LLM agents that call external tools must decide what to do when a tool call times out: retrying risks duplicate side effects, while not retrying can leave the task incomplete. We present RetryBench, an evaluation protocol for retry decisions and failure recovery in tool-using LLM agents. Each decision is scored on retry permission (is the retry justified by an idempotency-key guarantee or proof of non-execution?), settled effects (duplicates and omissions), and completion, by replaying it across all hidden execution worlds consistent with the observation. On four synthetic software-engineering tasks, we run two controlled prompt-sensitivity studies (1,728 decisions). Study 1 (idempotency-key descriptions): with the contract fixed, GPT-5.6 Luna reuses the valid idempotency key in 10/40 decisions when an extra key is described as a "Candidate recovery key" but in 39/40 when it is an "Alternative key string"; across three further phrasings per family the gap is 27/120 versus 113/120. GPT-5.6 Sol and Terra are at ceiling (40/40). Study 2 (action order): under underspecified lookup guarantees, retry rates are 17-27 points lower for GPT-5.6 Terra and GPT-5.4 when retry is listed first (option-order bias), mainly on two tasks; we make no claim about which behavior is better. Findings are limited to researcher-chosen phrasings, four tasks and unpinned model aliases. Code and all model responses are released.
Large language models (LLMs) on personal devices keep user data local, but limited memory restricts their general capability. Since most user-specific tasks do not require the breadth of frontier models, a practical solution is to adapt a shared on-device model into lightweight specialists. However, backpropagation is prohibitive in memory, and users rarely provide labeled data. We present MobileSpec, a system that trains specialists with forward passes alone and learns from automatically checked task outcomes, as in reinforcement learning with verifiable rewards (RLVR). Because forward-only updates are built from random perturbations, training is efficient only when these perturbations resemble the weight changes of RLVR, which largely preserve the singular values of pretrained weights and rotate their singular vectors. The core algorithm of MobileSpec, FuRA-S, produces such perturbations while training only the block-spectral coordinates of pretrained weights, namely the singular values of their column blocks. As we show theoretically and empirically, rescaling these blocks independently shifts their relative contributions to each weight matrix and thereby rotates its singular vectors. Since this update can be rewritten as a transformation of the layer input, FuRA-S also adapts INT4 models without modifying their weights. With specialists of only 1-5 MB that update 0.015-0.05% of the parameters, MobileSpec raises the success rate of a phone-use agent from 6-23% up to 89-100% and trains a personalized 8B model on an iPhone 16 Pro overnight. On RLVR benchmarks, FuRA-S is competitive with full-parameter evolution strategies (ES) and outperforms ES over LoRA adapters and quantized ES on INT4 models.
Mobile phones support many everyday agent tasks, but current mobile agents bundle capabilities into monolithic pipelines with no open recombination point. Users and third parties cannot publish, install, and compose new behaviors, leaving personalized daily-life requests underserved. We argue that the missing piece is a substrate: although phones are the Web's dominant access point, open publication and composition under decentralized authorship stop at the agent harness. We present FreeForm Harness (FFH), a mobile-native agent harness for open, composable extensions. A frozen base and a permanence-aware required layer handle mobile constraints, including location, proactive triggering, and scarce channels. A uniform plugin contract makes independently authored extensions composable, checkable, and replaceable over a frozen socket set. FFH also supports end-user authorship through a tiered commissioned-authoring path. When no existing extension matches a user's intent, a Web coding agent generates a candidate artifact, which is validated and admitted on device under the same contract as other extensions. An Android prototype spans device, market, and cloud. Evaluation shows that architectural boundaries are enforced at runtime; realistic daily-life requests fit the declared composition space without widening it; and user intent reaches on-device execution through retrieval or commissioned authoring, with final trust decisions remaining device-local. FFH advances the agentic Web from "everything is a plugin" toward "everyone is a developer" on mobile.
