Shijue Huang, Hangyu Guo, Guanting Dong, Chenxin Li, Junting Lu, Xinyu Geng, Zhaochen Su, Zhenyu Li 等 11 位作者
Multimodal deep search requires an agent to solve open-world problems by chaining search, tool use, and visual reasoning over evolving textual and visual context. Two bottlenecks limit current systems. First, existing tool-use harnesses treat images returned by search, browsing, or transformation as transient outputs, so intermediate visual evidence cannot be re-consumed by later tools. Second, training data is usually built by fixed curation recipes that cannot track the target agent's evolving capability. To address these challenges, we first introduce a visual-native agent harness centered on an image bank reference protocol, which registers every tool-returned image as an addressable <image:N> reference and makes intermediate visual evidence reusable by later tools. On top of this harness, On-policy Data Evolution (ODE) runs a closed-loop data generator that refines itself across rounds from rollouts of the policy being trained. This per-round refinement makes each round's data target what the current policy still needs to learn. The same framework supports both diverse supervised fine-tuning data and policy-aware reinforcement learning data curation, covering the full training lifecycle of the target agent. Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average. At 30B, ODE raises the average score from 30.6% to 41.5%. Further analyses validate the effectiveness of image bank reuse, especially on complex tasks requiring iterative visual refinement, while rollout-feedback evolution yields more grounded SFT traces and better policy-matched RL tasks than static synthesis.
Jiaxuan Li, Yao Liu, Yuewei Xue, Shijie Dong, Zixiang Jin, Xiaoyu Guo, Qiao Liu
Multivariate probabilistic forecasting requires modeling cross-variable dependence alongside marginal uncertainty. The challenge is to learn this dependence while controlling marginal noise scales and avoiding dense covariance costs. We propose Factor-Correlated Diffusion (FaCoDiff), which couples a compact conditional statistical estimator with structured denoising. History-conditioned factor loadings use distinct coefficients for forecast-center correction, scale adjustment and noise dependence. Diagonal compensation preserves the estimated marginal noise variances, while low-rank operators enable residual Gaussian likelihood evaluation and noise transport without dense covariance storage or factorization. Factor-derived feature modulation explicitly conditions the denoiser, and a geometry-aware objective accounts for errors after noise transport. The same conditional geometry governs forward perturbations and reverse sampling. Across nine datasets, FaCoDiff achieves the lowest mean variogram score on eight datasets and ranks among the top two on 43 of 54 dataset-metric combinations in the reported comparisons. Matched ablations on Solar and Weather show that factor covariance improves mean Energy and Variogram scores under all four combinations of center and scale correction. Disrupting cross-variable sample pairings while preserving marginal trajectories worsens joint and aggregate scores on both datasets, demonstrating the predictive value of generated cross-variable dependence.
As agents become more capable, they are moving beyond one-off tasks toward sustained assistance across successive user goals. Such assistance requires agents to discover how unfamiliar environments work, reuse what they learn across tasks, and keep that experience current as environments evolve. Existing benchmarks increasingly study experience reuse and adaptation, but provide limited insight into which parts of accumulated experience agents revise, retain, or lose. To address this gap, we introduce ReScopeBench, a benchmark for the experience lifecycle across 18 interactive environments. It distinguishes structural experience about how an environment is organized and accessed from rule experience about reusable principles and constraints that guide decisions across tasks. To study how experience is acquired, reused, and revised, we propose an experience-first construction approach. It builds task streams around shared experience and introduces silent environmental changes. We evaluate task performance and track changes in saved experience throughout each stream. Across 20 models, we find that retained experience improves task performance, but agents struggle to keep it current as environments change. Even when the required experience is correctly recorded, agents do not always complete tasks successfully. These findings identify selective maintenance and effective use as complementary challenges for long-lived agents.
