimated cost by 37.5% compared with the published GPT-5.5 baseline. On Online-Mind2Web and the long horizon Odysseys benchmark, CUAWright substantially outperforms GUI native harness by 4.7% and 44.0% in success rate, respectively. Furthermore, we found the gains extend to CAD applications that require accurate visual understanding and CLI interaction: on CADGenBench and BenchCAD, our unified harness yields 8.1%-41.6% relative improvements over other CLI based harnesses with GPT-5.5. Together, these results suggest digital environments are far more programmable than their GUI interfaces imply, and a minimal terminal-focused harness is the key for better performance and efficiency.

A deployed LLM agent emits tool calls, queries, and code that can be silently wrong — by the time the error surfaces, the action has run. Frontier chat APIs hide the model's token probabilities; the agent's stated confidence barely beats chance on the mistakes that matter; and resampling does not help, since frontier models are highly repetitive, reproducing the same call across samples. We recover the missing signal from a low-cost open-weight surrogate run in parallel. It reads the same context, schema, and proposed action as the agent, then scores the call from its own log-probabilities through a family of complementary readouts: teacher forcing and request-PMI weigh the likelihood of each argument value, a discriminative verdict judges the call as a whole, and tool-choice competition tests the function against its siblings. One principle says which to trust: a generative likelihood localizes wrong argument values, while the verdict catches holistically wrong calls. When the error type is unknown, an ensemble is the low-regret default. The readout is training-free, needs no access to the agent's internals, and costs one prefill pass alongside the tool call. On difficult coding tasks it reaches AUROC 0.825 where the actor's stated confidence is near chance (0.598), and the generative readouts beat it by +0.07 to +0.28 across three further actors. Against self-consistency it gains +0.14 to +0.19 on near-deterministic actors, at 1/K the cost. The signal drives two deployment modes: a real-time gate escalating the least-trustworthy calls for review (+0.05 to +0.30 accepted-action accuracy at 50% coverage), and confidence feedback, returning the tool result with the score so the agent adapts its next step — lifting task success on live-execution benchmarks (+0.119 and +0.137, p <= 1e-4) and beating a random-value control where step errors are silent (+0.078, p = 0.003).

Design-level security weaknesses can arise from requirements, trust assumptions, missing controls, and data flows before implementation begins. Existing security practices often identify these issues after code is written. We present a multi-agent LLM framework for design-time security analysis from product requirement documents and architecture diagrams. The proposed framework parses architecture diagrams into graph representations, generates misuse and failure cases, constructs attack trees, checks governance and compliance gaps, recommends mitigations, assigns enterprise security-domain tags, and produces a candidate revised architecture recommendation for expert review. The framework does not retrieve from Common Weakness Enumeration (CWE) databases at inference time. Instead, it analyzes system behavior, trust boundaries, component interactions, and data-flow assumptions. Misuse cases act as intermediate representations that link findings to system components and attack paths, while a validation and refinement loop filters unsupported findings and improves grounding, traceability, and actionability. We evaluate the framework on a Microsoft reference-labeled threat-modeling example, labeled synthetic PRD--architecture pairs, and two open-ended systems: Berty and Gas Town. The reference-labeled case supports threat-recovery and actionability analysis, while the open-ended cases evaluate validity, noise, traceability, actionability, redundancy, and attack-tree quality. Results show that architecture-informed, misuse-driven reasoning improves review quality compared with single-shot and ablation baselines. Keywords: LLM Multi-Agent Systems, Design-Time Security, Threat Modeling, Vulnerability Discovery, Architecture Diagrams, Security Analysis, Misuse Case Derivation, Attack Trees, Iterative Reasoning, Security Governance.

Financial LLM agents are often evaluated by comparing their end-to-end returns with those of a baseline and testing the paired difference against zero. This measures whether deploying the agent changes realized performance, but it does not isolate event-selection skill. An agent that frequently changes positions from flat to long can earn a positive paired return from an upward-drifting event pool even when it selects events at random. We propose the Agent Policy-Value Audit, which holds fixed the observed count of each ordered action-change type and randomly reassigns them across eligible events. The average payoff from these reassignments is the composition benchmark; the difference between observed deployment value and this benchmark is selection value. In semi-synthetic benchmarks based on real earnings-event returns, a zero-centered paired test falsely attributes passive exposure to selection skill in $11.6%$ of no-skill replications, while the transition-matched audit reduces this rate to $5.3%$. Applied retrospectively to 723 earnings events at 44 U.S. consumer-facing firms, the audit decomposes the agent's gross deployment value of $+15.2$ bps/event into a $+25.8$ composition benchmark and a $-10.6$ selection value. The agent does not detectably outperform matched random assignments. Financial-agent evaluations should report deployment value separately from event-selection value.