Modern learning systems increasingly combine pretrained models with downstream heads. Despite the growing importance of such modular designs, how the downstream model co-adapts to the geometry of pretrained representations remains relatively underexplored. We study this problem through vision-language-action models (VLAs), where a pretrained vision-language model (VLM) conditions an action model for control. We find that the action model progressively concentrates its sensitivity on high-variance dimensions of the VLM outputs. Through controlled experiments across different backbone-head models, we further show that counteracting this sensitivity concentration can be beneficial under representation--task mismatch settings. Motivated by these observations, we introduce \emph{Adaptive Interface Equalizer} (AdaIE), a parameter-free mechanism that reweights pretrained representations according to the evolving downstream sensitivity during training. AdaIE improves VLA success rates on simulation benchmarks, with gains of up to $+5.28$ percentage points (pp) on RoboCasa-GR1 and $+1.68$ pp on RoboTwin 2.0, and achieves an average gain of $+19$ pp across five real-world manipulation tasks. It further generalizes to diffusion policies and world models, improving downstream performance in both settings. Our results highlight interface co-adaptation as an important optimization problem when coupling pretrained representations with downstream models.
Koutian Wu, You Liu, Yue Liu, Zhiping Li, Ergan Shang
Scientific knowledge gains new uses when ideas, methods, and results move between research contexts. Understanding this process requires a model of both specialized knowledge production and the intermediary activities that connect it. We introduce Scientific Liquidity Agents (SLA), an agent-based model of knowledge arbitrage in scientific communities. SLA represents domain specialists and epistemic arbitrageurs within a shared environment of research problems, researcher interactions, and scientific artifacts. Individual decisions connect to collective state through a belief-desire-intention architecture, a prerequisite-based problem graph, a dynamic interaction network, and a persistent publication registry. The framework makes researcher roles and knowledge-transfer frictions explicit, providing a process-oriented account of scientific production and exchange. Two exploratory simulations with 10 and 18 agents produce 68 and 274 registered papers. In the smaller run, arbitrageurs take part in 16 of the 26 contact pairs that join this community into one connected component, while topical similarity alone splits it into seven. The same run yields three similarity edges, 26 contact pairs, and two coauthorship pairs, so measured community structure depends on what counts as a link. Paper production is abundant, but no recorded replication succeeds, which places the bottleneck in verification rather than production. Solved-problem sets correspond to the prerequisite reachability of the disciplines represented in each community. SLA makes knowledge liquidity, the movement and reuse of artifacts across domains, a quantity that agentic science can simulate and measure.
Ke An, Shizhe Xiang, Yue Liu, Liang Chen, Pei Fu, Shaojie Zhang, Xiaoxing Hu, Yupei Wang
Reinforcement learning with verifiable rewards (RLVR) has become an effective post-training paradigm for vision-language models (VLMs). Standard RLVR, however, assigns identical outcome-derived credit to all generated tokens regardless of their perceptual relevance. Recent perception-aware methods estimate visual reliance through image perturbation or masking. These indirect proxies require perturbation branches, while corruption sensitivity does not necessarily identify the visual evidence used for reasoning. We instead shape model perception through native attention, an intrinsic model signal. Rather than treating attention as a post-hoc explanation, we formulate it as a reward-anchored, token-aligned routing signal that encourages attention patterns supporting visually grounded reasoning. To this end, we propose NAS-PO, a Native Attention-Strategy Policy Optimization framework. NAS-PO analyzes native attention with a visual mass policy that quantifies visual allocation and a conditional visual policy that captures spatial routing over visual tokens, making the optimization well-posed. Based on this, Native Attention-Strategy Advantage Shaping (NAS-AS) draws on the visual-mass policy to redistribute signed policy-gradient credit for visually engaged rollouts. The Adaptive Attention-Coupled (AAC) loss then aligns the visual-mass policy on all response tokens and the conditional visual policy on visually active ones, consolidating the visual-routing patterns that drive successful trajectories. Experiments demonstrate that NAS-PO substantially surpasses leading RL-tuned VLMs across 7 out-of-domain reasoning and perception benchmarks, with its effectiveness consistently validated across 8B and 32B. Visual-Routing Sensitivity (VRS) analysis further indicates that the optimized attention prioritizes visually load-bearing regions.