Linwei Dong, Ruoyu Guo, Ge Bai, Quan Zheng, Yawei Luo, Changqing Zou
Diffusion distillation, exemplified by Distribution Matching Distillation (DMD), has shown great promise in few-step generation but often sacrifices quality for sampling speed. While integrating Reinforcement Learning (RL) into distillation offers potential, a naive fusion of these two objectives relies on suboptimal raw sample evaluation. This sample-based scoring creates inherent conflicts with the distillation trajectory and produces unreliable rewards due to the noisy nature of early-stage generation. To overcome these limitations, we propose GDMD, a novel framework that redefines the reward mechanism by prioritizing distillation gradients over raw pixel outputs as the primary signal for optimization. By reinterpreting the DMD gradients as implicit target tensors, our framework enables existing reward models to directly evaluate the quality of distillation updates. This gradient-level guidance functions as an adaptive weighting that synchronizes the RL policy with the distillation objective, effectively neutralizing optimization divergence. Empirical results show that GDMD sets a new SOTA for few-step generation. Specifically, our 4-step models outperform the quality of their multi-step teacher and substantially exceed previous DMDR results in GenEval and human-preference metrics, exhibiting strong scalability potential.
Shenghao Ye, Yuxiang Wang, Yu Guo, Tao Zhang, Shuangwu Chen, Zijian Wen, Qihui Zhu, Mengjie Zhang 等 11 位作者
Large Reasoning Models (LRMs) achieve strong performance on table reasoning tasks but incur substantial inference cost due to long reasoning traces. Stepwise model routing mitigates this issue by dynamically assigning reasoning steps to smaller or larger models. However, stepwise model routing for table reasoning remains underexplored. Through empirical analysis, we find that reasoning steps involving tables contain two types of tokens with distinct uncertainty distributions: table tokens grounded in table structure, such as cell values and headers, and text tokens representing surrounding natural-language reasoning. The uncertainty of both token types is correlated with the risk that the model makes an error in the next reasoning step. However, existing methods fail to model them separately, leading to suboptimal routing decisions. To address this, we propose EcoTab, a table-aware stepwise routing framework for efficient table reasoning. At each reasoning step, EcoTab separately estimates the uncertainties of table tokens and text tokens, maps them to next-step failure risks for the small model, and combines the two risks for routing. Experiments on multiple table reasoning benchmarks show that EcoTab consistently outperforms strong baselines and achieves a better balance between accuracy and efficiency.
Shenghao Ye, Yu Guo, Zhengheng Li, Qiang Huang, Shuangwu Chen, Wang Luo, Qian Rui, Jie Zhang 等 13 位作者
Stepwise model routing improves the efficiency of Large Reasoning Models (LRMs) by assigning each reasoning step to a suitable model. Recent methods formulate routing as a sequential decision process and train the router with reinforcement learning. However, although they model routing as a process, they still supervise the router with outcome rewards. Such rewards only reflect final answer correctness and fail to evaluate intermediate routing decisions, which can weaken performance and generalization. To address this gap, we propose RoRo, a rubric-guided process reward framework for stepwise model routing. RoRo first collects diverse routing trajectories and constructs preference pairs based on outcome, cost, and process quality. It then trains a Rubricor to generate a query-specific evaluation rubric and a Judge to score routing trajectories under this rubric through alternating optimization. The resulting process rewards are combined with outcome rewards to optimize the routing policy via GRPO. Experiments on five reasoning benchmarks under both same-family and cross-family settings show that RoRo consistently outperforms strong baselines and achieves better accuracy and cost trade-offs.
Fine-grained action annotations can support skill reuse and compositional generalization in robot policies. However, robot datasets typically provide task-level instructions without subtask annotations. Although some include subtask labels, inconsistent formats and conventions hinder cross-source joint training. We propose RoboAtom-Data, which adds unified, time-aligned atomic-action annotations to existing data from multiple sources, covering 1.45 million trajectories, 6.88 million skill-labeled segments, and 16,882.1 hours. Our pipeline combines a unified protocol, large-scale manual annotation, propagation across repeated executions with the same ordered skill sequence, and multi-stage quality control. We evaluate the dataset through automatic atomic-action annotation and robot control. We construct RoboAtom-Bench to standardize evaluation of skill recognition and temporal boundary localization. Trained on the annotated data, RoboAtom-Anno achieves better automatic annotation performance than all evaluated open- and closed-source models on both RoboAtom-Bench and the external WGO-Bench. For robot control, atomic-action supervision paired with hierarchical execution improves task success over task-level baselines on LIBERO-Long, LIBERO-Plus, and RoboTwin. We will release the data, benchmark, model, and construction pipeline to support further research on skill-level robot learning.
