AI 速览：综述机制可解释性研究，探讨“理解模型内部机制是否真能让AI更安全”，并提出统一评估框架与六类研究取向。
Neural networks are likely to be among the most consequential artefacts in human history. As these systems become more capable and autonomous, their potential to cause societal harm also grows, making AI safety one of the defining challenges of our time. Yet we continue to build increasingly powerful systems without fully understanding how they work, how they will behave in unfamiliar situations, or how to reliably control them. Mechanistic interpretability offers a path towards understanding these systems by studying their internal representations and computations, with the aim of evaluating, predicting, and controlling behaviour. We survey mechanistic interpretability for AI safety in order to answer the question: Does understanding how models work actually make them safer? We place mechanistic interpretability in the broader context of interpretability research and compare its mechanistic evidence with behavioural evaluation. We introduce the foundations of mechanistic interpretability, including features, polysemanticity, transformer computation, and circuits, and survey methods for feature extraction and analysis, circuit discovery and causal verification, explaining intermediate representations in natural language, learning dynamics, and training-data attribution. Beyond surveying existing methods, we propose a unified framework for evaluating mechanistic interpretability claims and characterise six research orientations by their goals, assumptions, and trade-offs. Finally, we examine how mechanistic understanding can support AI safety, including knowledge editing, steering, and monitoring, and assess its limitations, failure modes, open problems, and relationship to behavioural safety measures. We also release a repository of code notebooks with worked examples that reproduce key findings from surveyed methods, allowing readers to build intuition through hands-on experimentation.
Zhuangsheng Li, Liang Dong, Yu Wu, Dailin Li, Yu Xia, Yunhan C. Zhang, Rui Cai, Yi Wu 等 9 位作者
Collaborative-filtering (CF) scores provide graded, personalized rewards for training large language model (LLM) recommenders through reinforcement learning. However, these rewards can cause exposure collapse: even though the CF model itself produces diverse top recommendations across users, an LLM optimized with its rewards concentrates on fewer items. We decompose each score into an item-mean component, the item's average CF score across user histories, and a residual component, the current user's score deviation from the average. Through controlled interventions, we identify the item-mean component as a contributor to exposure collapse. Accordingly, we propose two reward corrections: per-item centering removes each item's estimated mean score, while per-item standardization further rescales the residual using its estimated standard deviation. Experiments on two datasets with two identifier schemes show that both corrections substantially improve catalog coverage and exposure evenness. Compared with training on the uncorrected CF reward, our corrections increase the number of distinct items ranked first across users by up to $9.8\times$, while maintaining comparable ranking accuracy. Code is available at https://anonymous.4open.science/r/LLM4Rec-CF-Bias-27EF/.
Shunyu Wu, Dan Li, Haozheng Ye, Weibin Feng, Jian Lou, Bo Zhang, Wenjie Feng, Chenjuan Guo 等 9 位作者
Assessing the quality of time series (TS) data is fundamental yet inherently challenging due to the multifaceted nature of quality dimensions. Recently, large language models (LLMs) have emerged as a promising paradigm for TS quality assessment via pairwise comparison and per-dimension evaluation. However, existing approaches rely on manually predefined quality dimensions and purely text-based reasoning, leaving it unknown whether LLMs can identify truly relevant quality dimensions or perform grounded and quantitative quality comparisons. To investigate this, we construct TSQBench, a dedicated benchmark for evaluating LLMs on two progressive capabilities: (i) understanding and identifying relevant quality dimensions, and (ii) performing quality comparison under specific dimensions. Our analysis reveals that current LLMs consistently struggle with both dimension identification and evidence-grounded quality comparison. To address these limitations, we propose TSQAgent, a novel agentic reasoning framework for TS quality rating consisting of three collaborative roles: Perceiver for focused dimension selection, Inspector for dimension-wise quantitative analysis, and Adjudicator that aggregates and refines the final judgment. In particular, we introduce an agentic reasoning strategy that instills the ability to identify and prioritize the most relevant quality dimensions, and further propose an agent workflow equipped with external analytical tools to enable precise quantitative comparisons over selected dimensions. Experiments on both the proposed benchmark and eleven real-world datasets demonstrate that our framework not only substantially improves LLMs' capabilities in quality understanding and quantitative comparison but also effectively translates these improvements into better quality-aware data selection, leading to enhanced downstream performance and data efficiency.
