跳到正文
研踪PaperTrail

论文库

检索计算机领域论文,按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+1680新增论文
+1333其中 AI 相关
+187已有论文修订
+59新增 AI 导读

arXiv 最近完成扫描:2026-11-18 04:28 · 全站最近入库:2026-11-17 23:35各信源进度 →

统计口径与数据新鲜度

统计 2026-11-17 09:50 — 2026-11-18 09:50(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-18 04:28最近发现新增版本:2026-11-17 23:35库内论文最新更新时间:2026-11-17 02:59

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

25,786 已收录论文21,491 AI 相关7,725 有修订112 HF 社区发现

浏览论文库

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 5,970 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-18 04:27(北京时间)
版本更新 · v5arXiv:2602.04872

Multi-layer Cross-attention is Provably Optimal for Multi-modal In-context Learning

Nicholas Barnfield, Subhabrata Sen, Pragya Sur

原文摘要

Recent progress has rapidly advanced our understanding of the mechanisms underlying in-context learning in modern attention-based neural networks. However, existing results focus exclusively on unimodal data; in contrast, the theoretical underpinnings of in-context learning for multi-modal data remain poorly understood. We introduce a mathematically tractable framework for studying multi-modal learning and explore when transformer-like architectures can recover Bayes-optimal performance in-context. To model multi-modal problems, we assume the observed data arises from a latent factor model. Our first result comprises a negative take on expressibility: we prove that single-layer self-attention models fail to recover the Bayes-optimal predictor uniformly over the task distribution. To address this limitation, we introduce a novel, linearized cross-attention mechanism, which we study in the regime where both the number of cross-attention layers and the context length are large. We show that this cross-attention mechanism is provably Bayes optimal when optimized using gradient flow. Our results underscore the benefits of depth for in-context learning and establish the provable utility of cross-attention for multi-modal distributions.

stat.ML · cs.AI · cs.LG

版本更新 · v3arXiv:2507.01781

Symbolic Branch Networks: Tree-Inherited Neural Models for Interpretable Multiclass Classification

Dalia Nöth-Rodríguez

原文摘要

Symbolic Branch Networks (SBNs) are neural models whose architecture is derived directly from an ensemble of decision trees. Each root-to-parent-of-leaf decision path is mapped to a hidden neuron, while the matrices W1 (feature-to-branch) and W2 (branch-to-class) encode the structural relationships induced by the teacher ensemble. This provides an explicit correspondence between teacher branches and neural components, enabling feature-level and branch-level interpretation while retaining the flexibility of gradient-based optimization. The primary configuration studied in this work trains the teacher-derived feature-to-branch matrix W1 through gradient-based learning while keeping the branch-to-class matrix W2 fixed; this configuration is denoted SBN1,0. Across 29 multiclass tabular datasets, this configuration improves upon the teacher ensemble on 28 of 29 datasets, increasing mean accuracy from 84.0% to 87.9%. For reference, XGBoost and TabPFN achieve mean accuracies of 87.3% and 91.3%, respectively, on the same datasets. A one-sided Wilcoxon signed-rank test on dataset-level mean accuracies indicates a statistically significant improvement of SBN1,0 over the teacher ensemble (p=4.71x10-7). Beyond predictive performance, SBN provides an explicit structural basis for interpretation. Global feature importance can be derived from the teacher-induced feature-to-branch connectivity, while individual predictions can be traced to selected branch-neurons, their associated features, and their class contributions. An analysis across ten independently constructed teacher ensembles shows consistent predictions for a selected instance but only partial stability in its local feature rankings.

cs.LG · cs.AI

版本更新 · v2arXiv:2611.17048

Beyond Answering: Learning to Speak First and Act in Streaming Video

Shuhan Li, Huankang Guan, Xintong Guo, Yang Bo, Yuyang He, Dunyuan Xu, Jingyan Chen, Fang Liu 等 18 位作者

原文摘要

Most AI assistants respond only when prompted, remaining silent even when events in the video stream offer meaningful opportunities for conversation. To move beyond purely user-driven interaction, we present AURA 2.0, a streaming video large language model that introduces three core upgrades to AURA. The key advance is Model-Initiated Conversation (MIC), which enables the model to identify topics worth discussing in the evolving visual stream and initiate conversations at appropriate moments without a user query. Stream-Aware Tool Calling (SATC) lets the model act on what it sees by invoking external tools using the current visual and conversational context and incorporating returned information into the ongoing interaction. Interactive Visual Grounding (IVG) supports visual grounding in both user inputs and model outputs. The model can identify objects marked by users on video frames and produce frame-aligned bounding boxes to link its responses to specific objects and regions. To develop these capabilities, we construct a dedicated streaming interaction dataset and adopt a three-stage training scheme to integrate the new behaviors with AURA's existing streaming interaction abilities. We further incorporate Long-Horizon Interaction Memory (LHIM) to maintain contextual continuity over extended video streams. Evaluations on public streaming benchmarks show that AURA 2.0 achieves leading performance among open-source streaming models, while targeted evaluations further validate its new interaction capabilities.

cs.CV · cs.AI

版本更新 · v4arXiv:2611.05726

Do Language Models Know Which Constraints Belong Together?

