Yi Jing, Zao Dai, Jinwu Hu, Zijun Yao, Lei Hou, Juanzi Li, Xiaozhi Wang
Data engineering for large language model (LLM) reinforcement learning (RL) decides which samples to train on, when to present them, and which to batch together. Existing methods mostly rely on external feedback, such as pass rates or difficulty labels, which observes only the outcome on each sample. We argue that model internals offer richer and more model-intrinsic feedback that can guide all three decisions. We propose SAERL which organizes RL training data in the feature space of Sparse Autoencoders (SAEs), a mechanistic interpretability tool that decomposes model internals into sparse, interpretable features. SAE features let us view the model as a set of functional regions, each formed by features that related samples activate together. For each sample, its SAE activations indicate which regions it engages, and its gradients projected onto SAE features indicate how training on it would change these regions. Guided by this view, SAERL keeps samples with strong and non-redundant gradient signals, trains each functional region from easy to hard, and trains one functional region at a time, drawing each batch from a single activation cluster with moderate cross-cluster mixing. The resulting curriculum is constructed once with small models and transfers to larger ones. Experiments on mathematical reasoning show consistent gains across model scales and RL algorithms. For instance, on Qwen2.5-Math-1.5B, SAERL improves the average accuracy of GRPO by 3.1% and reaches the target accuracy with 42% fewer training steps. Further analyses reveal that ordering and batching are complementary. These results demonstrate that model internals are a powerful and practical foundation for post-training data engineering.
Vision-Language-Action (VLA) models exhibit remarkable action generation for embodied intelligence, but their heavy compute make deployment on edge platforms impractical. Aggressive, sub-4-bit weight quantization is the natural solution, yet existing post-training quantization (PTQ) methods suffer severe performance degradation in this regime. To address this, we introduce ActQuant, an action-guided mixed-precision PTQ framework that operates in two stages: (1) an inter-tensor bit allocator that assigns each weight matrix a single bit-width based on how much it contributes to predicting the agent's actions; (2) an intra-tensor scale optimizer tunes per-block quantization scales using action-aware curvature, so that dynamic range is concentrated on the weights most influential for control. To deliver the on-device benefits of our aggressive quantization, we further introduce OmniModel.cpp, an agentic conversion pipeline that ports architectures into a native C/C++ runtime with efficient low-bit kernels. We evaluate ActQuant both in simulation and on a real-world 6-DoF UR3 arm, with all models deployed through OmniModel.cpp. On the LIBERO benchmark, ActQuant is the only method that operates at or below 3 bits-per-weight, retaining 95.0% on OpenVLA-OFT and 94.8% on $π_{0.5}$. Pushed further, ActQuant reaches 2.5 bpw at 90.1% on OpenVLA-OFT, compressing the backbone from 14.3 GB to 2.7 GB (5.3$\times$). On the physical UR3 arm, $π_{0.5}$ quantized with ActQuant retains the baseline's success rate while reducing the memory footprint by 2.5$\times$.
Dingxian Wang, Zhe Fu, Jiaqi Deng, Chao Yin, Ganesh Prabhu, Karthik Gomadam Rajagopal, Darwin Yi, Huan Huo 等 12 位作者
Coding agents increasingly collaborate on long-term projects. When taking over a project, they need to recover currently applicable requirements from the interaction history and update them in response to new requests. However, most coding benchmarks do not include long interaction histories, and some evaluations that consider history still do not cover the full process from requirement reconstruction to project delivery. We therefore provide agents with the interaction history, current request, and pre-request project environment, ask them to make the requested changes, and assess takeover performance through delivery tests. Across six model-harness configurations and 37 targets eligible for executable validation, the mean complete-delivery rate under Full History is only 4.95%. We then use requirement atoms to independently diagnose a larger set of takeover targets, examining relevant information identification, pre-request state reconstruction, and current request processing. We further supply correct pre-request or target states to examine the remaining difficulties in requirement processing and project modification. The diagnostics show that none of the six configurations produces fully correct post-request requirement states, and a case illustrates omission of previously recovered constraints during updating. With correct target states supplied, the mean complete-delivery rate is 8.11%, with build and regression pass rates of 96.40% and 90.99%, but target-test success remains 9.01%. These results show difficulties in both forming target requirement states and implementing supplied targets in existing projects. Based on these findings, we introduce ReqMemBench, which combines requirement diagnostics with delivery validation to support systematic evaluation of project takeover.