As preferences, goals, and facts change, LLM agents must use the current state while earlier versions remain in context. Yet they can answer with an old value of the same variable, a failure that we call stale binding. To study when models use outdated information and why, we introduce Controlled In-Context Memory (CICM), a benchmark for tracking and using updated information in conversations and agent logs. We observe that even frontier reasoning models can fail to recover the current state. We find that in open-source models probes can still recover the updated value when the model answers with an old one, pointing to a failure to select information that remains available. Component tests in Qwen and Pythia identify a mechanism for this selection failure: attention drift, where attention favors old values over the current one when producing an answer. We study a one-layer transformer to mathematically understand how this phenomenon happens: when attention scores are similar, several old values can together receive more attention than the current value. Guided by this explanation, we redirect attention toward the current value without further training. When the current value is requested directly, adjusting this intervention for each input corrects most old-value errors across various model families while preserving nearly all initially correct answers. Reliable context management therefore requires more than remembering updated information: models must use it to guide their answers.
Wenhao Liang, Liangwei Nathan Zheng, Lin Yue, Wei Emma Zhang, Mingyu Guo, Weitong Chen
Post-hoc class activation maps (CAMs) are a standard tool for inspecting the evidence behind an image classifier's predictions, yet nothing in ordinary training encourages these maps to be spatially appropriate. We study whether inexpensive spatial supervision can improve a classifier's own predicted-class Grad-CAM without ever optimizing an attribution map. Box-Guided Evidence Routing (BGER) trains a lightweight gate on the final feature map under box or mask supervision and routes classification through the gated features, while Grad-CAM is computed separately at the pre-gate representation, so the evaluated map never enters the training objective. With a BCE routing loss, BGER raises MaxBoxAccV2 from $0.584$ to $0.715$ on CUB-200-2011 and from $0.757$ to $0.832$ on Stanford Dogs at comparable accuracy. Matched controls attribute most of the ResNet-50 gain to the spatial supervision reshaping the backbone rather than to routing itself: when classification bypasses the gate, most of the improvement remains, and detaching gradients through the gate leaves the ResNet-50 result nearly unchanged. The same detachment preserves most of the gain in two DenseNet-121 chest X-ray settings but removes the apparent gain on Swin-T, and directly supervising the CAM reaches stronger localization at a larger accuracy cost. Overall, spatial supervision can improve separately evaluated post-hoc CAMs, but both the mechanism and the size of the benefit depend on the architecture and the evaluation setting.
We give a deterministic quasipolynomial-time construction of Reed-Solomon codes that are list-decodable beyond the Johnson radius. Specifically, for every $ε\in(0,1]$ and all sufficiently large $n$, we construct an $[n,k]$ Reed-Solomon code of rate $R=Ω(ε/(1+\log(1/ε)))$ over a field of size $(nk)^{quasi-poly(nk)}$ that is $(1-ε,O(1/ε))$ list-decodable. The construction runs in quasi-poly(nk) time. We also give a deterministic construction of $[n,k]$-MDS$(3)$ Reed-Solomon codes for all integers $1\leq k<n$, with the same asymptotic bounds on field size and construction time. In the constant-rate regime, our field-size bound substantially improves the bound of $O(n^{k^{2k}})$ achieved by Roth's construction (Roth, IEEE TIT 2022). Our constructions exploit a connection to polynomial identity testing for read-once determinants, together with the quasipolynomial isolating-weight construction of Gurjar and Thierauf for linear matroid intersection (Gurjar and Thierauf, STOC 2017 and Computational Complexity 2020).
Deformable image registration relies on accurate anatomical correspondences. However, explicitly coupling local matching reliability with spatial context selection and deformation update strength remains insufficiently explored in learning-based registration. This challenge is particularly pronounced in homogeneous, weakly textured, and severely deformed regions, where ambiguous matches may induce erroneous deformation updates. We propose ReliMorph, a reliability-routed symmetric framework for 3D deformable registration that couples correspondence assessment with spatial context selection and update strength.Local correlation distributions characterize correspondence ambiguity using shared anatomical features with modality-specific adaptation. These reliability cues guide spatial expert fusion and gate residual velocities during coarse-to-fine refinement. Updated inverse-inconsistency feedback guides routing and gating at finer pyramid levels. To our knowledge, ReliMorph is the first 3D deformable registration framework to jointly condition spatial expert routing and residual velocity gating on correspondence reliability. Extensive experiments demonstrate that ReliMorph consistently outperforms competing methods in anatomical overlap, highlighting its effectiveness across unimodal and multimodal registration tasks.
