Computer Science & Mathematics

Multi-stage analytical tasks require LLM agents to use intermediate evidence to decide whether to continue, redirect or stop. Self-reflection, in which an agent critiques its output, is widely studied, yet a critique does not show whether evidence changed the next action. We call that capability self-restraint. Existing agent benchmarks largely measure task completion or output quality and leave it untested. DMAIC-Bench evaluates it with 97 seeded tasks across 15 industrial cases in four sectors, built around Lean Six Sigma's explicit statistical prerequisites. Paired counterfactual variants require opposite actions at the same workflow position, distinguishing warranted restraint from indiscriminate stopping. We assess execution traces with a private scorer and final claims through independent expert review. Across 13 models from six providers, execution-level routing accuracy averages 92% on matched GO decisions and 19% on NO-GO decisions, counting any forbidden downstream computation as a route error. Execution traces show agents reporting correct diagnostics before performing analyses that those results disallow. Separately, blinded expert review of the 11 audited agents estimates that reports act on invalid evidence at 47% of reached NO-GO decisions, with an interval from 31% to 61%. Explicit stopping permission helps some models without consistently restoring correct decisions at both GO and NO-GO forks. Sixteen life-sciences tasks in bulk ribonucleic acid sequencing (RNA-seq) and case-control genome-wide association studies (GWAS) show that the same framework can evaluate evidence-dependent decisions beyond industrial workflows. Project code and data are available at: https://github.com/psptpx/DMAIC-Bench-public.

Multi-step retrosynthesis requires recursively transforming a target molecule through successive reaction steps into purchasable building blocks, and recent work has begun to draw on large language model (LLM) reasoning to tackle the task. In its latest evolution, LLM reasoning has entered the long-horizon regime: models pursue objectives across interdependent steps, interact with environments, and adapt to feedback. This shift has markedly advanced LLM capabilities, as illustrated by breakthroughs in AI-driven mathematical research. At this frontier, we introduce Solvent, an 8B LLM that successfully extends long-horizon reasoning to multi-step retrosynthesis. Through a two-round post-training curriculum, Solvent learns to interleave iterative synthesis planning, external availability verification, and adaptive route revision. Experiments show that, following a $2.1\times$ performance boost from RL, Solvent scores 61.86% across four benchmarks, surpassing mid-scale language models and prior search-based retrosynthesis systems, and rivaling frontier trillion-parameter LLMs.

Why can a task-preserving rewrite turn a correct answer into a failure while a larger internal change leaves the answer intact? We develop a direction-sensitive, readout-relative account of semantic fragility: heterogeneous language changes can share geometric and readout relationships without sharing an input-independent failure direction. Clean decision reserve, representation transformation, and response alignment provide complementary descriptions of this process. An exact longitudinal-transverse decomposition identifies directional information discarded by scalar transmission. A compact reserve-and-transmission model predicts failures across held-out families and items in Qwen3-32B and Gemma 4 31B, with useful but parameterization-dependent transfer to ReClor. Matched interventions show that orientation carries causal information beyond scalar geometry. Among naturally failed rewrites, attention-component tests on disjoint original items and subsequent answer-order controls favor the natural paired direction over tested random and same-answer-letter donor directions. Yet failures rarely persist across option orders, and current-order corrections use recomputed clean references. Shared predictive regularity thus coexists with context-dependent directional effects. Numerical path integration reduces, but does not eliminate, finite-patch approximation error within a fixed context; a larger discrepancy from the natural rewrite effect remains. A complementary fixed-CoT cue study extends this account to influence unacknowledged in recorded reasoning: readout associations transfer across domains, although the positive-mean causal criterion fails and signed-effect prediction remains post hoc. Together, these results support shared relations among reserve, transformation, and response while distinguishing local causal sensitivity from a complete explanation of natural failure.

The interior sound of an electric vehicle is a superposition of several distinct noise contributions - wind noise, tire noise, powertrain (electric drive unit and inverter) noise, and auxiliary-drive noise - whose relative balance governs both the overall loudness and the perceived sound quality in the vehicle interior. Decomposing a recorded vehicle interior signal into these source contributions is a prerequisite for targeted acoustic engineering, but requires effortful measurement campaigns as well as expertise in the field of transfer path analysis. This work formulates the decomposition as a supervised, binaural source-separation problem with four target source contributions and adapts a hybrid time-spectrogram transformer architecture, originally developed for music demixing, to the domain of sound source separation for technical systems. To disambiguate source contributions, conditioning is introduced in the form of operating-state information, including electric drive speed, electric drive load, vehicle speed, and auxiliary-drive speed. Since isolated real-world reference stems are costly to obtain, the model is trained on a physics-based parametric data generator that synthesizes labeled component sounds from randomized virtual-vehicle hardware configurations and driving scenarios, so that the ground-truth contribution of each source is known by construction. Validation on both synthetic and measured vehicle interior sounds probes the achieved separation performance and underscores the feasibility of applying the proposed method for acoustic contribution analysis in technical systems.