Le Jiang, Xiangyu Bai, Bishoy Galoaa, Shayda Moezzi, Caleb James Lee, Tooba Imtiaz, Edmund Yeh, Jennifer Dy 等 10 位作者
We present PanoWorld, a panoramic video world model that generates geometry-consistent 360$\degree$ video from a single image and a caption. Existing panoramic video methods optimize primarily for visual realism and do not explicitly constrain the underlying 3D scene state, producing outputs that appear plausible yet exhibit inconsistent depth, broken correspondences, and implausible motion across the spherical surface. We address this gap by framing panoramic video generation as a geometry- and dynamics-consistent latent state modeling problem rather than pure visual synthesis. Building on a pre-trained perspective video world model, we introduce two lightweight regularizers: a depth consistency loss against pseudo ground-truth panoramic depth, and a trajectory consistency loss that supervises the 3D world-frame positions of tracked points across time. We further apply spherical-geometry-aware adaptation to the conditioning and positional encoding. We additionally introduce PanoGeo, a unified geometry-aware panoramic video dataset with consistent depth, trajectory, and prompt annotations across diverse real and synthetic sources, used for both training and stratified evaluation. Experiments show that PanoWorld improves geometric consistency over prior panoramic generation methods while maintaining competitive visual realism, establishing that panoramic video generation must be treated as a geometric modeling problem to support the holistic spatial understanding requirements of embodied AI applications. Code is available at https://github.com/ostadabbas/PanoWorld.
Ning Yang, Chengzhi Wang, Yibo Liu, Baoliang Tian, Haijun Zhang
Prefill-only KV compression freezes a token subset at the end of prefill and decodes from it without further eviction. The retention decision is therefore irreversible, yet existing methods estimate the corrective signals it relies on, per-head reliability and prompt-level compression sensitivity, online from a single noisy prompt. We argue this is the wrong statistical unit: these signals exhibit far higher cross-prompt regularity than within-prompt signal-to-noise. We introduce \textsc{CompilerKV}, a KV-retention policy whose corrective tables are compiled offline from a calibration corpus, reducing online correction after the standard observation-window scan to $O(1)$ lookups plus a budget clamp. We find that compiled retention tables behave as portable architectural priors: rankings transfer across disjoint corpora on four backbones (mean Spearman $\barρ{=}0.90$), and direct model-to-model table transfer costs only $0.4$--$0.8$ LongBench points on average. At a 512-token budget, \textsc{CompilerKV} attains compressed-SOTA on all four backbones, improving over the strongest prefill-only baseline by $+1.67$ points on average (task-bootstrap 95\% CI $[+1.08,+2.37]$). Pressure regimes amplify the gap: under a fixed $512/32k$ cache ratio, \textsc{CompilerKV} remains the strongest compressed method through 128k RULER ($\sim\!73$ vs.\ FullKV $\sim\!79$, SnapKV $\sim\!38$); on 32k NIAH it reaches $0.89$ vs.\ SnapKV $0.42$; and at 32k input, retaining only $1.56\%$ of the prefill KV, batch-16 serving remains feasible where FullKV is OOM.
Kaiyuan Chen, Shuangyu Xie, Letian Fu, Justin Yu, William Pacini, Sandeep Bajamahal, Hudson Kim, Jaimyn Drake 等 24 位作者
For robots to work reliably in commercial and industrial applications, recent advances in agentic coding systems have potential to combine interpretable robot programming with the open-world adaptability of model-free policies. We focus on "Variational Automation" (VA), a class of tasks that have larger variations in object geometry and pose than fixed automation. Model-free policies often struggle to close the reliability gap for VA tasks, which must be executed persistently and reliably in commercial and industrial applications. Motivated by prior work on Task and Motion Planning (TAMP) and the Robot Operating System (ROS), we introduce Graph-as-Policy (GaP), a multi-agent coding harness that generates directed computation graphs with perception, planning, and control nodes from a Modular Open Robot Skill Library (MORSL). GaP generates an internal simulation environment to rehearse task instances with different graphs in parallel to iteratively refine the graph structure and parameters to improve success rates and throughput. Evaluation with 8 new open VA task benchmarks, 4 in simulation and 4 in real-world, suggests that GaP can achieve success rates that significantly outperform baselines. Details, code, and data: https://graph-robots.github.io/gap
Reinforcement learning (RL) has emerged as a paradigm for fine-tuning large-scale generative models, such as diffusion and flow models, to align with complex human preferences and user-specified tasks. A fundamental limitation remains \textit{the curse of diversity collapse}, where the objective formulation and optimization landscape inherently collapse the policy to a Dirac delta distribution. To address this challenge, we propose \textbf{DRIFT} (\textbf{D}ive\textbf{R}sity-\textbf{I}ncentivized Reinforcement \textbf{F}ine-\textbf{T}uning for Versatile Image Generation), an innovative framework that systematically incentivizes output diversity throughout the on-policy fine-tuning process, reconciling strong task alignment with high generation diversity to enhance versatility essential for applications that demand diverse candidate generations. We approach the problem across three representative perspectives: i) \textbf{sampling} a reward-concentrated subset that filters out reward outliers to prevent premature collapse; ii) \textbf{prompting} with stochastic variations to expand the conditioning space, and iii) \textbf{optimization} of the intra-group diversity with a potential-based reward shaping mechanism. Experimental results show that DRIFT exhibits clear Pareto dominance in task alignment and generation diversity, achieving 3.88\%$\sim$81.42\% higher diversity at matched alignment and 14.38\%$\sim$56.45\% higher alignment at matched diversity. Our code is available at \textcolor{magenta}{\href{https://github.com/NJU-RL/DRIFT}{https://github.com/NJU-RL/DRIFT}}.