Ejafa Bassam, Konstantin Garbers, Yingsheng Geng, Dalton Jens, Jiaxu Qian, Guyue Liu
Data augmentation improves visual recognition by exposing models to synthetic training examples that encourage generalization beyond the observed data. Mixed-sample methods extend this idea by combining multiple images. However, many common approaches either mix only pairs of images or rely on fixed small layouts, limiting control over source count, visible area proportions, and induced supervision. We propose TreemapMix, a distribution-controlled multi-image augmentation that samples source weights from a Dirichlet distribution and assigns them to area-proportional regions using a treemap partition. The resulting known region areas support two complementary objectives. TreemapMix with soft cross-entropy (SCE) treats areas as soft class-probability targets, while TreemapMix with a Plackett-Luce (PL) objective converts areas into ordinal supervision. On ImageNet-1K under matched training recipes, TreemapMix-SCE achieves substantially lower calibration error than evaluated mixed-sample baselines, while TreemapMix-PL achieves the strongest Top-1 accuracy. Transfer experiments on object detection and instance segmentation suggest that TreemapMix-pretrained features remain useful beyond classification. These results show that controlled multi-image composition can expose the same region-area information as either probability-matching or rank-matching supervision. Code is available at https://github.com/Ejafa/treemapmix.
As AI coding agents take on increasingly complex, long-horizon software engineering tasks, existing benchmarks are rapidly saturating and their evaluation quality has come under serious scrutiny: a recent audit found that nearly 60% of unsolved SWE-bench Verified instances contain flawed tests -- either overly narrow tests that reject correct solutions or overly broad tests that check unstated requirements -- and that frontier models can verbatim reproduce gold patches from training data. Code refactoring, which requires coordinated, behavior-preserving changes across many files, offers a substantially harder and more realistic test of agent capability, yet remains underserved by current benchmarks. We introduce SWE-Bench ProMax, an expert-curated, multilingual code refactoring benchmark of 170 instances drawn from real commits across seven programming languages (Python, Java, TypeScript, Go, C, C++, and Rust). Every instance undergoes rigorous, multi-stage curation that directly addresses the quality problems identified in prior benchmarks: issue descriptions are rewritten from scratch to provide precise, unambiguous specifications, and test suites are manually reviewed to remove overly narrow and overly broad tests. Tasks with insufficient complexity or limited cross-file scope are filtered out, yielding a benchmark of challenging, large-scale refactoring tasks that average 11.5 modified files and 261.6 lines of code per instance, substantially exceeding the scale of existing benchmarks. Experiments with frontier models under two agent scaffolds show that the best model achieves only 41.2% resolve rate, confirming that SWE-Bench ProMax presents a meaningful and unsaturated challenge for current AI coding agents. Our benchmark is available at https://huggingface.co/datasets/swe-bench-promax/SWE-Bench-ProMax.
Xinxun Zhang, Yuxiang Wang, Xiaokai Zhou, Jiayue Li, Yi Li, Yue Liu, Jun Zhang, Xiaoling Zhou 等 11 位作者
Click-through rate (CTR) prediction relies on historical behaviors and contextual features to estimate clicks on a target item. Existing Transformer-based CTR models often adopt index-based or temporal positional encodings and token-level causal interaction. However, we observe that most existing positional encodings do not outperform NoPE, while relaxing causal constraints within target-relative temporal groups also yields gains. These observations motivate a fundamental question: what positional encoding and interaction granularity does CTR prediction actually require? To address this question, we propose \textbf{TRIP}, which combines target-aware rank--time positional encoding with temporal-group interaction modeling. Specifically, Factorized Rank Encoding represents each behavior's ordinal position within the historical sequence through a low-rank factorization, while Target-Relative Time Encoding characterizes its temporal distance to the target on a coarse-to-fine scale. This joint design grounds each behavior's positional representation in its sequence context and temporal relation to the current target. Furthermore, Temporal Group-Causal Interaction enriches intra-group representations through bidirectional interaction and masked reconstruction while preserving inter-group causal order. Finally, behavior representations are condensed into compact group summaries for efficient group-level modeling in subsequent layers. Experiments on four public benchmarks show that TRIP improves AUC by up to 4.45\% relative to the strongest baseline.
