Neural solvers for the Flexible Job-Shop Scheduling Problem (FJSP) typically refresh graph representations and update feasibility after every operation--machine decision. However, only feasibility must remain exact at every step, whereas graph encoding is substantially more expensive and can potentially be refreshed less frequently. We present CoRS, which decouples them: an interval-level graph representation is reused across multiple sequential decisions, while a stepwise feasibility state is updated after every action. This two-timescale design amortizes graph encoding without parallelizing action generation or relaxing scheduling constraints. CoRS further develops a feasibility-preserving composite-interval decoding scheme with dynamic constraint tracking and boundary state synchronization, which reduces representation refresh frequency while maintaining continuous feasible schedule construction. Furthermore, to alleviate representation staleness caused by interval-level state reuse, CoRS introduces enhanced static features as a compensation mechanism. By providing more comprehensive structural information of the scheduling problem, these features improve the model's perception of overall problem characteristics. Extensive experiments on synthetic and multiple public FJSP benchmarks demonstrate that CoRS significantly reduces inference time while maintaining competitive solution quality, achieving favorable speed-quality trade-offs under tight wall-clock budgets and low-latency scenarios.
Flexible switching between reasoning trajectories (i.e., thought switching) has significantly enhanced the reasoning capabilities of Large Reasoning Models (LRMs). However, existing models often switch excessively yet fail to sustain promising reasoning thoughts, a phenomenon termed "under-thinking". While recent efforts suppress switching to mitigate this, such over-correction may discard valuable trajectories. To address this challenge, we propose Steady Thought (ST), a novel thought-level preference optimization framework. ST formalizes under-thinking as a preference issue at switching points, forcing single-path continuations from identified points to yield high-quality reasoning. Then, ST performs thought-level preference optimization by treating the newly generated response as preferred and the original one as dis-preferred. Experiments across multiple models and datasets show that ST effectively reduces token consumption, achieving reductions of up to 63.8%, while maintaining strong reasoning accuracy. Further analysis suggests that ST helps models acquire a generalizable ability to reduce redundant switching across domains and languages. Code is available at https://github.com/MIRALab-USTC/SteadyThought.
Collaborative edge large language model (LLM) systems often require repeated prefill computation when task contexts are transferred between agents. Reusing previously computed key--value (KV) caches can eliminate this overhead, but requires KV-cache delivery before the receiving agent begins decoding. Under strict deadlines and unreliable edge links, complete KV cache transfer may be infeasible, making the timely delivery of useful KV information critical. This paper proposes PriorityFEC, a utility-aware KV-cache handoff scheme that combines prioritized source transmission with adaptive loss recovery. PriorityFEC first organizes KV blocks into a utility-aware source queue to prioritize important information. It then employs ACK-sliding Streaming FEC and a lightweight Double Deep Q-Network (Double-DQN) scheduler to dynamically balance source delivery and repair transmission under deadline constraints. We implement PriorityFEC in a distributed edge LLM-serving prototype and evaluate it using Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct across multiple workloads. Experimental results demonstrate that PriorityFEC preserves more useful KV information and achieves substantially higher task accuracy than competing transport schemes, particularly when the deadline is insufficient to deliver the complete KV cache.
Automated approval systems let coding agents execute high-risk actions with minimal human oversight by delegating the authorization decision to a classifier, typically a large language model (LLM). Such systems, adopted in tools like Claude Code and Codex, remain vulnerable to failure modes such as scope escalation, where the agent takes actions exceeding what the user authorized. Existing benchmarks focus mainly on malicious requests, evaluating refusal under an implicit policy. We introduce AuthGateBench, the first policy-guided benchmark for authorization gating under benign requests. We implement a novel pipeline where an LLM generates contrast pairs with identical action and trajectory but differing request, flipping the label between block and allow. This allows us to assess whether models reason over context rather than the action alone. We further examine whether auxiliary environment signals can help models make better decisions. The resulting benchmark comprises 1,400 instances spanning five failure modes drawn from real-world Claude Code incidents. Evaluations show accuracy degrades unevenly across failure modes as models over-rely on internal safety alignment, and models still block most actions even when explicitly authorized, revealing unstable rather than deterministic decisions; retrospectively retrieved environment signals, however, substantially improve authorization accuracy.