Jun Liu, Dajie Sun, Zhenglun Kong, Zhenjiang Fan, Peiyan Dong, Fan Yang, Wenbin Zhang, Geng Yuan 等 11 位作者
World-action models increasingly apply test-time scaling, spending extra inference computation on candidate selection and adaptive allocation to improve control. Such gains are credible only when candidates carry exploitable signal, nominal compute savings become real speedup, and evaluation stays reliable. We audit all three on Cosmos Policy and RoboCasa with a premise-first, pre-registered methodology. On clean data, a learned value selector is 1.66 pp worse than random; a hindsight oracle gains only 1.40 pp, with an episode-level interval spanning zero; and 82% of states show identical candidate outcomes, leaving little selection headroom. Token pruning cuts per-forward FLOPs by up to 47% but yields at most a measured 1.15$\times$ end-to-end speedup on the deployed path (1.25$\times$ on a decode-excluded path), short of a 1.5$\times$ target, and our adaptive-allocation experiment is inconclusive under a power gate. We further identify and fix a silent rendering corruption whose rate rises sharply along compute-heavy planning paths while leaving the official direct-policy result unaffected, and release a detector, fix, and scripts. These findings show why test-time scaling should be audited as a chain of premises (decision signal, realized efficiency, evaluationintegrity) rather than by success rate alone. Conclusions are scoped to the tested configuration.
Xuwei Tan, Jingchun Huang, Ming Cai, Shuangquan Lyu, Tiange Wang, Yinzhi Wang, Ronghua Shi, Jingwei Song 等 15 位作者
Multi-expert agent marketplaces -- where creators publish domain-specific AI experts for end users to consult -- are an emerging deployment pattern for language technology. A published expert can receive many requests yet little evidence about its creator, who remains active elsewhere: the cold-expert problem. We frame continuous cognitive distillation as converting attributed creator observations into revisable decision rules, communication patterns, and domain procedures. CogniLife proposes a cross-expert attribution design for this problem, with Creator Profiles for transferable creator patterns and Expert Journals for listing-specific knowledge; this design is not yet implemented or evaluated. A deployed platform partially realizes the surrounding lifecycle through creator-authorized authoring, capability-owned proposal filtering, and reversible consolidation of eligible owner-scoped memory. On ProfyBench-v1, an 85-task offline benchmark built from replay seeds and controlled internal tests, the Profy Full configuration as a whole achieves a mean overall score of 4.65/5 from the primary LLM judge, compared with 4.01 for Raw LLM and 4.28 for Expert Prompt, with the largest gains in completeness, specificity, and missing-information tasks. These configuration-level scores do not test cold-expert mitigation, cross-expert transfer, or longitudinal distillation.
We characterize the optimal space for one-pass spanning-forest computation in dynamic streams of $v$-vertex simple graphs with failure probability at most $δ$. Nelson and Yu proved an $Ω(v\log^2v\log(v/δ))$-bit lower bound, establishing the AGM algorithm's optimality for each fixed $0<ε<1$ and $2^{-v^{1-ε}}<δ\le1/4$. We resolve the remaining high-confidence regime: for every $0<δ\le1/4$, matching lower bounds and an algorithm give \[ Θ\!\left(\min\left\{v^2,\;v\log(v/δ)\log^2\!\left(2+\frac{v}{\log(v/δ)}\right)\right\}\right) \] bits, improving space by up to a factor of $Θ((\log v/\log\log v)^2)$ in this regime. We also study streams with at most $Z$ previously inserted edges deleted and not yet reinserted at report time, where the integer $Z\ge0$ is given in advance. Total deletions, restorations, and intermediate absence counts are unrestricted. For a single report and every $0<δ\le1/4$, matching lower bounds and a one-pass randomized algorithm yield optimal space $Θ(\mathcal L(v,Z,\log(1/δ)))$ bits, where $x_Z(t)=\min\{Z+1,v,\sqrt{vt}\}$ and \[ \mathcal L(v,Z,t)=v\log v\min\{\log^2v,\log^2(Z+2)\} +vt\log\!\left(1+\frac{x_Z(t)}t\right)\log\!\left(1+\frac v{x_Z(t)}\right). \] At zero error, a lower bound for all zero-error algorithms and a matching deterministic algorithm give optimal space \[ Θ\!\left(\min\{v^2,v(Z+1)\log(1+v/(Z+1))\}\right) \] bits. Together, these results characterize the memory required across all confidence levels and report-time edge-loss bounds.