Geographic regions are multimodal: satellite imagery captures visual morphology, text conveys human-interpretable semantics, and map structures describe urban context. Existing geospatial contrastive methods typically align image-text or image-location pairs, underusing structured map data. We introduce GeoLEON-CLIP, a graph-anchored contrastive framework that aligns three views of each H3-indexed region: an OpenStreetMap-derived graph embedding, a satellite image crop, and a textual description. Frozen pretrained LEON graph embeddings provide a semantic anchor between vision and language. Frozen graph, image, and text encoders are connected by trainable residual projection heads and trained with a symmetric tri-modal contrastive objective. Experiments show strong graph-image and graph-text alignment; graph-image retrieval achieves the best cross-modal performance in a mixed-resolution pool. To evaluate the anchoring hypothesis, we propose chained cross-modal recall $\mathrm{R}@A@B$, which routes retrieval through the graph modality and compares it with direct retrieval under a matched candidate budget. At $\mathrm{R}@1@1$, image$\rightarrow$graph$\rightarrow$text and text$\rightarrow$graph$\rightarrow$image outperform direct image-text retrieval by 4.8 and 4.4 points, respectively, despite using a single intermediate cell. Downstream probes show task-dependent gains, with the largest improvements on Airbnb and Beijing housing. These results indicate that OpenStreetMap graph structure can anchor multimodal representations combining maps, imagery, and language.

Generalist robot manipulation requires a single vision-language-action (VLA) policy to execute diverse skills across scenes and instructions. However, current multi-skill VLA policies often leave two aspects of an instruction insufficiently modeled: which primitive skill to execute and where/how to contact and move objects. We propose MAST-VLA (Multi-skill Adaptation with Skill-conditioned experts and Trajectory-affordance priors), a dual-conditioning framework for robot manipulation. MAST-VLA introduces skill-conditioned LoRA Mixture-of-Experts (LoRA-MoE) adapters in the LLM backbone, where skill-query tokens and routing biases guide low-rank expert specialization, and converts a predicted affordance heatmap into a patch-level trajectory-affordance spatial prior for the action head. On the CALVIN ABC$\rightarrow$D long-horizon multi-skill cross-scene generalization setting, MAST-VLA achieves state-of-the-art performance, reaching an Avg. Len of 4.65 over five-subtask evaluation sequences. Real-robot multi-skill sequence experiments further validate its deployability under real perception and control noise.

AI coding agents are increasingly integrated into software development workflows. However, most coding agent studies rely on benchmarks similar to SWE-bench, which assume autonomous agents operating in a single trajectory. This assumption does not mirror real-world usage, where developers interactively collaborate with agents. We address this gap by studying human-in-the-loop agent trajectories in the wild, focusing on 7,581 Copilot-authored PRs with publicly accessible session data. We find that 58% of Copilot PRs have no human intervention, and that when developers do intervene, collaboration is brief (median of 2 Human-AI loops). Merged PRs also tend to have fewer initial trajectory iterations than unmerged ones. We also find that the trajectory following the first human intervention exhibits a significant drop of 20.8 iterations and 8.4 min in duration. However, trajectories following later interventions show minimal iteration drops and no significant duration change, signifying that most of the work happens in the initial trajectory. Our qualitative analysis of illustrative PRs explains this pattern: many developer interventions are small, focused requests that Copilot completes in shorter later trajectories. We also illustrate that developers collaborate with Copilot in multiple ways, including delegated work through feedback as well as adding their own commits. However, we find that developers may need to guide Copilot through multiple (sometimes repeated) interventions in the PR, or even across multiple editions of the PR.

This paper presents the engineering implementation methods of the SKG inference framework and provides complete case studies across four domains. Case 1, 2D Dissipative Mechanical System: OOD MSE is 1.69 to 4.12 times lower than the best neural baseline; the state-dependent metric yields a 10^8-order improvement under controlled conditions. Case 2, Target Detection Sim-to-Real: Real-machine mAP increases from 0.2564 to 0.4964 (+24.00 points, single-run experiment, 32 images, point estimate), and AP75 increases from 0.1736 to 0.5765 (+40.29 points), with no real-machine retraining required. Case 3, Sequence Modeling sine to square: OOD MSE is 0.0296 versus Transformer 0.0641 (mean of 5 seeds, p=0.024). Case 4, Molecular Dynamics MD17: ID@100 ADE is 0.0285 versus ESTAG 0.1467; OOD@100 is 0.0316 versus 0.1411 (single seed). For each case, we present the objective, applied SKG theory, parameter settings, baseline comparison protocol, data results, and sources of advantage. Following the cases, we provide the complete methodology: pre-validation, task invariant construction, three-stage training, residual training, and the four theorems and one proposition supporting these methods. This paper not only demonstrates effective configurations of SKG across four domains but also honestly reports the variance of single-run experiments, limitations of single-seed results, negative updates caused by parameterization mismatch, and practical experience of using engineering thresholds as diagnostic signals rather than hard constraints.