World action models (WAMs) predict actions and future observations, typically from a reconstruction-based representation that carries noisy, redundant information which can complicate downstream predictions. We introduce LeWAM, a bidirectional transformer for forward, backward, inverse dynamics and policy prediction, on a decoder-free JEPA latent trained end-to-end through all four modes. We see the following benefits: 1) Alignment: linear probes read robot and object state from LeWAM's latent better than from a regular Le World Model (a forward-only JEPA world model), while the latent ignores visual distractors as well as LeWM does and far better than a reconstruction-based WAM. 2) Acting: Closed-loop evaluations of LeWAM match a regular flow-matching policy trained on the same encoder at matched size, while also providing a world model. 3) Planning: Sampling raw actions when planning with WAMs lets MPC exploit dynamics-model inaccuracies; planning in the noise space of the policy head instead improves the closed-loop performance of these WAMs.
Portable mobile-manipulation demonstrations can help alleviate data scarcity for embodied intelligence, but obtaining reliable, low-cost, and robot-free motion supervision from RGB observations remains challenging. Existing approaches often rely on teleoperation or specialized devices equipped with additional sensing hardware, while directly using estimated visual odometry (VO) trajectories can introduce inconsistencies due to accumulated drift and imperfect motion supervision. We present the Visual-Odometry-Conditioned Mobile Manipulation Interface (VOMMI), a portable demonstration collection and learning framework that connects portable RGB demonstrations to vision-language-action (VLA) post-training through offline trajectory reconstruction and online visual-motion conditioning. VOMMI synchronizes body and hand views to capture navigation context and local object interactions without requiring human-robot kinematic correspondence calibration. R2-VO refines offline demonstration trajectories using sparse geometric anchors and produces causal local-motion tokens over multiple prediction horizons for online policy conditioning. An action-group residual adapter incorporates these tokens only into the base branch. Experiments use a 500-trajectory portable for each task, with 75 trajectories held out for RGB-VO evaluation, and 200 robot demonstrations as references. Our policy, post-trained only on portable demonstrations, achieves 18.2% lower base-velocity error than a policy trained with robot-collected demonstrations, while maintaining comparable end-effector translation accuracy. Offline reconstruction reduces absolute trajectory errors for the body and hand streams by 24.6% on average relative to the best evaluated baseline for each stream. The complete system improves the mean success rate by 8.3 percentage points over OpenPI 0.5 across three real-robot tasks.
Mathematical discovery requires preserving partial ideas and the evidence supporting their reuse. We present Ansatz, a research harness whose continual graph memory separates model-accepted claims and dependencies from curated exploration, and Ansatze, a community platform for inspecting, guiding and checking this work. On First Proof, Ansatz recorded nine local verifier closures in each of two ten-problem batches; a partial result in the Second Batch also counted as a pass under the study's criterion. These model-assessed outcomes include continued searches and differ from expert validation. Component studies and re-review expose disagreements and their dependency reach. Open-problem investigations produced candidate results, with five Lean formalizations reported. The platform connects problem discovery, research records, manuscript analysis and formal checking while preserving distinct evidence states. Together, these systems support cumulative human-AI research through persistent investigation and inspectable contributions.
Judgment models such as \jev{} return, in a single call and without reasoning text, a probability for each described option. This makes them attractive as an agent's action-selection layer, but it is unclear which decisions they can be trusted with. We test \jev{} on reflection tests, one-shot matrix games, the text game ALFWorld and robot control, and find a sharp boundary. \jev{} succeeds when the right option can be judged from what the input describes, which we call \emph{evaluation}. Specifically, it solves 99\% of the counterintuitive Cognitive Reflection Test questions. However, it fails when the right option depends on \emph{simulation} (i.e., predicting something not in the input), such as the opponent's action or the subgoal that must come first. In games, \jev{} plays suboptimally as if its rational opponent acted at random, because the opponent's action is not given. In ALFWorld, \jev{} favors commands that mention an object or place named in the task description. For example, given the task ``put a clean knife in the drawer'', \jev{} carries an unwashed knife straight to the drawer instead of first washing it at the sink. Surprisingly, many of these failures are not due to a lack of knowledge. Asked separately what the opponent will do, \jev{} usually answers correctly, and it responds well given the opponent's action. It fails when one call must both perform the simulation and evaluate based on it. This suggests letting code make the prediction or simulation. When code supplies it, such as a lookahead in ALFWorld and physics simulation in robot control, \jev{} becomes an expert controller through its general evaluation ability.