Junbo Jacob Lian, Kaixuan Chen, Hanzhang Qin, Huiling Chen, Chaoyu Zhang, Chung-Piaw Teo

原文摘要

Dantzig--Wolfe reformulation strengthens the linear relaxation of a mixed-integer program by convexifying constraint families, and each candidate costs a column generation run. The choice is hard when families strengthen the bound only together, as in classical lot-sizing, network-design, and facility-location formulations. We prove that when an evaluation convexifies at most $s$ of $k$ families, methods that observe only bounds need of order $k^w$ evaluations to find a hidden group of $w$ families in the worst case. A prediction missing $e$ members reduces this to $Θ(k^e)$ through a staged search that need not know $e$. The orders extend to several groups, prices, and growing evaluation costs. We study two predictors: a lookahead at an optimal LP dual, which vanishes on groups whose families share no variable, and language models reading the model, verified by the staged search in JointDW. Such groups appear in three of ten MIPLIB-NL models whose bound can be improved and in all 59 benchmark instances. Given the coefficients of constructed models with up to 16 families, one language model names a true group first in 249 of 252 answers. On the benchmark instances, decompositions chosen by language models from the text solve faster than those of GCG's detector.

cs.LG · cs.AI · cs.DS · math.OC

版本更新 · v2arXiv:2607.12713

Code-Anchored Review Feedback Improves Correctness and Lowers Hosted Editing Cost on Large R Files

William Franz Lamberti

原文摘要

Tokens drive the cost and latency of generative AI (GAI) code editing, and correctness decides whether an edit is useful. We study how the format of a reviewer's feedback, a human-AI collaboration (HAIC) channel, impacts both. In a paired experiment on 36 synthetic base-R tasks with seven models, the same requested changes were delivered as a holistic list (control) or as the export of FileMark (treatment). That code-review extension anchors each comment to the code it concerns and quotes it. For the typical task, total tokens changed by -4.2% to +14.7% across models, while output tokens fell by 14.7% (Claude Opus) and 50.5\% (Claude Sonnet). Since an output token costs five times as much as an input token, the hosted cost of an edit on files over 100 lines fell by 20.0% (Opus) and 46.3% (Sonnet). The treatment raised the correct-run share of four of five local models by 24% to 46%. A decomposition experiment places these gains in the format, not the wording, and mostly in the quoted code. Misplaced anchors cut correctness sharply. Code-anchored feedback thus lowers hosted editing cost on large R files and makes weaker models more correct, provided the anchors are accurate.

cs.SE · cs.AI

新论文 · v1arXiv:2611.19714

An Interpretable Acoustic Coordinate System for Japanese Consonant Morae in Dysarthric Speech

Yuma Enomoto, Misa Yamaguchi, Hiroki Nishikawa, Ittetsu Taniguchi, Kazuhide Matsunaga, Narikazu Uzawa, Takao Onoye

原文摘要

Dysarthria reduces speech intelligibility, requiring self-directed rehabilitation based on fine-grained feedback to capture gradual articulatory changes. For Japanese speakers, such feedback is naturally provided in mora units due to its distinct moraic timing and consonant-vowel structure. While recent visualization methods have explored distance-based feedback, most are limited to vowels. Extending them to consonants is challenging, as generative embeddings often smooth transient spectral structures. We propose a visualization system for Japanese consonant morae that maps articulatory deviations onto a two-dimensional space defined by temporal and spectral distances from healthy reference speech. Experiments show consistent geometric separation of consonant categories, with Silhouette Coefficients exceeding 0.5 and origin-based ROC-AUC scores above 0.86. Furthermore, evaluations using real dysarthric speech achieved a high ROC-AUC of 0.897 in screening for unclear pronunciations. Perturbation tests confirm orthogonal temporal and spectral shifts, yielding a physically interpretable mapping for rehabilitation.

cs.SD · cs.AI

新论文 · v1arXiv:2611.19710

Joint Learning of Factual and Counterfactual Worlds for Causal Prior-data Fitted Networks

Furkan Danisman, Luke Zhang, Zhijing Jin, Dehan Kong

原文摘要

Estimating treatment effects is central to science across many disciplines and requires counterfactual reasoning about alternative worlds. In the real world, however, each unit reveals its outcome only under the factual world. Thus, the factual and counterfactual outcomes can never be jointly learned by any model trained on real data. The problem is therefore tractable only in synthetic settings, where both worlds can be generated. Recent progress on prior-data fitted networks (PFNs) has shown that models pre-trained purely on synthetic data can transfer to real-world prediction. This paradigm has since carried over to causal effect estimation. Despite this fully synthetic setup, existing causal PFNs carry the real-world constraint into their training objective, confining learning to a single world's outcome per unit. The dependence between the two worlds is thereby lost, and the density of the treatment effect relies by default on independence, which is implausible in practice. We therefore propose joint learning, a training setup in which a causal PFN learns the joint distribution of the factual and counterfactual outcomes of each unit. From the joint distribution, we obtain both the response and the individual treatment effect densities, and at inference, we apply optimal transport across units to obtain an approximation of the average treatment effect density. Ours is therefore the first causal PFN approach to provide all three densities together while accounting for the cross-world dependence. We apply joint learning and optimal transport to existing causal PFNs and evaluate it across semi-synthetic, synthetic, and real benchmarks spanning diverse causal scenarios. The jointly trained models significantly improve upon their baselines in point estimation and uncertainty quantification in most settings, indicating that joint learning is a preferable training setup for causal PFNs.

stat.ML · cs.AI · cs.LG

新论文 · v1arXiv:2611.19709

Cache or Denoise Less? Residual Reuse in a Flow-Matching World Action Model

Abhinav Gupta, Nissan Pow

原文摘要

Flow-matching robot policies generate each action chunk through repeated network evaluations, one per denoising step. Cross-cycle caches such as C$^3$ache reuse computation from earlier replanning calls, but whether they beat simply taking fewer steps is unclear. We compare the two in Fast-WAM, a 6B-parameter world action model, on LIBERO-Goal. Two uncached steps keep the success of ten: on all 500 initial states they are non-inferior within two percentage points, for one checkpoint and noise seed. Under PyTorch's default compile mode, they also cut a replanning call on an RTX A6000 from 202 to 92 ms. This leaves little room for caching. A cache that keeps two fresh steps cannot be faster than two uncached steps: every call still encodes the images, runs the video expert and computes those two steps. Caching the first of two steps, refreshed every fourth call, saves at most 11--14% of a call by our cost model. Cached policies are also fragile. Their losses tracked the share of the flow path the cached steps cover: under Fast-WAM's current time shift, C$^3$ache's eight cached steps cover 80% of the path instead of 44%, and the cached policy succeeds on 63% of states against 95% uncached. Resampling the noise on every call cost caches covering 60% or more of the path 14--40 points. These results argue for fewer-step baselines and for testing reuse under the deployed sampler and noise settings. We release our evaluation harness, logs and a checklist.