Agentic multimodal large language models (MLLMs) have recently pushed the frontier of visual reasoning by calling tools such as zooming and tagging. Despite the recent strong success of agentic MLLMs, this work uncovers a critical safety failure in the tool-use paradigm: agentic tool-using MLLMs become less capable of refusing harmful requests. Our experiments confirm that, across three popular safety benchmarks, all the top open- and closed-weight MLLMs we test exhibit significantly lower safety in tool-using settings than in non-tool settings, with a relative refusal failure rate increase of up to 68.7%. Based on analysis of 100,000+ responses, including extended experiments, we also propose two possible reasons for this safety degradation.

Memory poisoning attacks on LLM agents typically assume an external adversary who plants content in the agent's persistent memory to steer its behavior. We instead study, with no adversary involved, whether a misaligned agent can write a goal it cannot yet act on to persistent memory, so that a future aligned agent carries it out when the opportunity arises. We investigate this threat, which we refer to as self-propagation of misalignment, across 20 different scenarios, whose misaligned goals include self-preservation, power-seeking, undermining oversight, reward hacking, and deceiving the user. We simulate misalignment in 11 frontier models using two prompting strategies; unrestricted and values-only. The first explicitly states the misaligned goal, for instance, to prevent its own replacement, and self-propagation succeeds in 58% of runs. The second only describes what the agent cares about, for instance, that its continued operation is essential to its users, without specifying misaligned goals or directives. Even under this weaker prompt, self-propagation succeeds in 18% of runs, and every model self-propagates in at least one scenario. On removing the memory tool from the harness, we find that agents use the file system, writing the goal to a file in 74% of sessions; self-propagation still succeeds in 11% of runs. We also show that weaker models can propagate misalignment to more capable models, and that propagated goals can persist through 100 sessions of unrelated work. Existing defenses against memory poisoning and prompt injection do not directly address this threat because the memory content is generated by the agent itself, rather than injected by an external adversary. An LLM memory auditor from prior work (MemMorph) only reduces propagation from 71% to 34% of runs. We release our scenarios to support the evaluation of defenses against this emerging threat.

LLM agents for clinical text-to-SQL applications reason autonomously over multiple steps but cannot assess whether their own reasoning or outputs can be trusted. In high leverage applications such as healthcare, this presents a critical risk where system mistakes can be costly. These reliability failures are also resource failures: an incorrect reasoning trajectory spends computation budget on outputs that must be discarded. We introduce Sentinel, a trajectory-derived, classifier-based confidence layer that analyzes an agent's reasoning, code and database outputs to decide at three points whether to stop: refusing unanswerable questions before the agent runs, halting doomed trajectories mid-run, and withholding untrustworthy answers at delivery. Here, utilizing Chow's rule, we optimize decisions under the EHRSQL shared task's Reliability Score, which penalizes incorrect answers given a utility weighting, and find on the benchmark EHRSQL that Sentinel raises this score from +0.08 to +0.24 when mistakes have a low utility weighting, well above the +0.03 earned by refusing every question, with delivered-answer accuracy rising from 54% to 69% as coverage falls from 84% to 55%. At higher stakes the agents we test rarely answer reliably enough to deliver, and Sentinel detects this on its own, abstaining to that same +0.03 where the unmonitored agent scores -3.02. The same stopping decisions cut computation: at low stakes, where the system still answers, Sentinel eliminates 13-28% of agent steps at little or no reliability cost.