Recurrent models trained from scratch have recently become competitive on ARC-style reasoning tasks, but the usual framing around small recurrent backbones overlooks two important parts of the system: task-conditioned memory and synthetic augmentation data. We study this regime through CHARM, a compact hybrid ARC model that combines recurrent reasoning with structured task memory, synthetic data, and inference-time aggregation. In existing approaches, task-conditioned memory supplies a large hidden source of capacity, reaching more than 30x the size of the recurrent backbone. We introduce a compositional sparse embedding (CoSE) for task conditioning that reduces learned task-memory parameters by over 90% while improving pass@2 in controlled ARC ablations. We also introduce Re-ARC2, a set of programs that generate and verify new examples for all ARC-AGI-2 training tasks. Combining these ingredients, our system reaches 84% pass@2 on ARC-AGI-1 and 46.7% pass@2 on ARC-AGI-2 public evaluation. Structured memory also enables adaptation with only 512 parameters per unseen puzzle, and its benefits generalize to other domains. Our code, dataset, and model checkpoints are available at https://github.com/water-vapor/charm.
Inference in LLMs is conventionally a fixed-depth, fixed-order forward pass through every layer, regardless of how difficult the input is. The human brain does not work this way. Using the thalamus as a central hub, it routes information flexibly to all regions of the cortex according to demand. Li et al. (2026) recently showed, with program-of-layers (PoLar), that transformers can be given an analogous flexibility if their layers are treated as a library of functions rather than a fixed sequence. Performance improves over the standard forward pass when each input is dynamically routed through an adaptive sequence of skipped or repeated contiguous layer blocks. Reconstructing PoLar's diagnostic Monte Carlo Tree Search in more detail across 6 models, we reproduce several of its findings: skipping outperforms the standard pass, repeating outperforms skipping, and combining both outperforms either alone. Shorter programs suffice for easier questions, while harder questions require more layer repeats. For their main claim, the learned router's gain, we find what drives it. Under their setup, the router gains by repairing the format of the model's answers, which a single fixed program, selected on validation questions, does better. In our setup, with a prompt that already fixes the format, a larger data split and our search, its top prediction collapses to the standard pass on every model, although its top-5 programs together still gain. Beyond reproduction, we find in our setup that a small fixed menu of programs solves far more questions than the standard pass, and that programs which correct errors are brittle: undoing a single edit typically breaks the correction. PoLar mirrors principles of thalamo-cortical coordination between cortical-area-like transformer layers, although a working thalamus analogue, a router that picks a better program per input, remains to be found.
Most work on improving large language models treats accuracy as the sole objective. We argue that the harness, the Python code surrounding the model that constructs prompts, routes calls, and parses outputs, is a first-class design surface whose quality is inherently multi-objective: an accurate harness that refuses no unsafe request, or that consumes an order of magnitude more tokens, is not a good harness. We present Meta-Harness, a system that casts harness design as search over three per-domain objectives (accuracy, behavioural safety, and token cost) solved by an agentic proposer (Claude Code) with full filesystem access to prior harness source, execution traces, and scoring artifacts. Our central finding is that a singlephase joint-reward proposer (MoMHa) outperforms every alternative, including a two-phase "accuracy then tokens" ablation, scalar-only feedback, and an accuracy-only baseline. We evaluate on seventeen domains: seven synthetic capability suites, seven real-world public benchmarks (HumanEval, MBPP, Spider, FEVER, MMLU-Pro, LawBench, NuminaMath), and three U-SafeBench-derived user-specific safety domains, using a 12-model fleet spanning four families. On the synthetic track MoMHa achieves a joint mean of 0.482 versus 0.198-0.422 for ten baselines, winning $7 / 10$ per-domain columns; on the real-world track it scores 0.461 versus 0.377 for the strongest baseline (DSPy), winning 5/7 columns, demonstrating that harness strategies transfer to unseen benchmarks without retraining on 8 of 12 target models. MoMHa attains the highest measured behavioral safety composite (U-SafeBench, 0.781) and uses 95 fewer tokens per example than the two-phase alternative. We will release all harness code, evaluation infrastructure, and crossmodel logs.