As Large Language Models (LLMs) move from curated training sets into open-ended real-world environments, a fundamental limitation emerges: static training cannot keep pace with continual deployment environment change. Scaling training-time and inference-time compute improves static capability but does not close this train-deploy gap. We argue that addressing this limitation requires a new scaling axis-evolution. Existing deployment-time adaptation methods, whether parametric fine-tuning or heuristic memory accumulation, lack the strategic agency needed to diagnose failures and produce durable improvements. Our position is that agentic evolution represents the inevitable future of LLM adaptation, elevating evolution itself from a fixed pipeline to an autonomous evolver agent. We instantiate this vision in a general framework, A-Evolve, which treats deployment-time improvement as a deliberate, goal-directed optimization process over persistent system state. We further propose the evolution-scaling hypothesis: the capacity for adaptation scales with the compute allocated to evolution, positioning agentic evolution as a scalable path toward sustained, open-ended adaptation in the real world. Our code is publicly available at https://github.com/A-EVO-Lab/a-evolve.
Daksh Mittal, Tommaso Castellani, Thomson Yen, Naimeng Ye, Minghui Chen, Tiffany Cai, William Zeng, Emmanouil Koukoumidis 等 10 位作者
We envision continually learning agentic systems that become more useful over time: as they encounter sequences of related tasks, they should infer the hidden structure shared across those tasks and use it to improve future decisions. This cross-task experiential learning capability is pivotal in domains such as personalization and interactive assistance, but existing training/evaluation frameworks do not provide shared, controllable latent structures and cannot measure whether or why agents improve. We introduce LatentGym: a controllable suite in which each environment is organized around a ground-truth latent variable governing the structure across tasks. Our construction yields metrics that separate exploration (whether the agent's actions gather information about the latent) from exploitation (whether the agent uses what it has gathered). We demonstrate our suite on empirical studies addressing three questions: how and why frontier models fail to adapt across related tasks; whether post-training on related task sequences improves general cross-task adaptation, and where those gains come from; and how design choices such as inter-task feedback shape training dynamics and generalization. Together, these results establish a controlled foundation for studying how LLM agents learn from experience across tasks, and for designing agents that adapt more reliably in sequential, personalized, and interactive settings.
LLM agents are increasingly deployed as systems built around editable external harnesses, including prompts, skills, memories and tools, that shape task execution without changing model parameters. Harness self-evolution adapts such agents by updating these harnesses from execution evidence. Yet it remains unclear whether a model's base capability in task-solving predicts its capabilities in harness self-evolution: which models produce useful harness updates, and which actually benefit from them? We analyze two harness self-evolution capabilities: (i) harness-updating, the capability to produce useful persistent harness updates from execution evidence; (ii) harness-benefit, the capability to benefit from updated harnesses during task solving. Our analysis reveals two findings. First, harness-updating is flat in base capability: models from different capability tiers produce harness updates that lead to surprisingly similar gains; even Qwen3.5-9B's updates yield gains comparable to those of Claude Opus~4.6. Second, harness-benefit is non-monotonic in base capability: weak-tier models benefit little from updated harnesses, mid-tier models benefit most, and strong-tier models benefit less than mid-tier. We trace low gains at the weak tier to two failure modes: weak-tier models may fail to activate relevant harness artifacts, or activate them but fail to follow them faithfully. These findings suggest investing capability budget in the task-solving agent rather than the evolver, and targeting harness invocation and long-horizon instruction following in agent training. Our source code is publicly available at https://github.com/A-EVO-Lab/a-evolve/tree/release/harness-evolution.
Recent Large Audio-Language Models (LALMs) have demonstrated promising abilities in understanding musical content. However, whether their responses are grounded in the correct temporal regions of the audio remains underexplored. This limitation is particularly critical for music understanding, where key information often occurs as temporally localized events, such as instrument entries and rhythmic transitions. To address this gap, we introduce MusTBench, a reliable benchmark with stem-wise data and music-expert validation designed to evaluate temporal grounding in LALMs through five temporally grounded question-answering tasks. To further improve temporal grounding in existing models, we propose MusT, a novel four-stage temporal optimization recipe spanning music encoder adaptation, LLM adaptation, LLM supervised fine-tuning, and RL-based optimization. Experiments on MusTBench show that existing LALMs face challenges in precise temporal grounding, while MusT consistently improves performance across different LALM backbones and scales. These results highlight temporal grounding as an important challenge in current LALMs and provide a benchmark and training recipe that may support future research on temporally grounded music understanding.
Computer-using agents are increasingly capable of performing complex tasks in the digital world. To advance such agents, supervised learning on annotated interaction trajectories has become a common approach to improving their task-solving capabilities. However, high-quality interaction trajectories remain scarce for specialized desktop applications, where limited application knowledge, complex GUI workflows, and incomplete exploration hinder automated collection. To address these challenges, we introduce \aks{}, a knowledge-guided automated trajectory collection framework. It combines tutorial-guided exploration with a hybrid GUI--Code action space, providing application-specific knowledge and simplifying GUI workflows through direct execution of application-native commands. The framework also checkpoints and revisits pivotal states during tree exploration to explore alternative workflows, complementing linear rollouts for broader coverage. We collect 10.9K automatically validated trajectories and use them to fine-tune open-source base models. The resulting model achieves 64.1\% accuracy on a benchmark of specialized desktop workflows, improving performance by 58.2 percentage points, establishing a new state of the art. The fine-tuned model also improves performance on general desktop workflows using GUI actions only, suggesting that the collected supervision benefits general desktop tasks as well. Comparative experiments and ablations demonstrate the efficiency of our collection framework and the value of code actions in specialized workflows.