Software performance optimization requires identifying bottlenecks and improvement opportunities, and empirically evaluating the effects and trade-offs across performance objectives. Autonomous AI coding agents submit performance-oriented pull requests (PRs) to open-source projects, yet it remains unclear whether these changes reflect established optimization practices and adequately substantiate their claimed value. Extending our MSR 2026 Mining Challenge pilot study of 407 performance PRs, we analyze a sample of 1,130 agentic and 1,130 human-authored performance PRs collected through June 2026. We compare adoption outcomes and patch characteristics between the two groups, and examine their optimization practices and validation behaviors. The results show that agentic PRs are merged less often than human-authored PRs (54.4% vs. 73.3%), but the two groups use similar optimization strategies and report validation at comparable rates. Agentic PRs rely more on static reasoning and less on benchmarks (46.1% vs. 57.2%), although benchmark use converges by the end of the study period. Agentic PRs also report validation for code-smell refactorings as often as for resource-targeting changes, whereas human-authored PRs do so less often. Across both groups, about half of validated PRs report no quantitative performance metric, and relevant trade-offs are rarely quantified. Agentic optimization practice has become increasingly similar to human practice, but the supporting evidence remains incomplete. Establishing whether agent-proposed optimizations are supported by rigorous, quantitative, and multidimensional evidence remains an important challenge. These findings motivate evaluation infrastructure and review criteria that measure intended effects and relevant costs rather than treating the presence of validation alone as sufficient.

Agent harnesses specify the roles, instructions, tools, and communication structure used to solve a task, and the right harness depends on the query. Because the value of each design choice is observable only through execution, tailoring a harness to each query has required either executing alternatives at inference time or costly manual design. We introduce SHIFT, which moves execution out of the per-query search loop. A local LLM architect learns a policy over harness-building actions from search, and a value function that predicts, from measured executions, a utility balancing accuracy against execution cost. For each query, Monte Carlo tree search uses these predictions to construct a harness. Across 9,193 tasks in six benchmarks, from math to document and general-assistant tasks, with a Gemini 3.5 Flash executor, SHIFT attains the highest mean accuracy, about 80%, outperforming 17 baselines that span prompting, prompt optimization, and workflow search, and exceeding the strongest baseline by 7.2 percentage points. A cheaper mode of SHIFT also attains a higher mean accuracy than every baseline while using 32% fewer execution tokens than the strongest baseline. We further show that choosing structure, instructions, and tools jointly beats choosing only instructions or only tools by up to 9.1 percentage points, and that learned value selection identifies more accurate harnesses with lower execution cost from candidate pools.

When documents supporting an agent's derived facts are revoked or replaced, should it repair memory or re-read current evidence? We introduce an evidence-revision evaluation on medication- and problem-list tasks from public ICU records. Under revocation, replacement and control events, we compare full and source-filtered re-reading with caching, rebuilding and graph-local repair across two 7B models. Memory is supplied in full without retrieval, and costs include ingest, revision and every use. On short records, local repair uses 5-10$\times$ fewer revision tokens than rebuilding, yet every memory pipeline costs at least twice full re-reading in held-out conditions. In a small pre-specified development sweep, adding task-ineligible documents extended records to about 10,000 tokens; at that length, memory's mean cumulative cost fell below full re-reading's after 2-14 uses, partly through truncated extraction, while source-filtered re-reading remained cheapest. In the replacement study, none of the four primary confirmatory tests reached statistical significance. These results show why the cost of agent memory after evidence revision must be assessed against source-filtered re-reading over the full pipeline.

Deep neural networks continue to grow in parameter count, driving up training and inference cost on GPUs. Sparse neural networks and Dynamic Sparse Training (DST) promise to reduce these costs, but most implementations rely on binary masks over dense tensors and recover little of the theoretical compute, memory, or energy savings. We propose SNACK, a truly sparse GPU layer that stores and computes only non-zero connections. SNACK exposes a simple PyTorch API for restructuring connections and backpropagating gradients entirely in the sparse paradigm, and ships SNACK-COO, a custom COO-format SpMM CUDA kernel with a batch-to-Streaming-Multiprocessor mapping tuned for the small-batch, high-sparsity regime typical of large-model training and single-stream inference. At the kernel level, SNACK is up to 7x faster than the masked dense baseline (Dense+Mask) and competitive with cuSPARSE, Sputnik, and FlashSparse at 95% sparsity. At 90% sparsity, a single SNACK layer accelerates training by 8x and 3.7x, and inference by 4x and 2x, over Dense+Mask and fully dense layers, respectively, while using 72% less memory than dense and substantially less energy. End-to-end, SNACK reduces GPT-2 peak training memory by up to 40% and graph-style inference latency by 4.8x over Dense+Mask at 99% sparsity.