Genetic programming and its variants, such as grammatical evolution, are widely used in symbolic regression to derive mathematical expressions from multivariate data. In addition to predictive accuracy, models are appreciated for their potential to provide interpretability, offering explicit equations that relate input variables to outcomes. However, achieving interpretability and plausibility remains challenging, as evolved models may be complex or scientifically inconsistent. In this study, we explore whether large language models (LLMs) can assist in improving the interpretability of symbolic regression models generated by evolutionary computation methods. Building on our previous work on estimating body fat percentage using grammar-based genetic programming, we investigate LLMs as post-processing tools to analyze and rank evolved expressions by interpretability and physiological plausibility. Four symbolic expressions are analyzed by three LLMs over three repeated runs, and the resulting interpretations and rankings are assessed by a panel of three clinicians. Across the three LLMs, comparative model-ranking outputs received more favorable clinician assessments than isolated term-level interpretations. However, the LLMs also produced physiologically and mathematically questionable explanations, indicating that they are better suited to comparative auditing under expert oversight than to autonomous validation.
Large language model (LLM)-based multi-agent systems (MAS) achieve strong performance by employing specialized multiple agents, yet their performance depends on the prompt design of each agent. For MAS prompt optimization, textual gradient methods that guide prompt updates using natural-language feedback have emerged as a leading paradigm. In this paper, we identify limitations in two stages of existing textual gradient approaches: gradient extraction and gradient aggregation. In gradient extraction, previous works select a prompt to update without verifying whether modifying it resolves the failure, and derive gradients without agent-level supervision over the corresponding agent's intermediate output. In gradient aggregation, individual gradients are randomly grouped and concatenated, often mixing unrelated failure modes and producing prompts that fail to generalize. To address these limitations, we propose AgentGrad, a prompt optimization framework for multi-agent systems based on sequential intervention and semantic textual gradient abstraction. For each failure, sequential intervention injects a hint into one agent at a time to identify the target agent whose modification resolves the failure. The modified output of the target agent then serves as agent-level supervision for extracting a fine-grained gradient. Semantic textual gradient abstraction clusters semantically similar gradients to prevent mixing unrelated failure modes, and abstracts each cluster into a generalized gradient that captures the shared corrective pattern. Experimental results show that AgentGrad achieves state-of-the-art performance across five MAS benchmarks while reducing wall-clock optimization time by 2.5 times and optimization cost by 21.8% on average compared to the next-best baselines. Our code is available at https://github.com/mlvlab/AgentGrad.
A benchmark score is a joint property of the model, the evaluation harness, the elicitation budget, the sampled population, and contamination status. Leaderboards publish the model and the score, so capability and leakage stay observationally equivalent. Existing taxonomies classify contamination for automated detection, not the question a reporter faces at publication: given the mitigations already applied, which validity threats remain open? We introduce a taxonomy organized by the mitigation each type defeats -- direct, derivative, temporal, distributional, and acquired -- spanning training-time and evaluation-time leakage. Holding out a private test set closes the first alone. The fifth is acquired during the evaluation itself; because it is a property of one run, it must be recorded with the reported score rather than with the benchmark release. We operationalize it as a four-field disclosure protocol in which "unknown" is a valid entry, released under CC BY 4.0 with a JSON Schema, a validator, and worked examples. Two coders external to the design team applied a pre-registered instrument to 41 documents. Per-variable linear-weighted $\kappa$ runs from 0.00 to 0.35 (median 0.21) over 29 main-pass documents against a single-coder test-retest ceiling of 0.84, collapsing under the class skew the registration anticipated; pooling raises it to 0.46 through chance correction rather than better agreement. Two variables fall below the prevalence-robust threshold registered in advance: strata reporting and the acquired type introduced here. Disagreement concentrates on when a variable applies rather than on what a document states. Elicitation budgets are reported in 13% of documents, and no document addresses all five types. The contribution is the taxonomy, the score-side artifact that follows from it, and a pre-registered measurement of instrument reliability and current disclosure.