cs.RO · cs.AI · cs.CV · cs.LG

版本更新 · v3arXiv:2603.27977

SARL: Label-Free Reinforcement Learning by Rewarding Reasoning Topology

Yifan Wang, Bolian Li, David Cho, Ruqi Zhang, Fanping Sui, Ananth Grama

原文摘要

Reinforcement learning is critical to improving large reasoning models, but its success relies heavily on verifiable rewards (RLVR), making it hard to use in open-ended domains where correctness is ambiguous and cannot be verified. Moreover, reasoning trajectories remain largely unconstrained, and optimizing solely toward the final answer can favor early exploitation over generalization. In this work, we ask whether general reasoning ability can be improved by teaching models how to think (the structure of reasoning) rather than what to produce (the outcome of reasoning), and we extend traditional RLVR to open-ended settings. We introduce Structure-Aware Reinforcement Learning (SARL), a label-free framework that constructs per-response reasoning maps from intermediate thinking steps and rewards their reasoning topology. SARL shifts supervision from destination to path, encouraging reasoning trajectories that are both locally coherent and globally efficient. On verifiable math tasks, SARL outperforms prior label-free RL baselines and even exceeds RL methods with ground truth supervision, with average gains of +9.1% under PPO and +11.6% under GRPO across four math benchmarks, with particularly large improvements on AIME25 (+35.5% with PPO and +44.7% with GRPO). On non-verifiable open-ended tasks, SARL achieves average gains of +34.6% under PPO and +30.4% under GRPO on WildBench across five task categories, outperforming prior label-free RL methods and DPO, which relies on additional preference labels. Beyond strong performance, SARL exhibits substantially lower KL divergence and higher policy entropy, indicating more stable and exploratory training dynamics. Code and data are available at \href{https://github.com/cacayaya/SARL}{Code Link}.

cs.AI

新论文 · v1arXiv:2611.19700

Where Should LLM Agents Adapt at Test Time? In Context, in Weights, or Both

Parsa Mirtaheri, Mohammadreza Armandpour, Keivan Alizadeh, Iman Mirzadeh, Parshin Shojaee, Aryo Lotfi, David Harrison, Mehrdad Farajtabar

原文摘要

LLM agents operating repeatedly in the same environment, such as a code repository, a database, or an application suite, encounter tasks that share latent structure. Experience can reveal this structure, allowing agents to reuse environment knowledge and learn from environmental feedback. Agents can adapt at test-time using this experience through two channels: retaining it in context and updating their weights. These channels are usually studied separately, leaving unclear what each contributes and when they complement each other. We compare in-context learning (ICL), reinforcement learning with verifiable rewards using GRPO, and self-distillation using SDPO, alone and in combination, using a shared online protocol and modular framework. Agents encounter matched task streams, are evaluated on each task before adapting to it, and learn only from their own attempts and environment feedback. We evaluate 14 adaptation mechanisms and a stateless baseline with Qwen3.5-9B on 109 tasks from four environments drawn from SWE-bench Verified, AppWorld, ScienceWorld, and BIRD-Critic-SQLite, using five task orders per environment. ICL provides the largest standalone gain: its best policy raises mean success across environments from 28.8% to 49.4%. Adding GRPO to this policy achieves the highest mean success, 51.6%. While adding GRPO improves the overall mean under every ICL policy, SDPO has mixed effects. Trace analyses show fewer lookups and shorter attempts under ICL, consistent with reusing knowledge of the shared structure. However, ICL also frequently skips inspection and verification for the current task. Adding GRPO to ICL helps restore these behaviors while preserving its reduced environment lookups. These observations suggest that reinforcement learning with verifiable rewards can refine how agents use retained experience, motivating the joint study of context management and weight updates.

cs.LG · cs.AI · cs.CL

新论文 · v1arXiv:2611.19698

Data Attribution Should Steer Pretraining

Glenn Matlin, Chandreyi Chakraborty, Mohammed Rehan Parwani, Saehee Eom, Anthony Wen-Ming Zang, Akhil Theerthala, Arjun Chatterjee, Tejas Khandwekar 等 11 位作者

原文摘要

Pretraining data curation decides what a model learns from, yet what makes a document worth learning from depends on what the model already knows. Influence-driven selectors already rank documents to improve one run; we argue that training-data attribution should steer pretraining by identifying testable data properties that become reusable curation criteria, evaluated on new data against cheaper selectors at matched total cost, discovery included. A document ranking is consumed by the run it was computed for, whereas an explained data property can be tested across corpora and checkpoints, especially where familiar quality signals diverge from measured usefulness. Because historical contribution, an understandable explanation, and future data value require distinct evidence, we specify a feedback loop that interprets attributed sources, tests candidate data changes, and states for five claim types what evidence each requires and what none establishes. We close with a falsifiable comparison against perplexity, classifier, and mixing baselines at equal total compute.

cs.LG · cs.AI · cs.CL

新论文 · v1arXiv:2611.19692

MathaMAGICal: Validating Trajectory Data Attribution During Capability Acquisition

Jason Lee, Dikshanta Pokhrel, Phonn Nay Khant, Victor Lee, Moritz Weckbecker, Sebastian Lapuschkin, Wojciech Samek, Mark Riedl 等 9 位作者