Olfaction contributes substantially to human perception, yet the olfactory-verbal gap often leaves people struggling to articulate what they smell. We present Smell with Genji, an AI-mediated interactive experience adapting Genji-kō, a traditional Japanese scent-matching game. This practice maps comparative judgments across a sequence of scents onto Genji-mon, providing visual patterns as shared reference points for communication. Within this framework, we introduce an AI co-smelling partner integrating multichannel olfactory sensing, a Transformer-based classification model, and a large language model (LLM)-driven conversational interface to translate sensor signals into descriptive dialogue. Rather than deploying machine olfaction purely for identification, the system positions the AI as a co-present partner sharing its sensory impressions while human judgment remains open. Placing the machine's perspective alongside human intuition creates a space for shared reflection, encouraging participants to deliberately evaluate and articulate their choices. This work explores how computational systems can support sensory interaction grounded in dialogue, fostering reflective articulation and preserving human agency in making sense of the olfactory world.
Speech large language models retain audio history to support multi-turn conversations. Yet remembering an event does not ensure that its information is used only when relevant. We investigate why irrelevant historical audio changes answers even when the queried event remains represented. Controlled experiments on MOSS-Audio and MiMo-Audio show that historical values direct errors toward specific competing answers, beyond a generic effect of context length. Causal interventions distinguish two paths. In open-ended question answering, most history-value errors disappear when subsequent computation cannot directly access the irrelevant source. In multiple-choice answering, many errors persist: early historical influence is carried by question-candidate states even after direct source access is blocked. Low-dimensional interventions further separate source-role information from answer-value information. Their effects peak at different depths, and changing the source-role signal does not eliminate the causal influence of competing values. On naturally occurring errors, strengthening the queried role and counteracting the historical value each improve answers; their joint intervention gives the greatest recovery. These findings support a failure of downstream evidence use beyond source identification alone: source-role information can influence a decision without exclusively controlling which historical values affect it. Reliable conversational memory therefore requires not only retention and retrieval, but also control over how retained information enters answer computation.
He Wang, Yeye Liu, Ruijie Li, Xuemei Jiang, Junyu Wu, Lining Liu Zhuoqi Man, Tao Li, Yanjie Song
Agile Earth observation satellites must coordinate imaging tasks within limited observation windows, while accounting for attitude maneuvering and onboard energy constraints. Competition among observation opportunities makes scheduling performance sensitive to the heuristic rules used for task selection. This letter proposes a large language model (LLM)-assisted adaptive heuristic search method for agile Earth observation satellite scheduling. The LLM generates an initial heuristic rule according to the characteristics of the scheduling instance. During the subsequent ant colony search, search performance and scheduling information are fed back to the LLM to refine the heuristic rule. The proposed modifications are evaluated through paired solution constructions before being incorporated into the search process. A unified scheduling procedure determines task selection, satellite assignment, and observation timing under the operational constraints. Experiments on six scheduling scenarios involving 50--200 imaging tasks demonstrate that the proposed method achieves 4.3\%--9.6\% higher mean observation benefit than the strongest baseline.
Detailed image captioning is usually framed as an amount problem: a model should mention more visual facts while avoiding hallucination. However, autoregressive VLMs generate captions as ordered sequences, and later generations may be produced under weaker or noisier visual grounding. We therefore argue that long-form captioning should be studied not only by what facts are mentioned, but also by where they appear. We introduce a position-aware rubric representation that decomposes dense human references into atomic visual claims, verifies generated captions against those claims, and induces a quartile-wise composition of supported, hallucinated, and generic content. This lens reveals two qualitatively different failures in long captions: models tend to produce unsupported specificity in the middle, while the final portion often escapes into vague, non-evidential language rather than simply accumulating more hallucination. A same-reference human control does not exhibit these patterns; additional human audits examine the visual validity of the induced labels. Motivated by this diagnosis, we propose PAGER (Position-Aware Grounded Evidence Reinforcement), a GRPO training objective with separate signals for supported fine-grained middle evidence, evidence-bearing tail continuation, claim coverage, and content-aware length control. The result is a closed loop: the same ordered claim representation identifies where long captions fail and supplies the position-specific reward needed to repair those failures. Experiments on detailed-caption benchmarks show that PAGER improves position-wise grounding and pairwise caption preference while preserving competitive global caption quality.