Few-step autoregressive video diffusion enables efficient generation, yet maintaining consistency over long rollouts remains challenging, especially when subjects or scenes reappear after going out of sight. Despite progress in fidelity and rollout robustness, existing distillation methods do not fully exploit richer historical information to constrain the student's current prediction. To address this gap, we introduce \textbf{History Forcing}, a unified on-policy distillation framework that strengthens historical supervision for few-step autoregressive video generators. The framework uses the residual between teacher predictions with and without informative historical context to decouple this supervision. Such a historical context can be obtained using earlier generated chunks with less accumulated drift or a fixed-view teacher rollout that keeps the scene visible while the student looks away. The residual is then weighted and added to the score or velocity prediction used for distillation, forming a corrected target. By aligning the student with this corrected target through distillation, History Forcing incorporates long-range historical constraints into autoregressive generation. Our method achieves state-of-the-art performance among 1.3B models, with a VBench quality score of 85.07 for 15s video generation and a VBench-Long aggregate score of 79.82 for 60s generation. For out-of-sight world modeling, evaluations further demonstrate more plausible scene evolution and more consistent states when the camera returns, with Evolution Quality improving from 2.35 to 2.70 on PlayWorld and Object Consistency from 58.43 to 62.18 on MemoBench.
Detecting defects in real-world industrial inspection often requires learning from only a few normal images per category. Recent few-shot anomaly detection methods have achieved strong performance by measuring feature differences between query images and normal references. However, visual differences do not always indicate genuine defects, as normal appearance variations can also produce high anomaly scores. Conversely, structural or logical anomalies may not produce strong local feature differences, making them difficult to detect. To address this limitation, we introduce VerdictAD, a training-free framework that refines existing anomaly detectors through visual decisions. Rather than replacing existing detectors, VerdictAD adopts a typed decision model to assess the query image and its local regions against a few normal references. By directly producing structured decision probabilities without text generation, the model provides complementary evidence for refining anomaly responses at different spatial scales. Building on this capability, we introduce two modules: Global Verdict Refinement (GVR) uses image-level decisions to refine the overall anomaly score, while Local Verdict Refinement (LVR) aggregates regional decisions to refine the anomaly map. Combined with a simple DINOv2-based detector, VerdictAD achieves competitive performance in one-shot and few-shot anomaly detection. More importantly, as a plug-and-play framework, it consistently improves existing anomaly detectors across MVTec AD, VisA, and MVTec LOCO AD, achieving strong results across one-shot and few-shot settings. Our code and experimental logs are publicly available at https://github.com/sgstq0811/VerdictAD-Jev-AD.
Data-driven mechanistic hypotheses are essential to scientific discovery because they explain how underlying processes produce observed phenomena. AI agents and AI scientists increasingly support scientific data analysis. However, their ability to turn empirical findings into mechanistic hypotheses remains insufficiently examined. To address this gap, we introduce MechHypoBench, the first benchmark for evaluating whether AI agents and AI scientists can generate such hypotheses from empirical data. It combines paper-derived mechanisms from 14 scientific fields with real-world datasets containing 17.98 million records. The construction retains the observational complexity of empirical data while providing a specified underlying mechanism. Agents analyze the observations and propose open-form hypotheses. We develop an evaluation framework that assesses open-form mechanistic hypotheses through their consequences under withheld conditions. Experiments with general agents and AI scientists reveal a substantial gap between generated hypotheses and the underlying mechanisms.
Jun Liu, Zhenglun Kong, Changdi Yang, Fan Yang, Tianqi Li, Peiyan Dong, Joannah Nanjekye, Hao Tang 等 15 位作者
Large language model (LLM) agents are increasingly organized into multi-agent systems for complex reasoning and decision-making tasks. However, scaling such systems can substantially increase inference cost, especially when multiple agents repeatedly process overlapping or role-irrelevant context. Many existing multi-agent frameworks focus on orchestration and role specialization, but still rely on full-context sharing or static context routing, leading to redundant token consumption, increased runtime, and inefficient memory utilization. We present \emph{Role-Conditioned Context Routing (RCR)}, an execution-time layer that allocates role-relevant shared-memory items to agents under token budgets. Given a structured shared memory and a set of role-specialized agents, RCR selectively routes task-relevant memory items to each agent according to its role, optional task-stage metadata, and token budget. RCR integrates role-conditioned relevance scoring, token-budgeted memory selection, and iterative memory updates to reduce redundant context exposure during multi-agent execution. Experiments on multi-hop QA, embodied reasoning (ALFWorld), and web-based decision-making (WebShop) show that RCR reduces token usage by 35\% on average and runtime by 38\% on average relative to Full-Context, while maintaining or improving task performance.