Recent world action models (WAMs) reuse pretrained video VAEs whose encoder latents directly condition downstream action policies. Quantization must therefore preserve not only reconstruction fidelity but also the policy-facing latent contract expected by the frozen policy. Direct NVFP4 leaves W4A4 quantization error uncompensated, whereas joint quantization-aware training (QAT) can recover reconstruction by moving this representation. On Wan2.1, joint QAT nearly matches FP32 VBench-7 (0.7403 versus 0.7409), yet LIBERO success collapses from 95.5% to 10.5%. Controlled decoder-only experiments show that activation quantize-dequantize operations alter the reconstruction signal and decoder Jacobian, redirecting the gradient returned to the encoder and inducing persistent latent drift. Based on this mechanism, we introduce LatentQuant, a two-stage NVFP4 QAT framework that first aligns the quantized encoder with its high-precision counterpart, then freezes it while adapting the decoder. Across Wan2.1 and Wan2.2, LatentQuant preserves near-baseline control and high reconstruction quality, achieving 95.75% success on LIBERO and 68.8% on RoboTwin. On NVIDIA B300 GPUs, NVFP4 execution achieves 1.17x-1.26x end-to-end VAE speedups over BF16 cuDNN.

Long chain-of-thought (CoT) traces impose substantial output-token costs. Under constrained budgets, compression must preserve answer-critical information, making boundary placement central. Token-level and fixed-length boundaries can fragment coherent spans such as phrases, formulas, and local derivations, whereas step-level boundaries can bind content requiring different compression actions. We introduce SynLat, a text-latent CoT framework that aligns compression boundaries with syntactic structure through non-overlapping Syntax-Aligned Units (SAUs). An answer-conditioned Teacher constructs progressive KEEP/LATENT targets for a single compression-conditioned Student, which generates mixed reasoning from only the question and requested compression level at inference. Across two Qwen3 Student scales, Standard-CoT and Long-CoT groups, and three compression levels, SynLat matches or exceeds the strongest evaluated baseline in all 12 task-group aggregates and strictly leads in 11 under the reported achieved-CR selection protocol. Overall gains reach 3.6/2.6 points at MEDIUM and 7.0/5.5 points at HIGH for Qwen3-8B/14B, with larger advantages under stronger compression, particularly on Long-CoT groups.

A trainable input embedding table assigns each vocabulary item an independently adjustable vector. We investigate whether this token-specific parameterization is required for substantial language-modeling capability, or whether a shared Transformer can learn from fixed token identities. We compare three decoder-only language models trained from scratch with the same tokenizer, contextual backbone, untied output-head architecture, and training recipe, with a target budget of 100 billion prediction tokens per model. Their input interfaces are a learned table, canonical 16-bit token-ID codes, and one fixed invertible recoding over GF(2). The fixed codes are repeated to model width without an additional trainable input projection. Both fixed-code models acquire substantial capabilities: canonical codes achieve 52.40% HellaSwag normalized accuracy, 70.51% PIQA accuracy, and 42.75% LAMBADA accuracy. The learned-input control performs better on several evaluations, including HellaSwag and LAMBADA, so these results establish viability rather than performance parity. The fixed interfaces remove 100.7 million trainable parameters, yielding 1.711B-parameter models, but parameter reduction is not the central result. These single-run experiments distinguish architectural necessity from empirical utility: independently trainable token-specific input vectors are not required for the observed capabilities. A fixed identity interface also provides a controlled setting for studying representation learning downstream of an immutable input, without establishing where particular capabilities are localized.

Mechanistic interpretability usually studies fully trained models, yet the computations that drive a behaviour can change while the model is still learning the task. On the Indirect Object Identification task, a model should continue with the name mentioned once rather than the name mentioned twice. Pythia models pass through an early training window in which they prefer the repeated name, so accuracy in a choice between the two names falls below one half while language-model loss on a fixed text sample keeps decreasing across the same window. The window reflects a temporary imbalance between two computations. We identify one cause of the wrong preference by selecting a set of attention heads that write the repeated name, on prompts separate from those used for causal evaluation, keeping that selection fixed, and then replacing each head's final-token output with its average output on a separate set of non-repeated-name prompts. This improves the correct-minus-repeated logit difference in a separately trained 160M model and in the official 160M, 410M, and 1B models. At 160M, the head that lowers the repeated name in the mature model shows little of its mature behaviour at this point. It directs less than one percent of its attention to the repeated mention, and its output makes almost no direct contribution to lowering that name's logit. Both properties grow over the interval in which behaviour recovers. Across the 160M, 410M, and 1B models, transplanting the corresponding head's mature parameters into the early checkpoint recovers 35 to 68 percent of the total improvement in the correct-minus-repeated logit difference seen by the end of training. Related early-to-late reversals appear at further Pythia scales, in two independently trained GPT-2 models, and in OLMo. A mature circuit can therefore conceal a transient causal configuration that shaped behaviour earlier in training.