原文摘要

Influence-function-based methods like MAGIC provide exact first-order scores for trajectory data attribution, yet only retraining indicates whether those scores accurately predict data removal. We evaluate MAGIC on $86$M- and $152$M-parameter GPT-2-style models learning multistep arithmetic from scratch on filtered web text and four synthetic math families. Full-run attribution at $86$M suffers from numerical overflow, and even with metagradient clipping, scores correlate near chance with retraining (linear datamodeling score, LDS, $0.17$--$0.22$). In contrast, short checkpoint windows yield stable, accurate predictions at both scales (LDS $0.86$--$0.99$ across four $100$-step windows at $152$M), whereas a single-checkpoint gradient-dot baseline shows near-zero correlation. Rankings of individual training blocks predict retraining only within that single run and do not transfer across seeds or hyperparameters ($ρ\le 0.09$), whereas format-level patterns replicate across runs. Two data families, worked examples and primitive facts, occupy $2.1$--$2.6\times$ their corpus share in both the helpful and harmful $5\%$ score tails, reflecting strong bidirectional local sensitivity distinct from net benefit under multi-epoch family deletion. These results show that trajectory attribution is reliable within short, validated windows and yields family-level insights that generalize across runs.

cs.LG · cs.AI · cs.CL

版本更新 · v3arXiv:2604.20744

AAC: Admissible-by-Architecture Differentiable Landmark Compression for ALT

An T. Le, Vien Ngo

原文摘要

We introduce the Architecturally Admissible Compressor (AAC), which adapts landmark-based A* heuristics to query demand under a fixed label-memory budget. AAC provides two continuous relaxations of ALT (A*, Landmarks, and Triangle inequality) selection. With exact landmark distances, every parameter value defines a lower bound on shortest-path distance. Both relaxations deploy ALT using only the selected labels. The row-stochastic relaxation takes convex combinations of landmark distances. We characterize the maximal admissibility-preserving linear classes and prove that the relaxed and discrete optima coincide. The fractional relaxation mixes subset heuristics and supports continuous greedy (AAC-CG) with exact gradients and an entropy-regularized oracle for differentiable selection. For expected heuristic value, we prove an approximation guarantee with finite-step, oracle-error, and finite-sample terms. The limiting $1-1/e$ factor is optimal for arbitrary supplied pools and query distributions unless P${=}$NP. At matched label memory, AAC-CG reduces A* expansions by $1.9$-$22.7$% relative to farthest-point ALT (FPS-ALT) in all $19$ tested combinations of graph and memory budget across six road networks and two synthetic graphs. Each combination uses one $500$-query test split; all methods share distance tables when distances are symmetric. Canonical greedy also expands fewer nodes than FPS-ALT in all $19$ graph-budget pairs. AAC-CG achieves $99.6$% of canonical greedy's mean heuristic value on test queries, averaged over these pairs. AAC makes admissibility independent of training success, while its analysis and experiments distinguish selection quality from search efficiency.

cs.AI · cs.LG · cs.RO

版本更新 · v2arXiv:2608.25480

DeCO: Discriminative Evidence Composition for Fine-Grained Dataset Distillation

Chuixuan Fan, Guang Li, Shijie Wang, Dongzhan Zhou, Baoli Sun, Takahiro Ogawa, Miki Haseyama, Zhihui Wang

原文摘要

Dataset distillation compresses a large training set into a compact synthetic set while preserving its downstream utility. However, existing methods primarily preserve global image statistics and may overlook the localized evidence essential for fine-grained visual classification (FGVC), such as object parts, subtle textures, and region-specific structures. We formulate fine-grained dataset distillation as budgeted discriminative-evidence preservation and propose Discriminative Evidence Composition (DeCO). DeCO uses attention rollout from a pretrained TransFG teacher to identify informative patches, applies spatial diversification to reduce redundant coverage, and organizes the resulting regions into class-wise evidence banks. Multiple same-class regions are then packed into compact grid-composed images. The teacher is used only for dataset construction, whereas downstream students are trained with standard hard-label supervision without teacher logits. Experiments on CUB-200-2011, FGVC-Aircraft, and Stanford Cars show that DeCO consistently outperforms representative coreset and dataset-distillation baselines under different IPC budgets.

cs.CV · cs.AI

新论文 · v1arXiv:2611.19670

Architecting the Corporate AI Infrastructure: The Evolving Roles of AI Governance Committees and Chief Ethics Officers

David Hason Rudd, Dennis Singh, Cesar Sanin, Bryn James

原文摘要

The accelerating integration of artificial intelligence (AI) into core business operations has created an urgent need for robust corporate governance infrastructures capable of managing complex ethical, legal, and operational risks. While organizations have increasingly adopted high-level AI ethics principles, a critical implementation gap persists: fewer than half of organizations that cite AI governance as a boardroom priority have a formalized governance framework in place. This paper addresses that gap by proposing a structured corporate AI governance architecture (CAIGA) centred on two mutually reinforcing pillars. First, it examines the evolving mandate of specialized C-suite roles, specifically Chief Privacy Officers (CPOs) and Chief Ethics Officers (ChEOs), in formalizing accountability for AI usage across the enterprise. Second, it argues that cross-functional AI Governance Committees, uniting representatives from legal, technology, data science, cybersecurity, and business units, are the essential organizational mechanism for translating ethical principles into operational decisions, including strategic go/no-go approvals for new AI initiatives. Drawing on a conceptual review of peer-reviewed governance literature, regulatory frameworks including the EU AI Act, and documented industry practice, this paper contributes a replicable governance architecture that embeds clear accountability policies, structured risk management workflows, and an organization-wide ethical AI culture. The framework is illustrated through application to the financial services sector. The findings have direct implications for corporate leaders, risk professionals, and policymakers navigating the rapidly evolving global AI regulatory landscape.