Video moment retrieval aims to localize a query-relevant temporal segment in an untrimmed video. Existing methods typically perform video-text alignment in Euclidean space, where representations are embedded in a flat geometry. However, accurate moment localization requires organizing query-relevant evidence across multiple grounding states, where target regions, boundary transitions, and contextual distractors exhibit distinct temporal structures and relevance patterns. Such Euclidean representations lack an explicit geometry to jointly capture these relationships, motivating us to explore structured temporal evidence organization from a geometric perspective. Specifically, we propose Hyperbolic Temporal Evidence Organization (HTEO), a geometric framework that organizes query-aware temporal memory in hyperbolic space. HTEO maps temporal representations into a Poincaré ball and organizes temporal evidence through Radial Evidence Organization (REO) and Orthogonal Angular Refinement (OAR). REO adjusts temporal tokens' radial positions to encode the structural differences among temporal evidence, while OAR enhances query-conditioned semantic alignment through orthogonal rotation. The organized representation is then projected back to the Euclidean tangent space and injected into the decoder through a lightweight residual memory update, enabling hyperbolic geometric cues to be incorporated into standard Transformer decoders while preserving the original cross-modal memory. Furthermore, we introduce radial structure regularization and angular relevance supervision to explicitly guide temporal evidence organization and query-conditioned semantic alignment, respectively. Extensive experiments and analyses across multiple benchmarks demonstrate that HTEO improves temporal localization while providing an interpretable geometric organization of temporal evidence.

System One models, or decision models, answer a question about an input with a calibrated probability instead of generated text. Within weeks of the first release, in September 2026, they were used as rerankers. We test whether the decision fine-tuning behind them makes a better reranker than plain reranker fine-tuning of the same model. For six Qwen backbones from 0.6B to 27B parameters and ten TREC DL and BEIR collections, we compare the untuned model, asked zero-shot, with its decision fine-tunes and its reranker fine-tunes, two of which we train ourselves. In our work, we draw the conclusion that decision fine-tuning never makes the better reranker. From 0.6B to 9B, reranker fine-tuning wins on every backbone by 5 to 29 points nDCG@10, also on zero-shot collections, and at 27B the two tie. Up to 9B, decision fine-tuning adds nothing over simply asking the untuned model. Moreover, we find that the calibration advantage of decision models depends on the question. Asked whether a passage answers the query, they are better calibrated than the published reranker fine-tunes from 4B on. Asked a slightly different question from prior work, they lose this advantage. Lastly, calibration does not make up for a worse ranking: up to 9B, a reranker fine-tune keeps the better passages at a fixed threshold. If you need a reranker, fine-tune a reranker.

Point-supervised Temporal Action Localization (PTAL) aims to localize action instances using only a single annotated timestamp for each instance. Such sparse supervision offers limited guidance for modeling the structural organization of an action instance. Existing PTAL methods primarily infer action regions from classification responses, leaving the structural organization of multiple action-related parts within each instance insufficiently modeled. To address this problem, we propose DeformStruct, an instance-level deformable structure modeling framework for PTAL. DeformStruct progressively constructs the structure of each action instance through three stages. First, Structural Anchor Mining discovers multiple anchors from annotated points and action-responsive regions, providing structural nodes grounded in the current instance. Second, Deformable Part Alignment associates latent structural parts with these anchors through deformation-aware matching, allowing their temporal supports to adapt to instance-dependent variations. Third, Relational Structural Organization explicitly models the relative temporal relations among the aligned parts, forming a relational representation that characterizes how different action-related parts are organized together. Extensive experiments on THUMOS14, ActivityNet-v1.3, GTEA, and BEOID demonstrate consistent improvements in point-supervised temporal localization and validate the effectiveness of deformable structural modeling.

LLM agents retrieve skills from public hubs that hold up to about 100,000 entries and serve many models. As models improve, a hub must find the skills a given model has internalized, meaning it can do without them, and retire them for that model, yet internalization has not been measured with known reliability. Published rules that compare runs with and without a skill retire on the gain alone. The gain is near zero both for an internalized skill and when neither the model nor the skill delivers the capability, a state we call uncovered. We instead withhold the skill, score the model on a task synthesized from it, and trust a high unaided score only when another model gains from the skill, since nearly half of high scores come from tasks no model benefits from. Under four models, a published gain rule applied with one task per skill removes over half of the uncovered skills. On SkillsBench, 28 of its 31 retirements rest on probes no run passed in full, while retiring on the unaided score costs no detectable reward. Unaided verdicts reproduce across tasks more often than the gain rule's (Cohen's kappa 0.42 against 0.23) but differ across models, so a hub must read them per model. We release 1,251 skills with tasks and item scores under four models.