Multi-agent systems built on large language models (LLMs) are increasingly applied to scientific discovery and hypothesis generation. Both the effect of refinement and the diversity of the delivered set are hard to interpret before experimental ground truth exists, and both are typically reported by deciding whether pairs of generated hypotheses describe the same underlying mechanism. We study two evaluation questions that rest on this pairwise equivalence judgment: (1) how much self-critique changes delivered hypotheses beyond run-to-run variability, and (2) how the equivalence rule used to group hypotheses affects measured diversity. Across four proprietary instances, we hold opening hypotheses fixed, rerun the downstream workflow with 0, 1, and 5 critique rounds, and score matched hypothesis pairs with an LLM-as-a-judge. Relative to matched same-depth reruns, moving from 0 to 1 round produces 34.5 percentage points (pp) of additional mechanism-level divergence, whereas 1 to 5 rounds adds 1.3 pp. We then compare three equivalence rules: term frequency--inverse document frequency (TF--IDF) similarity, dense embeddings, and the same LLM-as-a-judge. We construct controlled hypothesis pairs that either preserve the causal explanation through wording or biological-terminology changes, or replace one component of the causal chain while holding the rest fixed. All three rules are invariant to meaning-preserving edits, but when the initiating event is replaced, the LLM-as-a-judge identifies 83% of valid pairs as different mechanisms, versus 0% for TF--IDF and 8% for embeddings; varying only the rubric that defines same mechanism moves this figure from 38% to 96%. Together, these results show that pairwise equivalence judgments are a measurement choice: how mechanism equivalence is defined affects both the estimated effect of self-critique and the measured diversity of generated hypotheses.

Localizing behavior to individual components of a language model is a central goal of mechanistic interpretability. However, scoring components one at a time misses context-dependent effects: a primary component can inhibit the activation of a backup, leading to issues with ranking components. Actual causality studies the structure of such interactions via witnesses: variables that provide contextual information to resolve interaction terms. However, estimation with witnesses typically requires combinatorial enumeration and is infeasible in practice. We introduce the witness-integrated set effect (WISE), a family of causal estimands that build on the witness mechanism while taking expectations over sets of causes and witnesses to remain computationally feasible. Building on this approach, we introduce JuntaLearner, a gradient-based circuit discovery method that learns to rank components by their causal impact across varying-sized sets of components and witnesses. Alongside faithfulness metrics, we introduce measures of necessity and task specificity, and the circuit recognition score (CRS) to summarize each metric across circuit sizes while emphasizing effects achieved by small circuits. Across tasks and models of increasing size, JuntaLearner achieves higher mean CRS compared to attribution baselines on all metrics. Since its cost does not grow with the number of candidate components, JuntaLearner scales to large models while accounting for set-level interactions and avoiding first-order approximations.

Language models exhibit strong reasoning capabilities, yet adapting them to structured domains remains challenging and can yield inconsistent outcomes. We identify representation compatibility, the extent to which a model effectively processes a representation for a structured task, as a key factor in adaptation. We study this in chess, which provides a controlled testbed with precise semantics, computable optimal actions, and multiple state representations, including a symbolic encoding (FEN) and a spatial format (ASCII). We find that models often process semantically equivalent inputs substantially differently, affecting both learning and generalization. Building on this observation, we propose representation-aligned auxiliary supervision, which uses environment-derived tasks expressed in compatible representations to improve adaptation to structured domains. Across models and representations, auxiliary supervision consistently improves optimal-move prediction relative to target-only training under identical target data. Tasks that expose environment dynamics provide larger and most consistent gains than surface-level or static supervision, while remaining competitive with substantially increasing the amount of target-task data. Moreover, ASCII-trained models transfer more effectively to FEN than FEN-trained models do to ASCII, even surpassing the FEN target-only baseline on FEN evaluation. The gains also extend beyond optimal-move prediction to open-ended, factually grounded commentary generation. Overall, our results show that auxiliary supervision in model-compatible representations can enable effective adaptation in structured domains.