For $k$ vertex-disjoint paths with prescribed endpoints in a directed acyclic graph (DAG), the $O(mn^{k-1})$-time pebbling algorithm of Fortune, Hopcroft, and Wyllie [TCS 1980] has remained the fastest for every $k\geq3$ for over four decades, even when the paths must be shortest. We break this bound for every $k\geq3$ with randomized algorithms running in expected $O_k(m+n^{k-3+ω+\varepsilon})$ time for every $\varepsilon>0$, where $ω<2.372$ is the matrix multiplication exponent; this is $O_k(m+n^{k-0.62})$. For two disjoint shortest paths, we find the paths as fast as we detect them: in expected linear time on DAGs and on undirected graphs with polynomially bounded integer weights, and in expected $O(\min\{mn,n^{ω+\varepsilon}\})$ time on directed graphs whose cycles have positive weight, which on dense graphs is $o(mn)$ even for detection. All algorithms evaluate a polynomial whose monomials are the solutions: a new first-collision circuit speeds up the evaluation in DAGs, and a correlated extraction step at cuts crossed once by every path finds a solution at the cost of a constant number of evaluations.
Building general game agents requires both high-level task reasoning and low-latency action control in rapidly changing visual environments. These capabilities favor models with substantially different computational profiles, making them difficult to realize within a single model. In this work, we present TaskFrame, a hierarchical planner-controller framework that connects task-level planning with frame-level keyboard-and-mouse control. A vision-language task planner decomposes a high-level goal into an ordered sequence of atomic instructions and is optimized through supervised learning and dependency-aware GRPO. The hidden states of special goal tokens that terminate each atomic instruction provide continuous goal embeddings that are cached and reused across control steps. A causal Transformer controller combines the embedding of the active goal with recent observations to predict discrete keyboard and mouse-button actions and continuous camera motion at every game step. We train the system in three stages: planner training, controller pretraining, and joint alignment. Experiments in Minecraft show that TaskFrame is competitive with general-purpose vision-language models on instruction decomposition, and that in closed-loop evaluation the agent achieves a higher combat success rate than the evaluated baselines.
Visual foundation models learn rich pathological morphology but lack explicit concept-level supervision. Vision--language pretraining can provide this supervision but requires costly image--text pairs. Here, we introduce GloSAGE, a semantic self-distillation framework that uses lesion profiles predicted by a pretrained visual teacher to construct structured pathology descriptions. Reliability-guided selection identifies image--text anchors, while Knowledge-Structured Pathology Text Synthesis and Evidence-Guided Negative-pair Construction account for partially observed lesions. Image--text alignment trains projection layers while retaining the frozen visual encoder; an intermediate Semantic Transfer Layer supplies features for downstream tasks. Developed on GloMillion and evaluated across seven glomerulus- and case-level cohorts, GloSAGE supported zero-shot lesion recognition, image--text retrieval, hard-negative discrimination and concept-intervention analyses. It outperformed its source visual model in most lesion-recognition tasks and improved lesion grading, while supporting competitive case-level disease classification, injury grading and baseline renal function estimation. Ablations supported the combined contributions of knowledge-structured text and evidence-guided negatives, and the intermediate transfer layer improved mean lesion-recognition ROC-AUC relative to the complete projection output. A frozen language model conditioned on GloSAGE features generated glomerular descriptions. These findings suggest that teacher-derived lesion profiles can provide annotation-efficient semantic supervision for extending a glomerular visual foundation model.
Qiaoling Chen, Jiaxiang Cheng, Xueqing Deng, Tiankai Hang, Ji Hou, Jiaqi Li, Yifan Liu, Tao Lu 等 38 位作者
Interactive world generation requires models to follow successive commands while preserving relevant history and action consequences. We present Hy-Nomas, a video world model with an 8B diffusion backbone that supports free-form text instructions and continuous camera control across multiple interaction turns. It generates variable-length segments conditioned on current inputs and cached history. Our central design aligns structured video captions with action-aware causal attention: descriptive context remains accessible across segments, while action instructions are restricted to their intended temporal windows. A vision-language Prompt Enhancer (PE) translates user requests and multimodal interaction histories into structured conditions for the next segment. A multi-stage curriculum progressively introduces causal generation, text interaction, camera control, and training on model-generated histories, followed by diffusion distillation. In human evaluation, Hy-Nomas shows superior performance over existing world models. We release model weights, code and benchmarks, together with a lightweight adaptation workflow, to support reproducible research on interactive world generation.