Tianle Zhou, Yang Zhang, Tianyu Wu, Zhuohan Wang, William Nixon, Juncheng Yang
Looped transformers apply a shared block several times per token to improve accuracy without adding parameters. For models such as Ouro that keep a separate key-value (KV) cache per loop, cache memory limits the batch size that amortizes repeated weight reads. Uniform-bit quantization and fixed rules per loop do not allocate precision according to each layer's sensitivity in each loop. Measurements on Ouro show that quantization sensitivity varies across (loop, layer) cells, and the sum of individual cell damages approximates the overall output damage. Sharing damage also varies by cell, with later loops generally tolerating sharing better. Skipping low-attention values to save reads produces errors that compound across loops. LoopKV uses a one-time offline profile to choose 4-bit storage, 2-bit storage or reuse of the final loop's KV for each cell. The planner minimizes reads under budgets for predicted damage and storage. During decoding, each loop reads the high 2 bits of each 4-bit value and fetches the low bits only where its own attention weights require 4-bit precision. A single fused kernel in vLLM executes the plan. Serving Ouro-1.4B on one RTX PRO 6000, the memory-optimized LoopKV-M plan stores 35% less KV than uniform 4-bit quantization (KIVI-4) on the same fused kernel and CUDA graph. It sustains 1.7-2.2x KIVI-4's goodput under a latency service-level objective with a 16K-token maximum length. None of its averages on six public benchmarks falls significantly below BF16's.
Saurabh Kataria, Yi Wu, Zhaoliang Chen, Hyunjung Gloria Kwak, Yuhao Xu, Lovely Yeswanth Panchumarthi, Ran Xiao, Jiaying Lu 等 26 位作者
Time-series Foundation Models (FMs) are pretrained on either heterogeneous corpora or specific signal domains. Using Photoplethysmography (PPG) as a representative modality, we compare generalist time-series FMs, PPG-specialist FMs, and traditional machine learning (ML) baselines for PPG-based health assessment. We conduct a two-stage benchmark. First, we evaluate multiple FM families across 23 downstream evaluations (e.g., atrial fibrillation detection, vital-sign estimation). Second, we conduct a focused four-FM comparison (MOMENT, Chronos-2, GPT-PPG, MOMENT-PPG) at multiple model sizes on an expanded 51-task suite, under both frozen-backbone and full-model tuning, to study tuning strategy, model scaling, and cross-dataset transfer. The multi-model benchmark shows that PPG-specialist FMs achieve the strongest aggregate performance, but their advantage is not uniform: generalist FMs remain competitive on several tasks, and traditional ML baselines outperform FMs on selected clinical regression tasks. The focused analysis shows that the advantage of PPG-specific pretraining appears mainly with frozen backbones and on targets that the waveform directly reflects, such as atrial fibrillation and vital signs. Full-model tuning narrows this advantage, model rankings change with the tuning strategy, and the largest model variant is the best choice in only about one-third of cases. Additionally, controlled experiments with two generalist models show that initializing PPG pretraining from generalist checkpoints provides an early advantage over training from scratch, even when the pretrained features offer little downstream benefit. Together, these findings suggest that PPG-specific pretraining is useful but task-dependent, and that FMs should be evaluated against simpler statistical, traditional ML, and clinical history-based baselines.
On-policy distillation (OPD) trains a student on its own generations by matching the teacher's token probabilities. These probabilities entangle two kinds of information: which tokens the teacher prefers, and how confident it is. The latter reflects the teacher's own capacity and calibration, which a smaller student may be unable to reproduce. Because the teacher temperature changes these values without changing the ordering, OPD is also sensitive to the temperature choice. We ask whether the teacher's ordering of candidate tokens is enough, and propose Order-Only On-Policy Distillation (O$^{3}$PD). At each student-generated prefix, O$^{3}$PD trains the student to reproduce the order of the teacher's top-$K$ tokens, without fitting their probability values. The objective is therefore invariant to the teacher temperature and needs no temperature tuning. Across six teacher--student pairs, O$^{3}$PD outperforms OPD on the average over HMMT25, AIME24, and AIME25 in every pair, by up to 2.9 points, with gains of up to 5.6 points on AIME25. Other ranking losses yield similar gains, so the benefit comes from ordinal supervision itself. Analysis shows that O$^{3}$PD departs further from the teacher's probabilities than OPD, yet reproduces its ranking more faithfully and yields a stronger student. The ranking, not the probability values, thus carries the knowledge that on-policy distillation needs to transfer.