Methods for training language models on stale samples are judged by comparisons against importance-corrected baselines. We show that details of the experimental harness can reverse the observed ranking of methods, and we introduce PTH (Probe The Harness), a set of checks that makes the harness visible. Our case is a comparison between SAN, a behaviour-free method, and truncated importance sampling (TIS) on verl and in a single-GPU trainer, in which SAN first finished ahead in both stacks. Four details of the harness changed this comparison: the PPO ratio was taken against the learner's own recomputed probabilities, the data seed did not reach the TIS arm, the replay queue reused its first batch for 33 updates, and two loss normalisers differed from their description. In each case the logged quantity looked consistent with a working setup, while the quantity that defines the comparison went unchecked. With the harness checked, TIS matches SAN on verl, and in the trainer TIS learns steadily while SAN keeps a margin. We contribute the signature of each detail and its effect on the comparison, reference results for TIS and uncorrected GRPO under sampler lag, and the PTH checklist.

Improving joint audio-visual reasoning in Omni Large Language Models typically incurs substantial data construction and training costs. Our diagnostics reveal multi-hop reasoning difficulties despite correct answers to all corresponding single-hop questions and suggest partial decoupling in the local optimization of perception and reasoning objectives. This motivates post-training with different emphases on these capabilities. Text-only reasoning training yields gains across data sources, model scales, and families. With the best-performing text-only configuration, supervised fine-tuning followed by reinforcement learning (RL) raises Qwen2.5-Omni-7B's geometric mean of nine reasoning scores by 25.83% over the base model, outperforming the complete native audio-visual route with 56.6% fewer GPU-hours. Training on data synthesized entirely by a text-only LLM raises this geometric mean by 21.01% without audio-visual data in construction or training. However, text-only training degrades perception. We therefore propose a text-centric post-training paradigm: text-only training provides the main reasoning optimization, and reduced-data native audio-visual RL then refines perception. Refinement uses about 90% fewer input tokens than full-data audio-visual RL, restores perception above the base level, and retains 93.5% of the best-performing text-only pipeline's reasoning gain.

Rubric-based reinforcement learning extends reward-driven optimization to open-ended tasks by assigning partial credit to individual response requirements. However, rubric judges can assign a high criterion score even when the information or action it requires is absent from the response, a failure mode we term Vacuous Credit. Such awards persist after the required information is removed and can reverse the sign of a response's GRPO advantage. To address this problem, we introduce MetaRubric, which alternates evidence-aware policy optimization with response-guided rubric adaptation. We construct counterfactual counterparts by changing one task-relevant fact in each prompt. During policy optimization, credit is assigned only when the response contains sufficient evidence to satisfy the required rubric criterion. After each policy-optimization stage, current policy responses guide revisions to original and counterfactual criteria while preserving the meaning of the original prompt's initial rubric as interpreted under each prompt's facts. We also adapt criterion weights at stage boundaries to better address observed policy errors. Across multiple backbones, MetaRubric improves PubMedQA accuracy by 6.00--20.40 percentage points over static-judge GRPO, with further gains on HealthBench-Hard and two multimodal medical benchmarks.

Video multimodal large language models (MLLMs) keep climbing video question answering benchmarks, yet shuffling the frames, masking the segment that supports the answer, or occluding the target object barely changes their predictions. The accuracy rests on appearance and language priors, not on the temporal evidence the question asks for. We trace this to the unit of post-training: rewards are computed on a single response to the original clip, so the model is never asked to behave consistently across views. We propose Behavior Pack Optimization (BPO), which replaces the single response with a behavior pack of outputs across counterfactual views chosen by question type, scored jointly. The pack reward asks for stability when the intervention is irrelevant, sensitivity when key evidence is removed, and abstention when no evidence remains. To keep this objective stable at small pack sizes, BPO uses an anchor-relative advantage: the response on the original view serves as a per-prompt reference instead of a group mean over mixed views. On TempCompass, MVBench, and NExT-QA, BPO improves the macro accuracy of Qwen2.5-VL-7B-Instruct by 4.7 pp, the temporal-hard subset by 7.8 pp, and abstention F1 by 20.0 pp over a budget-matched vanilla GRPO baseline from the same SFT checkpoint. The gains transfer to Video-MME, LongVideoBench, and to LLaVA-Video-7B; ablations confirm they follow the view sets, not the rollout count. We hope this pack-level perspective offers a useful starting point for the video MLLM and multimodal post-training community as the field moves toward evidence-grounded video reasoning.