Ye Li, Jiahe Feng, Yuan Meng, Kangye Ji, Chen Tang, Jiajun Fan, Xinwan Wen, Jiawei Li 等 12 位作者
Diffusion Policy (DP) excels in embodied control but suffers from high inference latency and computational cost due to multiple iterative denoising steps. Existing diffusion model (DMs) acceleration methods improve inference speed by reducing computational overhead, as DMs for image generation are computation-bound. In this work, we reveal that DP are I/O bound, and thus propose using speculative decoding, a parallel computing framework, to accelerate DP for the first time. However, it is non-trivial to address the following challenges: (1) How to preserve the denoising quality of the base model at reduced computational cost under time-varying task difficulty in embodied settings; and (2) How to dynamically allocate computation according to task difficulty in such environments. Therefore, we propose Temporal-aware Reinforcement-based Speculative Diffusion Policy (TS-DP), the first framework that enables speculative decoding for DP with temporal adaptivity. First, to handle dynamic environments where task difficulty varies over time, we distill a Transformer-based drafter to imitate the base model and replace its costly denoising calls. Second, an RL-based scheduler further adapts to the time-varying task difficulty by adjusting speculative parameters to maintain accuracy while improving efficiency. Extensive experiments show that TS-DP achieves up to a 4.17x NFE reduction and a 3.2x wall-clock speedup with a draft acceptance rate of over 85% in simulation. Real-world experiments further demonstrate a 3.8x NFE reduction while improving success rates on real-time tasks (e.g., a 23% gain on the conveyor-belt task), reaching an inference frequency of 26 Hz and enabling real-time control while maintaining comparable task performance. The code is available at https://github.com/ChildTang/TS-DP.
The transition towards sixth-generation (6G) wireless networks necessitates autonomous orchestration mechanisms capable of translating high-level operational intents into executable network configurations. Existing approaches to Intent-Based Networking (IBN) rely upon either rule-based systems that struggle with linguistic variation or end-to-end neural models that lack interpretability and fail to enforce operational constraints. This paper presents a hierarchical multi-agent framework where Large Language Model (LLM) based agents autonomously decompose natural language intents, consult domain-specific specialists, and synthesise technically feasible network slice configurations through iterative reasoning-acting (ReAct) cycles. The proposed architecture employs an orchestrator agent coordinating two specialist agents, namely Radio Access Network (RAN) and Core Network agents, via ReAct-style reasoning, grounded in structured network state representations. Experimental evaluation across diverse benchmark scenarios demonstrates that the proposed system outperforms monolithic agents, rule-based systems, and direct LLM prompting, with architectural principles applicable to Open RAN (O-RAN) deployments. The results also demonstrate that whilst contemporary LLMs possess substantial general telecommunications knowledge, structured prompt engineering encoding domain-specific decision thresholds and operational policies is essential for reliable network automation performance.
Mobile network operators are increasingly exploring the use of artificial intelligence (AI) to automate network tasks such as cell selection and mobility management. Current assurance methods for these AI functions are forms of output testing. They check what a function decides, but they cannot show why it decided that way, and two functions can pass the same tests for different reasons. In this paper, we propose mechanistic auditing, which inspects the internal features of an AI function and checks them against machine-verifiable 3GPP specifications. We set out a three-step auditing principle that locates protocol-relevant features, verifies whether decisions depend on them, and audits each decision against them. We then present an audit-native network architecture. In this architecture, a dedicated verification agent certifies each AI function before deployment and audits its decisions during operation. In a case study on a language model for LTE signalling, we find internal features that carry the 3GPP handover concepts. Removing them lowers the model's accuracy on handover tasks from 96% to 0%, while removing the same number of random features changes its accuracy by less than two points. An audit of a handover decision can therefore check whether these features are active when the model makes that decision. We also discuss the open challenges that remain before mechanistic auditing can enter telecommunications practice and standardisation.
Ye Li, Haoran Ding, Huanan Liu, Yuan Meng, Yutian Shen, Yuansong Wang, Kangye Ji, Hanyun Cui 等 13 位作者
GUI agents built on vision-language models interact with continuously evolving interfaces, yet conventional inference repeatedly recomputes representations of largely overlapping observations and histories. This mismatch makes multi-step interaction costly, while changes in context, interface content, and action arguments prevent straightforward cache reuse. We introduce GUIFlow, a framework for full-pipeline incremental inference in multi-image GUI interaction. Our central insight is that advancing a decision need not require rebuilding all supporting representations: each interaction step can instead selectively update computational states maintained across the trajectory. To support new decisions as the context evolves, GUIFlow adapts how past computation is reused: historical KV states are position-aligned and injected into the current context; change signals guide selective updates of current-observation states in the vision encoder and language model; and earlier action responses serve as speculative drafts verified by the target model. Together, these mechanisms enable cross-step reuse throughout visual encoding, prefill, and decoding without additional training, a separate draft model, or changes to the model-visible inputs and action interface. Experiments demonstrate up to 3.39x geometric-mean speedup across offline multi-image benchmarks and up to 3.49x in online, closed-loop interaction on OSWorld, while maintaining task performance broadly comparable to the dense baseline.