Multimodal Large Language Models (MLLMs) are typically designed under the assumption that all modalities available during training will also be accessible at inference. However, many real-world settings violate this assumption, requiring models to operate under a privileged modality setting, where auxiliary modalities are available only during training. While these modalities contain valuable information, existing MLLMs largely fail to leverage them effectively, as they treat modalities as interchangeable inputs rather than sources of complementary supervision. We propose Mixture of Probes (MoP), a novel framework that disentangles modality-specific and modality-general signals within the MLLM, allowing the model to preserve modality-dependent structure while learning transferable representations across modalities. At its core, MoP achieves this through a structured probing mechanism that extracts and organizes information from intermediate representations of a shared modality encoder, rather than relying only on final-layer alignment as done in existing MLLMs. To support this disentanglement, we further introduce MoP Cross-modal Training (MoP-X), a training strategy for MoP centered around a probe disentanglement loss that prevents probe collapse and encourages cross-modal learning. We evaluate MoP across two domains spanning eight tasks and four modalities under a comprehensive evaluation protocol tailored to the privileged modality setting, where each modality is independently treated as the sole input at inference time. MoP consistently outperforms strong MLLM baselines, achieving up to 65% relative improvement, demonstrating that auxiliary modalities, even when unavailable at inference, can provide substantial gains when effectively leveraged during training. Code, model checkpoints, and evaluation protocols will be made available at https://github.com/Sony/MoP.
We establish tight lower bounds on the adaptive and non-adaptive quantum query complexities of Promised $k$-to-one Finding on an exactly $k$-to-one input string $x\in\mathbb Z_q^n$, for every fixed $k\ge2$ and every feasible alphabet size $q\ge n/k$. The lower bounds are $Ω_k(n^{1/2-1/(2(2^k-1))})$ for unrestricted adaptive algorithms and $Ω_k(n^{1-1/k})$ for one-round non-adaptive algorithms. The former matches the adaptive quantum algorithm of Liu and Zhandry [LZ19], whereas the latter matches classical birthday sampling even when the parallel queries are arbitrarily entangled. We also establish the non-adaptive lower bound $Ω_k(\sqrt{n/k})$ for $k$-to-one Testing when $q\ge n$, resolving an open question posed by Koiran, Landes, Portier, and Yao [KLPY10]. Our proofs generate the balanced input from a uniformly random permutation and use compressed permutation oracles to describe the information learned by the algorithm. The adaptive proof bounds the growth of partial collisions through successive queries. The one-round proof instead analyzes the state after the parallel queries directly, bounding the probability that the compressed database contains a full collision. In both proofs, a short verification step accounts for correct outputs that include unrecorded positions. For the decision problem, we also compare one-round states with independent and distinct output labels.
Before AI can be trusted to discover new science, it should be able to reproduce what science already knows. Reproduction provides a rigorous checkpoint for an AI scientist: the target claim is known, yet recovering it still requires translating an incomplete scientific description into an executable experiment and producing evidence that can be independently verified. Existing research-reproduction benchmarks largely focus on AI and machine learning, where experiments share relatively homogeneous software stacks. Extending this capability across scientific disciplines introduces a different challenge: agents must operate domain-specific scientific software, make physical and numerical choices, diagnose long-running computations, and distinguish valid evidence from merely plausible outputs. We introduce PAPERBENCHX, a benchmark of 93 expert-constructed paper-reproduction tasks spanning 12 research areas across electromagnetics, photonics, chemistry, materials science, biology, and robotics. Each task provides the source paper, a public task specification, and a domain-native scientific environment involving specialized scientific software. Agents must reconstruct a key experiment, execute and diagnose it, and deliver a runnable workflow that supports the target scientific claim. To evaluate scientific validity rather than surface-level agreement, we replay every submission from scratch and score only regenerated evidence against expert-validated criteria. Across 10 frontier agent configurations, the best configuration fully reproduces only 14% tasks under the 90% threshold. Meanwhile, mean rubric scores range from 27.53% to 59.08%, indicating that substantial partial progress rarely translates into complete scientific reproduction. These results reveal a large gap between paper understanding and reliably producing reproducible scientific evidence.
Rong Fu, Xiaowen Ma, Kun Liu, Wangyu Wu, Ziyu Kong, Jia Yee Tan, Tailong Luo, Xianda Li 等 10 位作者
Deploying expressive learning models directly on programmable dataplanes promises line-rate, low-latency traffic analysis but remains hindered by strict hardware constraints and the need for predictable, auditable behavior. Chimera introduces a principled framework that maps attention-oriented neural computations and symbolic constraints onto dataplane primitives, enabling trustworthy inference within the match-action pipeline. Chimera combines a kernelized, linearized attention approximation with a two-layer key-selection hierarchy and a cascade fusion mechanism that enforces hard symbolic guarantees while preserving neural expressivity. The design includes a hardware-aware mapping protocol and a two-timescale update scheme that together permit stable, line-rate operation under realistic dataplane budgets. The paper presents the Chimera architecture, a hardware mapping strategy, and empirical evidence showing that neuro-symbolic attention primitives can achieve high-fidelity inference within the resource envelope of commodity programmable switches.