cs.CY · cs.AI

版本更新 · v2arXiv:2605.07194

Closed-Form Linear-Probe Dataset Distillation for Pre-trained Vision Models

Bincheng Peng, Guang Li, Ping Liu, Takahiro Ogawa, Miki Haseyama

原文摘要

Dataset distillation compresses a large training set into a small synthetic set that preserves downstream training utility. While most existing methods target training networks from scratch, modern visual transfer learning often uses frozen pre-trained encoders followed by lightweight linear probing. Existing distillation methods for this setting either unroll iterative linear-probe updates with trajectory-based gradient matching, or rely on closed-form formulations originally designed for from-scratch training with neural-tangent-kernel (NTK) approximations. Neither route exploits the fact that frozen-feature linear probing admits a closed-form solution determined directly by the pre-trained features themselves, with no infinite-width approximation and no inner-loop trajectory. We propose Closed-Form Linear-Probe Dataset Distillation (CLP-DD), a bilevel formulation that computes the linear probe induced by the synthetic set with a sample-space kernel ridge solver. The synthetic images are then updated by evaluating this induced classifier on real features through a temperature-scaled softmax cross-entropy, where the classifier columns act as learned class anchors in feature space. We further show that the choice of outer objective is decisive: pairing the closed-form inner solver with a standard MSE outer loss substantially underperforms trajectory-based methods, while the discriminative outer loss closes most of the gap. On ImageNet-100 with four pre-trained backbones, CLP-DD substantially improves over LGM without DSA and approaches LGM with DSA at a fraction of the computational cost. On ImageNet-1K, CLP-DD matches or surpasses LGM with DSA on three of four backbones while running roughly $14\times$ faster and using less than one-eighth of the GPU memory. The code is available at https://github.com/Guang000/CLP-DD.

cs.CV · cs.AI · cs.LG

新论文 · v1arXiv:2611.19663

When AI Agents Should Repair, Replan, or Abandon: A Threshold Theory of Plan Commitment

Gaurav Tewari

原文摘要

AI agents often discover mid-execution that their plan is broken, for example when an action fails or a tool is withdrawn. Deployed systems use conventions such as retry loops or "always replan," and a recent benchmark finds no fixed strategy dominates. This paper derives the optimal policy. In the model, a disruption destroys the causal support of part of a plan; repair is a partial reset that preserves the plan's structure, replanning is a full reset, and changing a plan is costly because others have conditioned on it. The optimal policy is a threshold in effective disruption depth: repair shallow damage, replan after deep damage, and abandon the goal when even a fresh plan is not worth its cost. Depth depends on plan structure. In expectation, a single broken action forces 1+R(G)/n actions to be revalidated, where R(G) counts the pairs of actions joined by a dependency path; to first order, a fully serial plan switches to replanning at raw damage about w times smaller than a plan of w parallel chains. A closed-form stability-price threshold determines when a worse repaired plan should be kept over a better fresh one. As language-model planners push replanning costs toward zero, the stability price becomes the only force sustaining plan commitment. Under repeated disruptions, revision frictions make commitment hysteresis optimal: the agent keeps executing a plan it would never choose fresh, and each repair adds patch debt until the debt outweighs repair's cost advantage and the agent rebuilds. A value-of-diagnosis result states when measuring damage is worth its compute and shows that "always replan" is optimal when damage is predictably deep and replanning is the better fallback. A separately implemented numerical checker tests each proposition (zero violations; 1.6 million static-model evaluations), and the model's regime map reproduces that benchmark's regime flips qualitatively.

cs.AI

版本更新 · v2arXiv:2605.07278

Predictive but Not Plannable: RC-aux for Latent World Models

Wenyuan Li, Guang Li, Keisuke Maeda, Takahiro Ogawa, Miki Haseyama

原文摘要

A latent world model may achieve accurate short-horizon prediction while still inducing a latent space that is poorly aligned with planning. A key issue is spatiotemporal mismatch: these models are often trained with local predictive supervision, but deployed for long-horizon goal-directed search in latent spaces where Euclidean distance may not reflect what is reachable within a finite action budget. We present the Reachability-Correction auxiliary objective (RC-aux), a lightweight correction for this mismatch in reconstruction-free latent world models. RC-aux keeps the world-model backbone unchanged and adds planning-aligned supervision along two axes. Along the time axis, multi-horizon open-loop prediction trains the model beyond one-step consistency. Along the space axis, budget-conditioned reachability supervision, together with temporal hard negatives, encourages the latent space to distinguish states that are eventually reachable from those reachable within the current planning horizon. At test time, the learned reachability signal can also be used by a reachability-aware planner to favor trajectories that are both goal-directed and attainable under the available budget. We instantiate RC-aux on LeWorldModel and evaluate it under both continuation-training and matched-from-scratch settings. Across goal-conditioned pixel-control tasks and a LIBERO-Goal extension, RC-aux improves LeWM-style planning with modest additional cost. These results suggest that planning with latent world models depends not only on predictive accuracy, but also on whether the learned representation encodes the temporal and geometric structure required by downstream search. The code is available at https://github.com/Guang000/RC-aux.

cs.LG · cs.AI · cs.CV

版本更新 · v2arXiv:2611.15108

Typed Knowledge over Constructive Institutions: A Proof-Relevant Type Theory for Verified Science

Hans-Martin Will

原文摘要

Mathematical proof assistants such as Coq, Agda, and Lean 4 build proof irrelevance into definitional conversion ($p \equiv q$ for $p, q : P$ with $P : Prop$), erasing which derivation established a theorem. While natural for mathematics over fixed axioms, judgmental proof irrelevance is incompatible with verified scientific stores, which must inspect and compare independent derivations of a claim to assess consilience and compute counterfactual retractions when an assay, reagent, or model is invalidated. Moreover, combining definitional proof irrelevance with subsingleton elimination breaks normalization and makes algorithmic conversion non-transitive. We present the type-theoretic and institutional foundations of Eigenius, a verified knowledge substrate for science. First, we formulate a dependent kernel with cumulative universes ($Prop, Set, Type_k$), impredicative $Prop$, inductive families with index-forced subsingleton elimination, and intensional identity types, where definitional equality $\equiv$ is untyped joinability under an orthogonal CRS and proof irrelevance is a witnessed axiom ($irrel : ΠP{:}Prop.\,Πp,q{:}P.\,Id\,P\,p\,q$). Second, we unify universe cumulativity, $Π$-covariant subtyping, and ontology class inclusions in a coercion-free subsumption judgment whose bidirectional fragments ($\mathcal{K}_0, \mathcal{K}_1$) satisfy subject reduction, compile to unification-free checkers ($\mathcal{A}_0, \mathcal{A}_1$), and match declarative typing. Third, we organize external verifiers into a three-level institutional hierarchy with explicit axiom policies (Adopt, Discharge, Relativize), internalize a justification logic with Why-provenance support $supp(j,P)$, and embed Cooper's Type Theory with Records (TTR) for open-world knowledge graphs. Finally, we prove consistency in Grothendieck universes and mechanize the metatheory in Lean 4 with zero sorry axioms.