Xihan Ma, Bo Zhang, Boxiong Li, Yuhan Bao, Ziyue Liu, Yucheng Wang
HoloBoxing presents a simple recipe for learning humanoid boxing by building on a pretrained whole-body controller. The recipe separates reusable motor skills from task decisions: a compact planner requests future body-point targets, while the frozen controller handles joint-level execution. A learned decoder maps 32-dimensional actions to structured target windows, and measured-state imitation initializes the planner before reinforcement learning. Task learning then refines interaction behavior while reusing the decoder and controller, avoiding the need to relearn whole-body control. This combination keeps task optimization focused on a compact upper policy and supplies an initial behavior before reward-driven exploration. Starting from that available controller, the recipe accounts for about seven minutes of additional training, including reused representation fitting and a single-GPU task run. This excludes controller pretraining, data preparation, evaluation, and search. Simulated boxing experiments and component controls evaluate the resulting behavior and the roles of imitation initialization and the target representation.
Zehui Xiao, Yi Zheng, Guixing Fu, Wenkai Lu, Lingfei He, Tao Tan, Yue Liu, Silun Wang 等 10 位作者
Cross-subject generalization remains a major challenge for EEG-based emotion recognition due to pronounced inter-subject variability in neural activity and emotion-related spatiotemporal patterns. Existing methods mainly alleviate inter subject differences through distribution alignment or contrastive learning, while paying limited attention to representation bias associated with fixed spatiotemporal modeling orders. We hypothesize that high-level emotion representations of the same EEG signal should remain consistent across different spatiotemporal modeling paths. Based on this hypothesis, we propose a Dual-Path Spatiotemporal Contrastive Learning (DP-STCL) framework for cross-subject EEG emotion recognition. DP-STCL employs two complementary encoders with independent parameters, namely a Temporal-to-Spatial (TS) path and a Spatial to-Temporal (ST) path, to model EEG spatiotemporal information under different integration orders. During representation learning, the same emotional stimulus is used to establish cross subject associations, while three complementary relations jointly constrain representation variations across subjects and modeling paths. This promotes stable and discriminative stimulus-related emotion representations. To further handle limited labeled target subject data, we introduce a LoRA-based few-shot adaptation strategy with low parameter overhead. Extensive cross-subject experiments on SEED, SEED-IV, SEED-V, and FACED show that DP-STCL consistently outperforms the compared methods, demonstrating its effectiveness in improving cross-subject EEG emotion recognition.
Hongcheng Gao, Zelin Zheng, Alex Zhou, Jay Zhu, Shijia Ge, Yazhe Wang, Xiping Li, Jianshu Zeng 等 21 位作者
Embodied agents are expected to accomplish real-world goals by deciding for themselves how and when to interact with their surroundings. For many such goals, changing the physical state of the world is the means rather than the end: the agent changes the state of a system through interaction, observes the feedback, and explores until it either achieves the goal itself (e.g., discovering physical laws) or learns what the goal requires; it acts to know. Manipulation benchmarks stop at acting to change, with the target state given; scientific discovery benchmarks evaluate knowing but replace physical experimentation with symbolic actions or numerical queries. We introduce EmbodiedLab, a benchmark in which a robot must act to know. Quantitative science experiments make this setting controllable: the parameters of the system are resampled for every instance, so the result cannot be known in advance, yet it can be scored exactly. Across 30 experiments in five disciplines on 14 apparatus platforms, the agent receives an objective and the apparatus, chooses its own procedure, and operates a physics-based simulated laboratory, reading the instruments through its cameras. Evaluation checks what the agent came to know, grounded in readings recorded during the episode, as well as the physical state the task requires; every experiment is verified to be solvable. We evaluate six frontier models in their developers' coding agents with native observation and control, with added perception and motion planning, and with manipulation skills distilled from the strongest agents. With native control, even the strongest agent succeeds less than half the time; for every agent, most failures arise in interacting with the apparatus, and manipulation skills help weaker agents more than perception and planning. The benchmark and its evaluation code are available at https://github.com/HexaFuture/EmbodiedLab.