A streaming estimator must compute from the evidence that fits in memory. We determine when storing an accurate mean is easier than computing it. For Gaussian mean estimation over the Euclidean unit ball in $\mathbb{R}^d$, we establish the minimax rate under squared error with $K$ persistent states, a known horizon, and a public time index: $\min\{1, K^{-2/d}+\frac{d}{n} (1+\frac{\log(ed)}{1+(\log K)/d})\}$. Unrestricted computation followed by $K$-level encoding has rate $\min\{1,d/n+K^{-2/d}\}$. At a fixed sufficiently small error, statistically optimal sample size therefore requires order $d\log d$ working bits, while order $d$ bits suffice to store the answer. The lower bound tracks repeated state compression through an entropy potential obtained by Gaussian smoothing. A matching construction combines a multiscale modular counter with workspace reused across coordinate groups. In one dimension, it attains the terminal rate $n^{-1}+K^{-2}$. Any fixed confidence requirement that the realized loss not exceed the zero baseline preserves these rates. Increasing confidence creates a further memory requirement: at fixed nontrivial risk in every fixed dimension, failure probability $e^{-γn}$ requires and permits order $\log n$ working states for sufficiently small fixed $γ>0$. In the scalar problem, a fixed terminal alphabet suffices under the same protection. These bounds separate the memory needed to represent an answer from the memory needed to accumulate evidence and validate a decision.
Multi-agent LLM systems increasingly rely on persistent memory and inter-agent communication. These capabilities create privacy risks that extend beyond individual interactions: a release appropriate in isolation may become inappropriate when combined with information previously disclosed to a recipient or propagated to downstream agents along the planned workflow. Existing Contextual Integrity (CI) mechanisms for agentic systems typically rely on local judgments that assess each disclosure in isolation, with or without regard to the recipient's history, but do not account for how information propagates downstream. However, compliance at the current boundary does not guarantee appropriateness across the workflow. We introduce Task-Graph Contextual Integrity (TG-CI), to our knowledge the first task-graph-native formalization of Contextual Integrity for multi-agent systems. TG-CI represents a workflow as a Multi-Agent Task Graph (MATG) and extends CI from atomic transfers to stateful, graph-level information flows by jointly considering an item's intended propagation scope and recipients' evolving disclosure states. We operationalize TG-CI through propagation-scoped authorization under two complementary policies: Full-Scope, which requires admissibility across the complete downstream scope, and Auto-Region, which confines each item to its maximal admissible region using lightweight frontier gates. We further introduce MAPLE-Bench, a benchmark of 144 cases across four scenarios that jointly vary MATG topology, disclosure history, compromised-agent placement, and item-level privacy policies. In our primary experimental configuration, TG-CI reduces the mean unauthorized disclosure count by 85.8-94.1% relative to the Local-History CI baseline, while retaining 93.6-96.1% of the baseline's task accuracy, with the two enforcement policies providing complementary privacy--cost--utility trade-offs.
Thomson Yen, Gilbert Yang, Naimeng Ye, Minghui Chen, Dong Xia, Tianyi Shen, Yechan Jeon, Yaqin Chen 等 11 位作者
Language-model agents differ fundamentally from prior intelligent systems in their ability to learn from natural language in context. This capability enables agents to iteratively refine their answers in response to feedback, a prevalent mode of operation in modern agentic workflows. However, the community primarily ranks models by metrics such as Pass@1 or Pass@K, which cannot capture how they utilize feedback on prior attempts. We introduce Seq@K, a natural evaluation protocol that measures whether a model can solve a problem within K sequential attempts, provided feedback after each previous failure. Unlike multi-turn agentic benchmarks that examine long-horizon planning, Seq@K isolates and formalizes the core capability of refinement from feedback. We evaluate Seq@K on a suite of 7 benchmarks and 16 frontier models, and show that Seq@K reveals a materially different model-selection landscape: under Seq@K, (1) model rankings often differ, (2) the Pareto frontiers admit more efficient configurations and (3) the gains from choosing the optimal model configuration can widen or shrink relative to those of Pass@K. We further show that analyzing model trajectories under Seq@K helps characterize iterative refinement behavior, such as the rate at which models address issues raised in feedback, or how rigidly they hold onto their own answers. Finally, we study how the nature of feedback affects models' iterative refinement, showing that practitioners must carefully design and benchmark the feedback loops of iterative agentic systems. Together, these results establish how Seq@K offers a more faithful metric for intelligent systems that typically operate in such feedback loop, and the necessity to consider Seq@K in evaluating models for these systems.