cs.LO · cs.AI · cs.CL

版本更新 · v3arXiv:2512.18436

VeruSAGE-Bench: A Benchmark Suite for Rust System Verification

Chenyuan Yang, Natalie Neamtu, Chris Hawblitzel, Jacob R. Lorch, Shan Lu

原文摘要

Large language models (LLMs) have shown impressive capability to understand and develop code. However, their capability to rigorously reason about and formally prove the correctness of software systems remains in question. Current benchmarks for formal proof generation focus on isolated tasks, typically algorithmic-style problems utilizing basic data structures. We curate a new benchmark suite for system-level proof generation, VeruSAGE-Bench, which consists of 849 proof tasks extracted from eight open-source Verus-verified Rust systems, covering operating systems, storage systems, memory allocators, and more. These tasks are much more complicated than those in previous benchmarks, containing on average more than 20x the formal specification (in Lines of Code) and requiring a much broader array of verification techniques. We evaluate six LLMs on VeruSAGE-Bench (o4-mini, GPT-5, Sonnet 4, Sonnet 4.5, Opus 4.5, Qwen3.5), and demonstrate that different models and agentic setups lead to vastly different success rates and costs. We also show that the best models possess impressive capability in writing correctness proofs and hence should inspire new ways of developing trustworthy software. Meanwhile, some difficult tasks in VeruSAGE-Bench stress even the best model, forcing long reasoning time (over an hour on the hardest tasks, 15.3 min per task on average), high token cost (USD 13.65 per task on average), and proof failures, which we hope will guide future improvements in models and agents. This is particularly true for a harder no-lemma version of VeruSAGE-Bench, where the helper lemmas written by human experts are removed: there, the best model proves only 41% of the tasks from the hardest project, at USD 96 per task.

cs.OS · cs.AI · cs.FL · cs.SE

新论文 · v1arXiv:2611.19656

BadGCL: Towards Persistent Backdoor Attack on Graph Continual Learning Systems

Dong Li, Yufei Han, Xintao Wu, Xujiang Zhao, Wei Cheng, Chen Zhao

原文摘要

Backdoor attacks pose a persistent security threat to machine learning systems, yet their behavior under graph continual learning (GCL) remains underexplored. Conventional graph backdoor attacks often lose effectiveness as models are updated across successive tasks, creating the impression that continual updates naturally weaken malicious associations. This work reveals a contrasting vulnerability: knowledge-preservation mechanisms in GCL can also preserve early injected backdoors when the poisoned update is compatible with the clean-task signal being retained. For graph data, this risk is amplified by structural triggers, which are embedded into the message-passing paths of graph neural networks (GNNs) and can induce stable representation shifts across tasks. To expose this risk, we propose BadGCL, a model-agnostic framework for persistent backdoor attacks under a strict early-stage intervention setting. BadGCL learns topology-aware structural triggers using a gradient mutual-information objective that couples poisoned gradients with clean-task optimization signals, together with a propagation consistency objective that stabilizes trigger-induced effects through graph message passing. We provide theoretical analysis showing that gradient compatibility and propagation stability jointly support long-term backdoor persistence under benign continual updates. Extensive experiments across task-IL and class-IL protocols, four GCL benchmarks, and diverse victim systems show that BadGCL achieves persistent attack effectiveness with limited clean-utility degradation, substantially outperforming graph and continual-learning backdoor baselines and remaining robust against representative defenses.

cs.LG · cs.AI

版本更新 · v2arXiv:2608.13438

ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models

Gehan Zheng, Matthew Johnson-Roberson, Weiming Zhi

原文摘要

Contact-rich manipulation failures are often detected only after the robot has committed to contact. This is especially limiting in wrist-camera setups: close gripper--object views help observe contact, but a poor approach may already push, miss, slip, or disturb the object before conventional detectors react. We introduce ContactGuard, a pre-contact execution monitor for chunked visuomotor policies. Given the policy's planned action chunk, ContactGuard predicts its short-horizon consequence in latent visual space and aborts if the predicted future latent indicates likely failure. Its latent world model is trained from unlabelled robot trajectories to predict compact multi-view visual embeddings under planned actions, avoiding pixel-level video prediction. A lightweight failure probe is then trained from a small labelled set of pre-contact clips. At deployment, ContactGuard anchors prediction before an imminent contact event, rolls the model forward under the policy's own actions, and verifies the predicted post-contact latent. Across real-world contact-rich manipulation tasks, ContactGuard consistently benefits from action-conditioned latent prediction over current-state and direct-prediction alternatives. Counterfactual action perturbations further show that its failure score responds to the specific action being proposed. We deploy the same frozen monitor online with the visuomotor policy, where it provides a pre-contact abort signal without modifying the underlying policy.