Effective collaboration requires language models to understand what others know and believe, and to act accordingly. Models that assume others share their information may omit necessary communication and undermine coordination. Many theory-of-mind (ToM) benchmarks assess mental-state inference through story questions, but correct answers do not establish that models can use these judgments in subsequent tasks. We introduce a multi-level evaluation built on existing, verifiable tasks. We extend household collaboration and calendar scheduling with asymmetric information while retaining their original constraints and validators. Independent interfaces assess world understanding, information access, beliefs, decisions, and execution. Controls supply personal evidence or correct beliefs to distinguish state recovery from state use. Across six model configurations, these abilities diverge. On matched scheduling cases, Claude recovers all agent beliefs correctly but produces subjectively optimal plans in only 67.9% of cases. In household collaboration, supplying beliefs raises Qwen-27B success from 55.0% to 75.0%, yet Qwen-4B answers every fixed supplied-belief source question correctly while completing only 20.0% of the corresponding continuation condition. These results show why ToM evaluation must connect perspective inference to independently verified decisions and task consequences.
Knowing when they are reliable is essential for language models to support human-AI collaboration. Yet interaction feedback is not always complete or correct, and may mislead models' understanding of their own capabilities. We introduce LADDER to study how language models assess themselves under unreliable feedback. Our findings show that models can accurately report supplied performance information while still struggling to form trustworthy self-assessments from interaction experience. Apparently accurate assessments can obscure capability differences and do not necessarily predict future task performance. These findings suggest that trustworthy self-assessment should be evaluated against observed task performance and its predictive value. Our study provides empirical evidence of the limits of model self-knowledge and a basis for evaluating capability awareness in reliable human-AI collaboration.
Haoyue Liu, Zhichao Wang, Yongxin Guo ByteDance Inc., Haoran Shou, Xiaoying Tang The Chinese University of Hong Kong, Shenzhen
Automated prompt optimization is crucial for eliciting reliable reasoning from large language models (LLMs), yet most API-only prompt optimizers iteratively edit monolithic prompts, coupling components and obscuring credit assignment, limiting controllability, and wasting tokens. We propose Adaptive Prompt Structure Factorization (aPSF), an API-only framework (prompt-in/text-out; no access to model internals) that uses an Architect model to discover task-specific prompt structures as semantic factors. aPSF then performs interventional, single-factor updates: interventional factor-level scoring estimates each factor’s marginal contribution via validation-performance changes, and error-guided factor selection routes updates to the current dominant failure source for more sample-efficient optimization. Across multiple advanced reasoning benchmarks, aPSF outperforms strong baselines, improving accuracy by up to +4.29 percentage points on average, and reduces optimization cost by 45–87% tokens on MultiArith while reaching peak validation in 1 step.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Evaluating the grammatical abilities of large language models (LLMs) is important for both NLP and linguistic theory. We investigate the ability of large language models (LLMs) to perform acceptability judgments in a forced-choice paradigm. We evaluate a subset of LLMs on 150 minimal sentence pairs sampled from Linguistic Inquiry and categorized using BLiMP linguistic phenomena. Our results show that while LLMs approximate human judgments, performance varies across models and phenomenon types, with stronger alignment on morphosyntactic phenomena than on linguistically and semantically demanding phenomena. Prompting strategies have minimal impact.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop) · 2026 · 正式发表
Diffusion language models generate text by iteratively denoising all tokens in parallel, but when and where their hidden states encode whether the output will be functionally correct remains unknown.We present the first probing study of DLM internals, training linear classifiers on hidden states to predict functional correctness.Across two models (LLaDA-8B, Dream-7B) and four tasks, we find that DLMs uniquely accumulate correctness signal across denoising steps (AUC gains of 0.08–0.11 on reasoning tasks), absent in single-pass AR decoding. However, step-0 signal reflects prompt difficulty rather than diffusion-specific computation. Signal emergence is task-dependent: structural tasks show flat profiles while reasoning tasks show gradual buildup. The two models exhibit distinct layer dynamics, with LLaDA concentrating signal in upper layers while Dream redistributes toward lower layers. We further show that probe confidence can identify likely failures, enabling selective generation that avoids 36–98% of wasted compute.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop) · 2026 · 正式发表