Zhuoxiao Li, Xinyi Liu, Hongye Hou, Jing OU, Kang Du, Shuai Zhang, Dongli Wu, Taoyu Wu 等 11 位作者
Long-sequence panoramic reconstruction requires detailed scene geometry and globally consistent camera poses across distant revisits. Panoramic SLAM and incremental SfM rely on feature correspondences, but repeated, similar structures and small partial overlap make reliable long-range connections difficult to establish, leaving accumulated drift unresolved. Feed-forward models offer detailed local geometry, but their high computational costs limit inference window size, necessitating additional constraints for global consistency. This motivates combining continuous feature-based tracking and sparse mapping with rapid, high-quality local reconstruction from feed-forward models to achieve accurate localization and globally consistent dense mapping over long sequences. We therefore propose a panoramic reconstruction framework that couples an ORB-based sparse panoramic mapping frontend with a fine-tuned feed-forward model. Our key insight is that predicted local geometry enables 3D verification of feature matches across distant revisits, yielding long-range camera constraints even between regions. We calibrate predicted depths with sparse landmarks and use verified connections to optimize the global pose graph, updating dense geometry with the corrected camera poses. We also introduce PanoRecon, a seven-scene benchmark with repetitive architecture, large and small loops, and reference camera poses and LiDAR geometry for evaluating global consistency. Our method achieves the lowest mean absolute trajectory error and high F-1 Score among the evaluated methods. Project Page:https://zhuoxiaoli.github.io/HoloRecon_page/
Self-driving laboratories (SDLs) are laboratory systems that couple experimental planning, automated execution, and measurement feedback in a closed loop. Integrating flexible robot control with the typed application programming interfaces (APIs) and continuous measurements exposed by laboratory instruments remains an execution-interface challenge. To address this challenge, we propose a traceable framework comprising four components: LAM (Lab Action Model), OpenSDL, Safe-SDL, and L3-Bench. LAM is a proposed measurement-conditioned policy architecture that extends vision-language-action paradigms to generate both continuous robot trajectories and typed instrument API calls. OpenSDL is an execution substrate with an initial control implementation, designed to map role-level operations to heterogeneous devices, coordinate execution, and record timestamped causal events. The proposed Safe-SDL runtime gate combines pre-dispatch schema and resource checks with signal temporal logic monitoring during execution, including extended waiting periods. L3-Bench (Level-3 Laboratory Benchmark) specifies an episode-level evaluation protocol covering physical outcomes, timing, reproducibility, safety violations, and fault recovery. Together, these components provide an end-to-end design for linking hybrid action generation, monitored device execution, measurement feedback, and experimental evaluation, addressing coordination and traceability challenges in SDL workflows.
Bimanual robot manipulation plays an important role in real-world robotic applications, but the high cost of collecting bimanual demonstrations limits the scalable development of bimanual policy learning. Meanwhile, existing single-arm manipulation data are difficult to directly transfer to bimanual systems due to the Single-to-Dual gap in action spaces, coordination requirements, and arm-side spatial preference. To address this issue, we propose RoboSTD, a zero-shot Single-to-Dual data construction method for bimanual robot learning. RoboSTD first applies sagittal-plane mirroring to observations, proprioception, and actions to construct executable contralateral pseudo-demonstrations. It then uses an LLM to generate structured spatio-temporal coordination constraints that guide the reconstruction of pseudo-bimanual supervision across workspace regions and task stages. We evaluate RoboSTD on the AgileX ALOHA platform in both RoboTwin simulation and real-world settings. Experiments demonstrate the effectiveness of RoboSTD for bimanual policy learning. Without requiring additional bimanual demonstrations, RoboSTD increases the average number of collected cups to 2.11x that of the single-arm baseline in the cup collection task. Project page: https://huahungy.github.io/RoboSTD/
Frequency-domain information is beneficial for semantic segmentation. However, existing frequency-based methods often suffer from insufficient frequency utilization, inefficient fusion, and high computational complexity. To address these issues, we propose FGD2HF, a dual-domain fusion network that leverages frequency features to guide spatial-domain fusion and enhancement. Specifically, we introduce a novel Hybrid-Attention Encoder (NAHA) that jointly models local and global relationships through neighborhood and hypergraph attention, enhancing the network's ability to extract effective features. We further introduce a Wavelet Transform and Enhancement (WTE) module to extract and refine low- and high-frequency cues from preceding-stage features, followed by Frequency-guided Dual-domain Fusion (FGDF) modules that use frequency cues to condition current-stage spatial grouping and feature propagation. This design incorporates both global and local frequency guidance into spatial feature modulation, enhancement and fusion with low computational cost. Extensive comparisons with numerous state-of-the-art semantic segmentation methods demonstrate that our method is highly competitive. Our full FGD2HF-B model achieves 48.7% mIoU on ADE20K (+0.3% over SegMAN-B with 18% fewer GFLOPs), 83.2% mIoU on Cityscapes (comparable to SegMAN-B with 21% fewer GFLOPs), and 1.1% higher mIoU than VWFormer-T on SBD with 21% fewer GFLOPs, while also achieving significant improvements in boundary segmentation accuracy.