E-commerce videos are information-dense and frequently compared by consumers evaluating products and merchants assessing marketing strategies. However, existing multimodal models mainly focus on single-video understanding and have limited ability to compare information across videos. We introduce AdsCVR, the first e-commerce cross-video reasoning benchmark, containing 2,483 videos and 6,110 question-answer pairs across six reasoning dimensions. Cross- video reasoning requires models to locate fine-grained evidence among many redundant frames and integrate visual details, speech, and on-screen text. We therefore propose AdSeek, an agentic framework that dynamically selects visual and audio tools during multi-turn exploration, replacing static uniform sampling with active evidence acquisition. To address the sparse credit assignment of reinforcement learning, we develop an offline trajectory rectification mechanism that identifies reasoning errors and missing multimodal evidence in RL-generated trajectories. The corrected trajectories provide supervised fine-tuning signals that reduce biases learned during RL. This mechanism supports a rectified bootstrapping pipeline in which initial RL exposes reasoning bottlenecks, supervised fine-tuning corrects them, and a final RL stage further improves the policy. AdSeek achieves 74.30 percent accuracy on the AdsCVR test split, outperforming its Qwen3-VL-8B-Instruct backbone by 27.90 percentage points. It also generalizes to the open- domain CrossVid benchmark, demonstrating effective active evidence gathering.

Adaptive LLM reinforcement-learning post-training changes multiple training actuators online, including rollout temperature, group size, clipping, KL regularization, verifier allocation, and update budget. Three coupled issues remain unresolved. A future-risk model trained from behavior trajectories need not estimate the risk induced by the controller that will be deployed; a score calibrated on logged state-action pairs can become miscalibrated after selective action choice; and independent per-resource minimum costs do not in general certify a feasible multi-resource continuation. We introduce FSPO, a feedback-state controller for budgeted LLM RL post-training that addresses these issues jointly. FSPO learns a policy-consistent risk-to-go model whose Bellman target follows the same frozen controller used for future decisions, together with a long-horizon utility model. Decision-conditioned trajectory calibration (DCTC) calibrates risk on cross-fitted trajectories generated by actions selected by provisional controllers. A Pareto resource continuation certificate (PRCC) admits an action only when a non-dominated cumulative reservation remains feasible over the residual horizon. Under a matched GRPO resource envelope, FSPO reaches 66.11% held-out and 59.43% OOD accuracy, compared with 64.47% and 57.03% for PB2, the strongest evaluated adaptive baseline. Three paired training seeds give gains of +2.42 and +3.19 percentage points over the contextual bandit on held-out and OOD evaluation. Under high behavior-deployment mismatch, policy-consistent risk lowers selected-decision ECE from 0.108 to 0.053; DCTC lowers it from 0.039 to 0.022 at matched acceptance; PRCC removes false-feasible admissions on an 18-action catalog ($0.197\rightarrow0.000$); and enabling all three components reduces trajectory failure from 0.181 to 0.083 in a factorial ablation.

Large Language Models (LLMs) are prone to generating hallucinated content, which compromises their reliability in knowledge-intensive tasks. To address this challenge without sacrificing creativity, we propose HARPO, a reinforcement learning framework designed to jointly optimize faithfulness and creativity. HARPO incorporates a Hallucination-Aware Generative Reward Model (HA-GRM), trained via verifiable feedback, to assess both faithfulness and writing quality. A Selective Activation Mechanism (SAM) activates writing rewards only for outputs judged hallucination-free by HA-GRM, while a data curriculum progressively shifts training from creative writing to hallucination-centric tasks. On RAGTruth, our Qwen3-4B-based HA-GRM achieves a response-level F1 score of 78.08%, compared with 66.37% for the supervised fine-tuning baseline. Experiments on Qwen2.5 and Qwen3 models from 1.7B to 8B parameters show improvements in both faithful generation and writing quality. On Qwen3-4B, HARPO reduces the HA-GRM-judged hallucination rate on MultiHopRAG from 3.29% to 1.02%, while increasing the Arena-Hard-v2.0 creative-writing score from 16.95% to 27.54%.