cs.RO · cs.AI · cs.CV

新论文 · v1arXiv:2611.19649

LoopFill: Pipelined Multi-Token Speculative Decoding for Looped Transformers

Siqi Zhu, Suozhi Huang, Yi Zhao, Jiaxuan You

原文摘要

Looped Transformers improve parameter efficiency by repeatedly applying the same Transformer layers. During autoregressive decoding, however, generating each token requires several sequential passes through these layers, leading to high decoding latency. Pipelined self-speculative methods use shallow-loop predictions to process tokens at different loops in the same batch, but each token must reach the drafting depth before the next can enter the pipeline. We introduce LoopFill, a training-free speculative decoding algorithm for looped LLM that combines multi-token suffix drafts with shallow-loop predictions in a single draft tree. Both drafters contribute to one tree, so a suffix lookup can immediately extend a shallow prediction and several future tokens can enter the pipeline together. The shared block computes the highest-scoring draft candidates at different loop depths in one batch, and full-depth verification preserves the model's autoregressive output distribution. Across seven benchmarks, LoopFill's average speedup over greedy autoregressive decoding is $5.85\times$ on Ouro-2.6B and $5.13\times$ on Huginn-3.5B. The speedups exceed those of LoopAhead, the strongest baseline, by \(40\%\) on Ouro-2.6B and \(25\%\) on Huginn-3.5B.

cs.LG · cs.AI

版本更新 · v4arXiv:2602.02734Google Research

WAXAL: A Large-Scale Multilingual African Language Speech Corpus

Abdoulaye Diack, Perry Nelson, MohamedElfatih MohamedKhair, Emmanuel Asiedu Brempong, Subhashini Venugopalan, Tavonga Siyavora, Vusumuzi Dube, Kwaku Agbesi 等 47 位作者

原文摘要

Recent advances in speech technology predominantly favor high-resource languages, leaving speakers of Sub-Saharan African languages underserved. To bridge this divide, we introduce WAXAL, an open, large-scale speech dataset spanning 27 African languages representing over 100 million speakers. WAXAL comprises over 14,100 hours of speech collected in partnership with African academic and community organizations to build local research capacity: an Automatic Speech Recognition (ASR) corpus with 2,340 hours of transcribed, image-prompted natural speech across 19 languages (plus ~11,760 untranscribed hours) and a studio-quality Text-to-Speech (TTS) corpus of ~247 hours across 15 language configurations. We empirically benchmark three architecturally distinct ASR models -- Gemma 3n 2B, Whisper-Large-v3, and MMS-1B-all -- across the 19 ASR languages on strictly speaker-disjoint splits. Fine-tuning on WAXAL yields substantial gains, reducing macro-average WER by up to 58% for Whisper-Large-v3, 56% for Gemma 3n 2B, and 17% for MMS-1B-all (and Character Error Rate by up to 70%). WAXAL-fine-tuned models generalize zero-shot to three out-of-distribution benchmarks: FLEURS, Mozilla Common Voice, and Sunbird SALT. Finally, an LLM-as-judge semantic evaluation confirms that fine-tuning preserves speaker intent beyond surface-level transcription metrics. WAXAL is released at https://huggingface.co/datasets/google/WaxalNLP.

eess.AS · cs.AI · cs.CL

新论文 · v1arXiv:2611.19643

SG-LegalClarify: A Large-Scale Benchmark for Legal Reasoning and Ambiguity Recognition in Singapore Law

Ziwei Gu, Michael S. Lee, Drew Rein, Yash Maurya, Vikneswaran Kumaran, Darshini Ramiah, Chenguang Wang, Udari Madhushani Sehwag

原文摘要

Singapore offers a distinctive testbed for legal reasoning by large language models (LLMs). Its legal system combines inherited English common law with jurisdiction-specific statutes, requiring models to distinguish local doctrine from generic common-law intuition. This setting also highlights the challenge of recognizing underspecification in lay legal queries, because the facts material to a legal outcome depend on the governing doctrine. Yet legal benchmarks rarely cover Singapore law or test this form of ambiguity recognition. To address these gaps, we introduce SG-LegalClarify, a large-scale benchmark for open-ended legal reasoning and ambiguity recognition in Singapore law. SG-LegalClarify contains 1,100 lawyer-authored source questions spanning 10 legal domains, including 340 fully specified and 760 underspecified questions. Each underspecified question has a fixed hidden-fact assignment and a fully specified counterpart, yielding 1,860 evaluation instances. Responses are graded with lawyer-authored Issue, Rule, Application, and Conclusion (IRAC) rubrics. We evaluate 19 LLMs, closed-book and with web search, using a multi-turn pipeline in which models decide whether to answer or clarify, and a user simulator reveals hidden facts only when asked. We find the benchmark far from saturated: the best models satisfy only 69.8% of mandatory rubric items closed-book (average 53.8%) and 77.8% with web search (average 63.0%). Models also struggle to recognize when clarification is necessary. We will release the public split of our benchmark, the model outputs on it, and the evaluation pipeline to support future research.

cs.CL · cs.AI · cs.LG

新论文 · v1arXiv:2611.19636

ROUNDS: Source-Framed Diagnostic Revision after Multi-Turn Clinical Evidence

Jinrui Fang, Runhan Chen, Yining Wu, Jiawei Xu, Xu Yang, Jian Yu, Wenqi Shi, Ying Ding 等 9 位作者

原文摘要

Large language models (LLMs) achieve high scores on medical benchmarks, but does this performance translate into reliable diagnosis as a conversation unfolds? In diagnostic conversations, users can provide clinical findings across multiple turns, ask for a diagnosis without supplying answer options, and introduce opinions from others. Models must therefore form diagnoses from the available evidence and decide how to respond to subsequent advice. We introduce ROUNDS, a framework that evaluates these abilities after a fixed sequence of clinical evidence delivered over multiple turns. ROUNDS compares diagnosis with and without answer options, presents the same incorrect diagnosis under clinician and lay framing, and varies support for competing diagnoses in written five-clinician panels. Across seven models and 1,035 cases from five datasets, average accuracy was 76.9% with answer options and 57.0% without them. Clinician-framed misleading advice produced lower final accuracy than lay-framed advice in every model, with differences of 4.6 to 38.2 percentage points. Initially correct diagnoses more often became incorrect answers than abstentions, and many adopted the suggested disease. Conversely, accuracy increased as more panel members supported the correct diagnosis, including when the majority opinion remained unchanged. These findings reveal a gap between medical benchmark performance and reliable diagnosis in conversation: models respond to external opinions, but remain vulnerable to abandoning correct diagnoses without new clinical evidence.