LLM-based machine translation evaluation can closely match human judgments, but in practice it remains largely diagnostic, with the signals rarely translating into direct quality improvements under real production constraints. We propose a two-stage, evaluator-guided automatic post-editing framework that turns MQM-style evaluation into targeted repairs: a retrieval-augmented LLM evaluator outputs structured, span-level MQM diagnoses under an explicit edit contract, and a separate LLM post-editor applies minimal edits restricted to those diagnoses. This separation improves controllability and reduces paraphrastic drift compared to one-stage "judge-and-refine” baselines. In a systematic study involving seven LLMs spanning three model providers and seven languages, our best configuration consistently improves both reference-based COMET and CometKiwi scores over one-stage post-edit methods, while the evaluator’s error spans and severities show strong agreement with human MQM annotations and human editor preferences.
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 6: Industry Track) · 2026 · 正式发表
Conversational agents, such as ChatGPT and Doubao, have become essential daily assistants for billions of users. To further enhance engagement, these systems are evolving from passive responders to proactive companions. However, existing efforts focus on activation within ongoing dialogues, while overlooking a key real-world bottleneck. In the conversation initiation stage, users may have a vague need but no explicit query intent, creating a first-message barrier where the conversation holds before it begins. To overcome this, we introduce Conversation Starter Generation: generating personalized starters to guide users into conversation. However, unlike in-conversation stages where immediate context guides the response, initiation must operate in a cold-start moment without explicit user intent. To pioneer in this direction, we present IceBreaker that frames human ice-breaking as a two-step handshake: (i) evoke resonance via Resonance-Aware Interest Distillation from session summaries to capture trigger interests, and (ii) stimulate interaction via Interaction-Oriented Starter Generation, optimized with personalized preference alignment and a self-reinforced loop to maximize engagement. Online A/B tests on one of the world’s largest conversational agent products show that IceBreaker improves user active days by +1.84‰ and click-through rate by +94.25‰, and has been deployed in production.
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 6: Industry Track) · 2026 · 正式发表
Bowen Yang, Kaiming Jin national university of singaore, National University of Singapore, Zhenyu Wu, Zhaoyang Liu Hong Kong University of Science and Technology and Shanghai AI Laboratory, Qiushi Sun University of Hong Kong, Zehao Li University of Science and Technology of China, JingJing Xie, Zhoumianze Liu 等 15 位作者
While Vision-Language Models (VLMs) have significantly advanced Computer-Using Agents (CUAs), current agentic frameworks struggle with robustness in novel domains and long-horizon workflows due to the absence of visual-aware tutorial retrieval and the lack of granular control over historical visual context curation and pruning. To bridge these gaps, we introduce OS-Symphony, a holistic framework that comprises an Orchestrator coordinating two key innovations for robust automation: (1) a Reflection-Memory Agent that utilizes milestone-driven long-term memory to enable trajectory-level self-correction, effectively mitigating visual context loss in long-horizon tasks; (2) Versatile Tool Agents featuring a Multimodal Searcher that adopts a “SeeAct” paradigm to navigate a browser-based sandbox to synthesize live, visually aligned tutorials, thereby resolving fidelity issues in unseen scenarios. Experimental results demonstrate that OS-Symphony delivers substantial performance gains across varying model scales, establishing new state-of-the-art results on three online benchmarks, notably achieving 65.84% on OSWorld. All research assets will be made publicly available.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Qiuyuan Ai, Zenghuang Fu, Zhaoyang Li, Ping Jiang Peking University, Haoyu Wu Mininglamp.com, Jie Song Peking University, Guannan He Peking University
Scaling LLM-based agents to long-horizon deep research is constrained by the context-noise trade-off, where linear history accumulation degrades reasoning and dilutes fine-grained evidence. To address this, we introduce the Cognitive Scaffold, a factorized memory architecture that decouples the cognitive state into a Fluid Working Context for immediate reasoning and a persistent Knowledge Graph for long-term retention. Unlike unstructured summarization, our framework employs a Rejection Sampling Fine-Tuning (RFT) pipeline to crystallize saturated context into structured event snapshots, strictly enforcing atomic constraints to preserve numerical values and entities. During reasoning, a thought-driven dual-path retrieval mechanism enables the agent to proactively recover precise evidence. Empirical evaluations on Xbench-DeepSearch, BrowseComp-ZH, and GAIA demonstrate that Cognitive Scaffold consistently outperforms baselines, achieving 74.7% Avg@3 and 87.0% Pass@3 on Xbench-DeepSearch, 48.5% Avg@3 and 65.9% Pass@3 on BrowseComp-ZH, and 72.8% Avg@3 and 88.3% Pass@3 on GAIA, while reducing compression hallucinations to 5.3%. We open-source our codebase to facilitate future research.
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