Vision-Language-Action (VLA) models remain brittle in long-horizon, contact-rich manipulation because success-only imitation provides little supervision for execution drift, while failed rollouts are often discarded. We introduce RePO-VLA, a recovery-driven policy optimization framework that assigns distinct roles to success, recovery, and failure trajectories. RePO-VLA first applies Recovery-Aware Initialization (RAI), slicing each corrective suffix into a new episode so corrective actions depend on the current adverse state rather than the preceding failure. It then learns a Progress-Aware Semantic Value Function (PAS-VF), aligning spatiotemporal trajectory features with instructions and successful references. The resulting labels salvage useful failure prefixes via reliability decay, while low-value labels mark drift and terminal breakdowns, teaching differences among nominal, failed, and corrective actions. The data engine turns adverse states into planner-generated or human-collected corrective rollouts, teaching recovery to the success manifold. Value-Conditioned Refinement (VCR) trains the policy to prefer high-progress actions. At deployment, a fixed high-value token biases actions toward the learned success manifold without online failure detectors or heuristic retries. We introduce the Failure-Recovery Benchmark (FRBench), with standardized error injection and recovery-focused evaluation. Across simulated and real-world bimanual tasks, RePO-VLA improves robustness, raising adversarial success from 20% to 75% on average and up to 80% in scaled real-world trials. Code and video results are available at https://1018weijia.github.io/RePO-VLA-Web/.
Multimodal intelligence can be viewed as human-guided co-evolution between worlds and agents through interaction. However, scaling this process faces two bottlenecks: constructing heterogeneous, verifiable environments is costly, and manual Harness iteration cannot keep pace with growing task diversity and complexity. A key insight is that tasks with different modalities and surface forms can share the same underlying capability requirements. We therefore introduce The Trinity, a recursive self-improvement framework coupling world generation with Harness evolution through three roles: Creator, Explorer, and Evolver. Drawing on verified execution trajectories, Evolver shifts the basis of task organization from surface categories to shared requirements, reorganizing the original task tree into a hierarchical capability tree. We introduce Capability-Driven Real2Sim2Real: for heterogeneous tasks at the same node, Creator abstracts their common requirements into simplified, verifiable simulated worlds for capability learning, while evaluation on original tasks measures transfer. Along this capability tree, Explorer progressively evolves reusable capability modules through selective inheritance, forming a Harness tree whose validated updates accumulate across iterations within a fixed topology. The resulting Harnesses specialize in different capabilities, motivating a Capability-Aware Harness Router that uses public task information to select compatible module configurations from the tree. Our evolved Harnesses improve mean performance by 34 points across 28 training benchmarks and 26 points across 18 unseen benchmarks, including two real-world embodied tasks.
Frontier multimodal large language models (MLLMs) can control robots without robot-specific training, but how their capabilities compare and interact with post-trained vision-language-action (VLA) and world action model (WAM) policies across manipulation demands remains to be thoroughly evaluated. In this report, we evaluate GPT-6-Astra on EBench as the core of an agent that controls a robot through an agent harness under the same visual and proprioceptive observation protocol as post-trained policies. We study two agent configurations: direct control, in which the agent generates waypoint commands directly, and harness with policy, in which it can also invoke a post-trained VLA or WAM policy for action execution. Under direct control, the agent achieves competitive overall performance relative to post-trained policies. It leads on short-horizon tasks and exhibits the least performance variation across generalization dimensions, whereas it underperforms post-trained policies on long-horizon tasks, tabletop manipulation, and high-precision contact-rich tasks. In the harness with policy, invoking a post-trained VLA or WAM combines the complementary strengths of both components, raising overall task success to the level of the strongest compared post-trained policies while substantially reducing token cost. However, performance remains below the per-task upper bound of the standalone components due to suboptimal policy invocation, fixed-instruction execution without subgoals, and grasp disruption when the agent resumes control. Overall, our study characterizes the capability boundary of a frontier MLLM agent and suggests directions for improving future agent harness design. Project page: https://internrobotics.github.io/EBench-Astra-Tech-Report/
Shenghao Ye University of Science and Technology of China, Yu Guo University of Science and Technology of China, Dong Jin, Yuxiang Wang, Yikai Shen University of Science and Technology of China, Yunpeng Hou, Shuangwu Chen, Jianyang University of Science and Technology of China 等 9 位作者
Table question answering (TableQA) is a fundamental task in natural language processing (NLP). The strong reasoning capabilities of large language models (LLMs) have brought significant advances in this field. However, as real-world applications involve increasingly complex questions and larger tables, substantial noisy data is introduced, which severely degrades reasoning performance. To address this challenge, we focus on improving two core capabilities: Relevance Filtering, which identifies and retains information truly relevant to reasoning, and Table Pruning, which reduces table size while preserving essential content. Based on these principles, we propose EnoTab, a dual denoising framework for complex questions and large-scale tables. Specifically, we first perform Evidence-based Question Denoising by decomposing the question into minimal semantic units and filtering out those irrelevant to answer reasoning based on consistency and usability criteria. Then, we propose Evidence Tree-guided Table Denoising, which constructs an explicit and transparent table pruning path to remove irrelevant data step by step. At each pruning step, we observe the intermediate state of the table and apply a post-order node rollback mechanism to handle abnormal table states, ultimately producing a highly reliable sub-table for final answer reasoning. Finally, extensive experiments show that EnoTab achieves outstanding performance on TableQA tasks with complex questions and large-scale tables, confirming its effectiveness.