Reinforcement learning with verifiable rewards (RLVR) improves multimodal reasoning, but uniformly assigning a single response-level advantage to all tokens overlooks the distinct roles of visual observations and subsequent reasoning. Visual affinity alone is insufficient: predictive states may align with both image and prompt-context representations, and upweighting valid visual observations can amplify their negative-advantage updates when later reasoning fails. To address this issue, we propose \textbf{Outcome-Conditioned Visual Credit (OVC)}, a plug-in framework that separates visual-association measurement from token-level credit assignment and is adaptable to policy-optimization methods with uniform response-level advantage allocation. Relative Visual Encoding constructs a context-relative visual coordinate by contrasting each predictive state's affinity to image and prompt-context representations. Outcome-conditioned Evidence Calibration combines this coordinate with uncertainty-scaled predictive features and calibrates their contributions to outcome-discriminative evidence through mean contrasts between positive- and negative-advantage responses, scaled by regularized within-side variation. Visual association thus informs credit through observed outcome differences rather than directly determining update priority. Normalized Credit Allocation conditions the calibrated evidence on the response's advantage sign to produce positive token weights with unit global mean. These detached weights preserve advantage signs while modulating individual token-gradient contributions. OVC reuses existing rollouts without auxiliary reward models, process annotations, or inference-time modules. Instantiated through GSPO-token, OVC retains sequence-level clipping and consistently improves six-benchmark mean accuracy over GSPO across three models, with gains of 1.86-2.39 percentage points.

Music-driven group dance generation must decide not only how each dancer moves, but also what formation the group should form, when it should change, and who should execute each transition. Existing methods can generate orderly and rhythmically synchronized group motion, but remain limited in formation evolution and inter-dancer interaction. In human choreography, formation transitions often align with changes in high-level musical structure, such as the progression from an introduction to a chorus or from a chorus to a climax. We present ChoreoStudio, which makes choreography planning explicit. A music disentangler first decomposes the audio into frame-level rhythmic features and segment-level structural features. Given the segment-level features and sparse labels, a frozen omni-modal model acts as a scriptwriter, producing a global semantic description that guides the Director's choreography planning. Based on this context, the Director produces three coupled interaction decisions: a centroid-relative point set specifies what formation to produce, a transition gate determines when it should change, and a mobility allocation indicates who should execute the transition. The transition gate and mobility allocation are further conditioned on the planned geometric change, coupling formation targets with their temporal and dancer-specific execution. At a finer temporal scale, frame-aligned music features condition the diffusion generator to control detailed body motion and preserve beat-level alignment. The segment-level interaction plan is injected in parallel through frame-aligned instruction tokens and cross-dancer attention biases, coordinating how the generator realizes each transition. Experiments on AIOZ-GDance demonstrate that ChoreoStudio follows the planned interactions while achieving state-of-the-art performance in motion quality, group organization, and formation dynamics.

Large Language Model (LLM) agents have enabled automated code evolution for algorithm discovery. However, current recursive self-improvement (RSI) research for sequential recommendation faces a dilemma: general-purpose frameworks lack domain-specific feedback, while recommender-specific frameworks rely on linear evolution. Crucially, naively integrating them fails when applied to Generative Recommendation (GR). Unlike end-to-end models, GR relies on a module-cascaded pipeline (tokenization, pre-training, and alignment). In this paradigm, existing evolution methods suffer from: i) cost blindness, squandering budgets by ignoring stage-asymmetric retraining overheads; and ii) short-sighted evaluation, discarding promising upstream innovations that temporarily regress metrics before downstream adaptation. To bridge these gaps, we propose TGR-EvoDAG, a multi-agent code evolution framework tailored for GR. TGR-EvoDAG models code evolution as a Directed Acyclic Graph (DAG) to maintain multiple evolution branches and merge complementary innovations, unifying advanced search with rich GR diagnostic feedback. We further introduce stage-aware cost planning to optimize GPU-hour budgets, and active downstream co-adaptation to navigate temporary regressions and unlock cross-stage gains. On public Amazon datasets, TGR-EvoDAG enables the first autonomous evolution of a complete GR pipeline, improving offline metrics by up to 37.4% and Agent4Rec simulated user satisfaction by 19.7% over the strongest baseline. Furthermore, in industrial deployment on Tencent's large-scale commercial media platform, the evolved GR achieves a 12.16% relative improvement in offline Hit@1 and a statistically significant 0.40% increase in online click-through rate (CTR).