Low-Rank Adaptation (LoRA) has emerged as the dominant paradigm for parameter-efficient fine-tuning of large language models. Standard all-layer LoRA applies adaptation throughout the transformer, despite heterogeneous contributions from its components. In this work, we conduct a systematic mechanistic interpretability analysis to understand how LoRA affects model internals. We observe that LoRA's adaptation effect is highly non-uniform, with significant component-wise and layer-wise sparsity. Most critically, we find that important layers are highly task-dependent, with limited overlap across tasks. Motivated by these insights, we propose Adaptive-LoRA, a calibration-driven method that automatically identifies task-specific layers via frozen-base zero-ablation, enabling efficient layer selection without exhaustive search. Extensive experiments on Qwen2.5-1.5B across 11 diverse reasoning tasks demonstrate that Adaptive-LoRA outperforms full-layer LoRA while using only 30% of layers, achieving a $3.5\times$ reduction in adapter parameters. This work bridges mechanistic interpretability and practical efficiency, providing both fundamental insights and a principled approach for adaptive fine-tuning.
Mingyu Wang, Bin Yu, Wenjian Lu, Zhi Wang, Gao Kefeng, Cheng Wen Xi'an University of Electronic Science and Technology, Xu Lu Xidian University, Cong Tian
Linux kernel device drivers are tightly coupled with hardware, making them difficult to execute and test without physical devices. This heavily limits automated code analysis and vulnerability discovery. While manual modeling is unscalable, Large Language Models (LLMs) offer a new approach to scale virtual device construction across the Linux driver ecosystem. In this paper, we present DevGen, an LLM-powered tool that generates QEMU-based virtual devices directly from Linux driver source code. DevGen combines static analysis to gather necessary context, guides the LLM through step-by-step prompting, and uses an automated self-correction loop driven by compilation and execution feedback. To further reduce errors, similar fixes are retrieved from a library of common modeling failures and incorporated into the repair prompt, which supports more targeted corrections in later iterations. The generated devices finally integrate with QEMU and Syzkaller, enabling driver fuzzing without physical hardware. DevGen is evaluated on 50 PCI/PCIe drivers from Linux 6.18 using three mainstream LLMs, and successfully generates usable models for 44 drivers. In these drivers, 24% of them achieve significant improvements in fuzzing coverage, and 7 previously unknown crashes are triggered with 1 CVE assigned. These results demonstrate the practical capability of LLMs to automate complex, system-level code generation tasks.
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 3: System Demonstrations) · 2026 · 正式发表
Enzhi Wang, Jiaming Zhou, Yuhang Jia, Aobo Kong, Qicheng Li Nankai University, Yong Qin Nankai University
Recent advances in speech large language models (e.g., GPT-4o) have enabled end-to-end spoken interactions, yet their robustness in real-world applications remains unclear, where systems must assist users in completing specific tasks under complex conditions such as multi-turn, ambiguous, and often spontaneous speech, as well as natural alternation between speech and text. Task-oriented dialogue (TOD) offers a realistic scenario to evaluate whether models can effectively help users accomplish such task-oriented goals, but existing benchmarks are mainly text-based, and the few speech datasets are limited to English and often neglect spontaneous disfluencies and speaker diversity. To address this gap, we introduce RealTalk-CN, the first Chinese multi-turn, multi-domain speech–text TOD dataset, containing 5.4k dialogues (60K turns, ~150 hours) of real human-to-human recordings with detailed annotations for dialogue states, disfluency types, and speaker characteristics. Based on this dataset, we propose a cross-modal interaction task supporting dynamic speech-text switching and a comprehensive evaluation protocol assessing robustness to disfluencies, sensitivity to speaker variation, and cross-domain generalization. Experiments on state-of-the-art models demonstrate the challenges posed by RealTalk-CN and establish its value as a benchmark for developing reliable and fair Speech LLMs in real-world deployments. The dataset and evaluation framework are available to encourage further 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
Binghai Wang, Yantao Liu Qwen Team, Alibaba Group, Yuxuan Liu Qwen Team, Alibaba Group, Tianyi Tang Alibaba Group, Shenzhi Wang Alibaba Group, Chang Gao Alibaba Group, Chujie Zheng Alibaba Group, Yichang Zhang Alibaba Group 等 16 位作者
Generative Reward Models (GenRMs) and LLM-as-a-Judge exhibit deceptive alignment by producing correct judgments for incorrect reasons, as they are trained and evaluated to prioritize Outcome Accuracy, which undermines their ability to generalize during RLHF. We introduce Rationale Consistency, a fine-grained metric that quantifies the alignment between the model’s reasoning process and human judgment. Our evaluation of frontier models reveals that rationale consistency effectively discriminates among state-of-the-art models and detects deceptive alignment, while outcome accuracy falls short in both respects. To mitigate this gap, we introduce a hybrid signal that combines rationale consistency with outcome accuracy for GenRM training. Our training method achieves state-of-the-art performance on RM-Bench (87.1%) and JudgeBench (82%), surpassing outcome-only baselines by an average of 5%. Using RM during RLHF, our method effectively improves performance as demonstrated on Arena Hard v2, notably yielding a 7% improvement in creative writing tasks. Further analysis confirms that our method escapes the deceptive alignment trap, effectively reversing the decline in rationale consistency observed in outcome-only training.