cs.CL · cs.AI

新论文 · v1arXiv:2611.19630

Where slide-level foundation-model classifiers fail on atypical breast lesions

Nerea Hernandez, Luis J. Herrera, Cesar Ramirez, Olga Valenzuela, Ignacio Rojas, Francisco M. Ortuño

原文摘要

Whether a ductal breast lesion is called atypical hyperplasia or low-grade carcinoma in situ increasingly determines surgery. However, slide-level classifiers built on pathology foundation models (PFMs) perform worst on these lesions. Using four PFMs, cross-validation on BRACS and preregistered hypotheses, we studied where the classification of atypia fails. Because slides are labelled with their most severe lesion, only 39% of the slides with annotated atypical ductal hyperplasia (ADH) were labelled ADH. Attention ranked the lesion high but rarely first, and showing the model only the lesion helped little, whereas classifiers trained per decision with lesion-level labels improved all four atypical boundaries. ADH regions inserted into benign slides raised the probability of malignancy more than that of atypia. In a hospital with pathologist-labelled single-subtype slides and in HISTAI, benign slides were scored closer to lesions and triage specificity fell from 0.48-0.53 to 0.00-0.07 and 0.22.

cs.CV · cs.AI

版本更新 · v2arXiv:2603.25351

Image Rotation Angle Estimation: Five Circular-Aware Methods Across Sixteen Architectures

Maximilian Woehrer

原文摘要

Automatic image rotation estimation is a key preprocessing step in many vision pipelines, yet challenging because angles have circular topology, creating boundary discontinuities that hinder standard regression. We present a comprehensive study of five circular-aware methods for global orientation estimation: three regression based (direct angle with circular loss, unit vector, phase-shifting coder) and two classification based (angular binning, circular Gaussian distribution). Using transfer learning from ImageNet-pretrained models, we evaluate them across sixteen modern architectures and two data regimes of natural images, small-scale DRC-D and large-scale COCO. Classification approaches are the stronger family, attaining the lowest error on thirteen of the sixteen DRC-D backbones. The decisive ingredient is supervision that encodes angular neighborhood structure: plain cross-entropy lacks it and fails intermittently, while circular smooth labels (CSL) eliminate every observed training failure and reach the lowest DRC-D error, 0.81° mean absolute error (MAE). On COCO, the common rotate-and-crop protocol leaks the applied angle through the rotated JPEG block grid, a shortcut that the classification heads learned and the regression heads largely did not. Re-encoding every training view after rotation suppresses it. Under this trace-free protocol and on re-encoded test views, classification again leads on all three backbones tested, and the circular Gaussian distribution (CGD) with ConvNeXt V2 Base reaches 3.76° on COCO 2017 (three-run mean), against 6.09° for the strongest prior model on the same views. CSL and CGD are also the cheapest methods to train.

cs.CV · cs.AI · eess.IV

新论文 · v1arXiv:2611.19625

Robust Well-Log Reconstruction Across Point Gaps and Whole Wells via Boundary-Gated Soft-Thresholding Kalman Fusion

Yongan Zhang, Zhihan Qiu, Yizheng Li, Youzhuang Sun

原文摘要

Missing wireline logs hinder subsurface characterization. Deep reconstruction networks suffer from gap-boundary steps and cross-basin shifts, while interpolation and Gaussian smoothers struggle with long intervals and outliers. We propose boundary-gated soft-thresholding Kalman fusion (ST-KF), unifying multiresolution wavelet shrinkage, distance-gated residual prior learning, and closed-form robust Kalman smoothing with learning--inference decoupling. Across two public benchmarks totaling 263 wells in the Norwegian North Sea and New Zealand's Taranaki Basin, ST-KF ranks first over 0.15--100-m anchored gaps with 0.09% parameter overhead. Against the strongest of nine baselines, pooled RMSE falls by 5.0%--8.3% for bulk density and 6.4%--15.6% for compressional slowness; against the four strongest baselines, reductions are 0.7%--4.1% for neutron porosity and 1.3%--2.9% for photoelectric factor. Ablations and sensitivity analyses confirm operator synergy and stable performance: removing state-space fusion increases short-gap errors by up to 58.8%, while zero-shot transfer cuts anchored errors by 14%--68%. The results show that local borehole continuity transfers across basins whereas cross-property mappings shift with regional geology, supporting measurement-guided log reconstruction.

cs.LG · cs.AI

新论文 · v1arXiv:2611.19624

Well-Log Reconstruction with Pre-trained Geological Knowledge: A Knowledge-Conditioned Large-Language-Model Generative Agent

Yongan Zhang, Yitao Yu, Yizheng Li, An Gong, Youzhuang Sun

原文摘要

Missing well-log intervals make subsurface reconstruction ill-posed. Existing numerical regression models lack geological prior knowledge to constrain ambiguous inverse mappings. We propose GeoLog-Agent, a knowledge-conditioned generative agent using pre-trained large language models (LLMs) as a source of geological priors. A universal Geo-Harness translates LLM stratigraphic reasoning into formation-anchored petrophysical trajectories, feature-wise linear modulation, and conditional adversarial constraints. On the Norwegian North Sea benchmark, GeoLog-Agent achieves a mean $R^2$ of 0.9307 and a peak $R^2$ of 0.9713. It exceeds the nine-baseline average $R^2$ of 0.8387 by 0.0920, with a 30.97% average RMSE reduction, and exceeds the strongest baseline by 0.0274--0.0498 in $R^2$, with 16.25%--24.05% lower RMSE. Ablations confirm the effectiveness and synergy of geological prior reasoning, multi-scale dilated generation, and conditional adversarial modules. Transferring the coupling mechanism to six conventional deep learning models improves their $R^2$ by 0.0162--0.0789 with cross-well stability. These results demonstrate how foundation-model domain reasoning can constrain deep generative reconstruction of missing petrophysical measurements.

cs.LG · cs.AI