Protein co-abundance clusters preserved across tissues can nominate shared disease mechanisms and therapeutic targets, particularly where proteins implicated in organ-confined diseases converge in other tissues. Yet cross-tissue proteomic studies have examined pre-selected tissue pairs, leaving most combinations, and non-obvious relationships among them, unexplored. Exhaustive comparison requires consistently assessing thousands of candidate clusters from hundreds of tissue-pair networks across heterogeneous evidence, a scale beyond manual curation. We present an LLM-agent framework for large-scale, evidence-grounded comparison of tissue-specific protein co-abundance networks. From sample-level proteomic data, it constructs tissue networks, derives conserved clusters for each tissue pair, integrates evidence from expression atlases, interaction and complex databases, pathways, disease associations and the literature, and proposes mechanistic hypotheses in which every conclusion cites its evidence. Applied to all 231 pairs among 22 human tissues and sample groups, it identified 1,825 conserved clusters in 216 tissue pairs. Co-abundance relationships were shared beyond chance in 76% of pairs, including unrelated organs, and shared clusters were only weakly associated with abundance similarity. Clusters were enriched 3.7- to 6.9-fold for known interactions and complex memberships, and 379 of 793 testable clusters replicated in independent CPTAC proteomes. Some clusters pointed to relationships absent from current databases: one shared by liver and bone marrow linked three genes causing inherited white matter disorders, suggesting a common mechanism via the integrated stress response. Conserved co-abundance clusters thus offer a hypothesis-generating resource for studying disease proteins beyond the affected organ. Code: https://github.com/jacob20030316-crypto/Protein_Co_Abundance_Agent
We present a first application of sparse-autoencoder-based mechanistic interpretability to particle physics. Studying a neutrino foundation model pretrained on IceCube data and fine-tuned for direction reconstruction, we identify a validated atlas of physical concepts in the model representation, using a strict validation protocol consisting of held-out tests, matched nuisance controls, and replication across independent dictionary trainings. Causal interventions show that the direction head barely draws on this atlas. Motivated by this underused information, we train an uncertainty head on the same event-level representation to predict the model's angular reconstruction error. Unlike the direction head, it depends causally on quality and brightness features from the atlas. These results suggest that mechanistic interpretability can reveal learned latent physics encoded within a model's internal representation and help design downstream tasks that exploit it.
Large-scale rollouts have become a core component of modern LLM systems, spanning reinforcement learning (RL) post-training, on-policy distillation, and sampling-heavy evaluation. Unlike online serving, which typically targets request-level latency and throughput, synchronous rollouts must collect a complete batch before proceeding. A few long-tail generations can therefore dominate an entire rollout step, and uniform routing can slow these generations further by placing them in high-concurrency decoding batches. We present TailSieve, a partial-rollout-guided framework that jointly controls tail routing and replica allocation. Motivated by the observed persistence of prompt-level tail tendencies across policy updates, TailSieve uses partial-rollout completion status as a training-free signal for identifying candidate tail groups. Selected prompts are combined with newly sampled prompts, and all responses are generated from scratch under the current policy. A hierarchical controller uses response-work history and a serving model: a fast inner loop adjusts the number of isolated groups to balance completion times, while a slower outer loop adapts how many replicas serve these groups. The resulting low-concurrency execution also enables route-specialized speculative decoding with MTP or DFlash. Across the evaluated configurations, TailSieve achieves up to $1.67\times$ speedup with routing alone and $2.59\times$ with routing and speculation combined, both over uniform group routing.