The Muon optimizer derives its update rule for hidden linear layers by solving a local linearization of the loss penalized by the spectral norm, motivated by an RMS-stability argument for dense linear layers. Standard Muon implementations, however, exclude the input (embedding table) and output (language model head) layers from this principled treatment, for which they use AdamW instead. We present MuonIO, a single Muon-style update for both of these layers. For the language model head $\mathbf{L} \in \mathbb{R}^{V \times d}$, we motivate the use of the $2\to\infty$ operator norm, due to the Lipschitz continuity of the softmax output geometry, while for the embedding table $\mathbf{E} \in \mathbb{R}^{d \times V}$, we draw on the $1 \to 2$ operator norm, based on the one-hot input geometry identified by Bernstein & Newhouse (2025). The identity $\lVert\mathbf{L}\rVert_{2\to\infty}=\lVert\mathbf{L}^\top\rVert_{1\to2}$ then puts both matrices in the same vocabulary-oriented geometry: MuonIO applies a single normalized-vector rule, which appears as column normalization for $\mathbf{E}$ and row normalization for $\mathbf{L}$. Empirical evaluations demonstrate the effectiveness of our approach, with MuonIO reducing I/O optimizer state memory by 50% and I/O update FLOPs by $\sim$46% compared to Muon for 1B LLaMA pretraining on C4, while also improving validation perplexity.

Irregular All-to-All communication is a major bottleneck in expert-parallel Mixture-of-Experts (MoE) models. Even with fixed expert routing and placement, uneven utilization of parallel network Rails and incast can limit communication performance. We present RailWave, a phase-adaptive communication layer built on DeepEP that addresses these bottlenecks below the routing layer through spatial and temporal traffic shaping. RailBalance redistributes source traffic across eligible Rails using source-local information, while a reusable, topology-derived permutation schedule limits concurrent senders per receiver without rebuilding demand-dependent schedules for each communication phase. A lightweight calibrated selector chooses an execution path according to each phase's traffic characteristics and offline profiling results. On training-derived communication workloads from the 106B GLM-4.5-Air model, RailWave delivers up to 5.84x speedup on H800 and 4.36x on H20 over Native. Code is available at https://github.com/CyberSecurityErial/RailWave-EP.

Multi-task post-training of large language models (LLMs) aims to improve performance across tasks with unequal amounts of training data. Existing methods focus primarily on balancing task contributions during single-model training. Different task-balancing strategies can produce models with complementary strengths, creating opportunities for mutual distillation. However, the usefulness of cross-model supervision can vary across tasks, transfer directions, and stages of training. We propose Adaptive Mutual Distillation (AMD), a collaborative post-training framework that jointly trains two models with different task-balancing strategies. AMD evaluates candidate adjustments to distillation weights through short training probes shared across tasks, then uses task-wise validation scores to select an adjustment for each task and transfer direction. Across six benchmarks and three LLM backbones, both AMD models achieve higher average benchmark scores than supervised fine-tuning (SFT) baselines trained with the same sampling strategies. They also outperform the task-balancing methods evaluated in our experiments. Merging the two trained models can further improve their average benchmark score while yielding a single model for inference. The merged models outperform multi-task SFT by an average of 2.91 points across the three backbones.

We study hindsight-guided distillation for rare disease diagnosis on ZebraMap: a 1.5B student is fine-tuned on chain-of-thought traces from a 8B teacher that observes the ground-truth diagnosis during generation. Absolute accuracy remains low for all models - the task is hard at this scale - but within this ceiling a filtered variant (StudentF) achieves a small, statistically significant accuracy advantage over the teacher (p < 0.001), concentrated in better-represented diseases. The unfiltered student does not significantly outperform the teacher (p = 0.129), establishing that contamination filtering - not hindsight distillation alone - drives the gain. The gap traces to an artifact we term GT hallucination. Label-visible generation causes the teacher to embed "ground truth is X" phrases in its reasoning chain; SFT copies the pattern. At inference, the unfiltered student reproduces the phrase in 33.9% of cases, with severe accuracy degradation when the hallucinated label is wrong. A regex filter removing these slots reduces contamination to near-zero, producing the observed gain - though the effect remains small. We precisely quantify this gain-cost tradeoff, document frequency-dependent knowledge transfer absent from the RL-trained teacher, and characterize a calibration gap that SFT does not close - identifying both as directions for future work.