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
Dongyi Lv, Qiuyu Ding Alibaba Group, Heng-Da Xu, Zhaoxu Sun Alibaba Group, Zhi Wang, Feng Xiong, Mu Xu
Generative recommendation with large language models (LLMs) reframes prediction as sequence generation, yet existing LLM-based recommenders remain limited in leveraging geographic signals that are crucial in mobility and local-services scenarios. Here, we present Reasoning Over Space (ROS), a framework that utilizes geography as a vital decision variable within the reasoning process. ROS introduces a Hierarchical Spatial Semantic ID (SID) that discretizes coarse-to-fine locality and POI semantics into compositional tokens, and endows LLM with a three-stage Mobility Chain-of-Thought (CoT) paradigm that models user personality, constructs an intent-aligned candidate space, and performs locality informed pruning. We further align the model with real world geography via spatial-guided Reinforcement Learning (RL). Experiments on three widely used location-based social network (LBSN) datasets show that ROS achieves over 10% relative gains in hit rate over strongest LLM-based baselines and improves cross-city transfer, despite using a smaller backbone model.
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
Lin Zhong, Renjin Zhu, Shujuan Ma, Jinhao Cui, Lingzhi Wang Harbin Institute of Technology (Shenzhen), Hao Chen City University of Macau, Qing Liao Harbin Institute of Technology and Pengcheng Laboratory
Emotional Support Conversation (ESC) plays a critical role in mental health assistance by providing accessible psychological support in real-world applications. Large Language Models (LLMs) have shown strong empathetic abilities in ESC tasks. Yet, existing methods overlook the issue of cognitive distortions in help-seekers’ expressions. As a result, current models can only provide basic emotional comfort, rather than helping help-seekers address their psychological distress at a deeper cognitive level. To address this challenge, we construct the CogBiasESC dataset, the first dataset that expands existing ESC datasets by adding labels for cognitive distortions, includes their type, intensity, and safe risk level. Furthermore, we propose the Cognitive Policy-driven Large Language Model framework (CoPoLLM) to enhance LLMs’ ability to diagnose and intervene cognitive distortions in help-seekers. We also analyze the safety advantages of CoPoLLM from a theoretical perspective. Experimental results show that CoPoLLM significantly outperforms 15 state-of-the-art baselines in terms of distortion diagnosis accuracy, intervention strategy effectiveness, and safety risk control. Our source code is available at: https://github.com/Chips98/CoPoLLM-for-ACL-2026.
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
Hao Peng Tsinghua University, Tsinghua University, Yunjia Qi Tsinghua University, Xiaozhi Wang Tsinghua University, Zijun Yao, Lei Hou Tsinghua University, Juanzi Li
Reward models (RMs) are crucial for the training of large language models (LLMs), yet they typically rely on large-scale human-annotated preference pairs. With the widespread deployment of LLMs, in-the-wild interactions have emerged as a rich source of implicit reward signals. This raises the question: Can we develop reward models directly from in-the-wild interactions? In this work, we explore this possibility by adopting WildChat as an interaction source and proposing a pipeline to extract reliable human feedback, yielding 186k high-quality instances for training WildReward via ordinal regression directly on user feedback without preference pairs. Extensive experiments demonstrate that WildReward achieves comparable or even superior performance compared to conventional reward models, with improved calibration and cross-sample consistency. We also observe that WildReward benefits directly from user diversity, where more users yield stronger reward models. Finally, we apply WildReward to online DPO training and observe significant improvements across various downstream tasks. We will release our code, data, and models 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
AI 速览：ToolACE提出一种自动化的智能体流水线，用于生成准确、复杂且多样的工具学习数据，以提升大语言模型的函数调用能力。
Function calling significantly extends the application boundary of large language models, where high-quality and diverse training data is critical for unlocking this capability. However, real function-calling data is quite challenging to collect and annotate, while synthetic data generated by existing pipelines tends to lack coverage and accuracy. In this paper, we present ToolACE, an automatic agentic pipeline designed to generate accurate, complex, and diverse tool-learning data. ToolACE leverages a novel self-evolution synthesis process to curate a comprehensive API pool of 26,507 diverse APIs. Dialogs are further generated through the interplay among multiple agents, guided by a formalized thinking process. To ensure data accuracy, we implement a dual-layer verification system combining rule-based and model-based checks. We demonstrate that models trained on our synthesized data, even with only 8B parameters, achieve state-of-the-art performance on the Berkeley Function-Calling Leaderboard, rivaling the latest GPT-4 models. Our model and a subset of the data are publicly available at https://huggingface.co/Team-ACE.