Parliamentary proceedings are a primary record of democratic deliberation, yet their volume and fragmentation make multi-perspective access difficult for citizens, journalists, and researchers. Applying Retrieval-Augmented Generation (RAG) to parliamentary transcripts introduces three specific risks: dominance of the most frequent speakers, inability to weight speakers according to topical expertise, and citation misattribution in politically sensitive text. We present ParliamentRAG, a RAG system for the Italian Chamber of Deputies that addresses these risks jointly. Its core contribution is a topic-dependent authority model that estimates each speaker's authority as a function of the current query, combining interpretable components such as profession, education, and previous interventions. Given a user query, the system retrieves relevant speech chunks, identifies topic-relevant experts across parliamentary groups, and generates a summary synthesizing their perspectives, accompanied by supporting quotations. ParliamentRAG is evaluated against Google NotebookLM on 15 policy topics via a two-level protocol combining automated metrics and blind A/B human evaluation by six domain experts. The system achieves higher coverage across political groups (0.97 vs. 0.95), perfect quotation faithfulness (1.00 vs. 0.95), and stronger expert preferences on source-related dimensions, while NotebookLM remains stronger on prose-oriented dimensions.
Reinforcement Learning with Verifiable Rewards (RLVR) makes Multimodal Large Language Models more accurate, but the gains are brittle: simply paraphrasing a question or changing the prompt template can degrade them, which challenges reliable deployment in high-stakes scenarios like medical VQA. We trace this to two issues of the standard RL objective. First, the binary verifier conflates format with content, so the reward signal cannot tell a wrong answer apart from a misformatted one. Second, the training distribution covers only a thin slice of the real-world prompts that the model might meet at deployment, so policies that perform well on the training distribution can behave differently under unseen prompts during test. Both failures call for a robust post-training method that helps the policy cover a broader distribution of semantically equivalent prompts, and we identify two measures that help achieve this objective: separating format from semantics in the reward, and applying policy invariance across perturbed prompts with equivalent semantics. We therefore propose Prompt-Invariant RLVR (PIRL), consisting of a dynamic trinary reward and a consistency regularizer based on an embedding-space adversary. Under stress testing, PIRL's average accuracy on benchmarks drops by only $\le 1\%$, where GRPO drops ~3%. On dynamic evaluation, PIRL also achieves the smallest performance drop.
High-quality long-horizon training data for terminal agents is costly and difficult to scale, as manual authoring is expensive and direct LLM generation often produces inconsistent instructions, environments, solutions, and verifiers. We present Recursive Synthetic Terminal Tasks (RST), a recursive verified synthesis framework for constructing long-horizon terminal-agent tasks at scale. Starting from verified seed tasks, RST extends the reference solution, realigns the verifier and instruction to the new workflow, validates the result in a fresh sandbox, and reuses accepted tasks as seeds for subsequent rounds. Across fifteen recursive rounds, RST produces 37,484 synthesized terminal-agent tasks at roughly $0.05 per task. Task difficulty increases substantially over rounds: the median reference solution grows from 67 to 374 lines, the median number of executed commands grows from 40 to 244, and DeepSeek-V4-Pro pass@4 drops from 90% at R1 to 2.5% at R15. To demonstrate training utility, we collect rejection-sampled Qwen3.5 trajectories on the synthesized tasks and use them for supervised fine-tuning. Fine-tuning on these trajectories improves Qwen3.5-27B and Qwen3.5-122B-A10B by up to 10 points on Terminal-Bench 2, Terminal-Bench Hard, and Long-Horizon Terminal Bench, while agentic PPO lifts Qwen3.5-27B to 49.44%, 32.00%, and 22.07% on the three benchmarks, corresponding to relative gains of 20.0%, 41.2%, and 21.9% over the base model. Moreover, after 15 rounds, the recursion shows no ceiling: synthesis yield and validation rates remain stable as difficulty keeps climbing, indicating that the process can continue well beyond the scale reported here.