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
Ziyun Zhang, Zezhou Wang, Xiaoyi Zhang Microsoft, Zongyu Guo Microsoft Research, Jiahao Li Microsoft Research Asia, Bin Li Microsoft, Yan Lu Microsoft Research Asia
GUI agents that interact with graphical interfaces on behalf of users are a promising direction for practical AI assistants, yet training them is hindered by scarce suitable environments. We present InfiniteWeb, a system that automatically generates functional web environments at scale for GUI agent training. While LLMs perform well on generating a single webpage, building a realistic and functional website with many interconnected pages faces challenges. We address these challenges through unified specification, task-centric test-driven development, and combining website seed variation with reference design images. Our system also generates verifiable task evaluators enabling dense reward signals for reinforcement learning. Experiments show that our system surpasses commercial coding agents at realistic website construction, and GUI agents trained on our generated environments achieve significant performance improvements on OSWorld and Online-Mind2Web, demonstrating the effectiveness of the proposed system.
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
Zhaokun Wang, Jinyu Guo University of Electronic Science and Technology of China, Jingwen Pu University of Electronic Science and Technology of China, Hongli Pu, Meng Yang, Xunlei Chen, Jie Ou, Wenyi Li 等 10 位作者
Large language models (LLMs) trained on unfiltered corpora inherently risk retaining sensitive information, necessitating selective knowledge unlearning for regulatory compliance and ethical safety. However, existing parameter-modifying methods face fundamental limitations: high computational costs, uncontrollable forgetting boundaries, and strict dependency on model weight access. These constraints render them impractical for closed-source models, yet current non-invasive alternatives remain unsystematic and reliant on empirical experience. To address these challenges, we propose the Controllable Alignment Prompting for Unlearning (CAP) framework, an end-to-end prompt-driven unlearning paradigm. CAP decouples unlearning into a learnable prompt optimization process via reinforcement learning, where a prompt generator collaborates with the LLM to suppress target knowledge while preserving general capabilities selectively. This approach enables reversible knowledge restoration through prompt revocation. Extensive experiments demonstrate that CAP achieves precise, controllable unlearning without updating model parameters, establishing a dynamic alignment mechanism that overcomes the transferability limitations of prior methods.
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
Vision-Language Models (VLMs) have demonstrated impressive capabilities in code generation across various domains. However, their ability to replicate complex, multi-panel visualizations from real-world data remains largely unassessed. To address this gap, we introduce RealChart2Code, a new large-scale benchmark with over 2,800 instances grounded in authentic datasets and featuring tasks with clear analytical intent. Crucially, it is the first benchmark to systematically evaluate chart generation from large-scale raw data and assess iterative code refinement in a multi-turn conversational setting. Our comprehensive evaluation of 14 leading VLMs on RealChart2Code reveals significant performance degradation compared to simpler benchmarks, highlighting their struggles with complex plot structures and authentic data. Our analysis uncovers a substantial performance gap between proprietary and open-weight models and confirms that even state-of-the-art VLMs often fail to accurately replicate intricate, multi-panel charts. These findings provide valuable insights into the current limitations of VLMs and guide future research directions.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Jie Ou, Jinyu Guo University of Electronic Science and Technology of China, Shiyao Guo, Yuang Li, Ruiqi Wu University of Electronic Science and Technology of China, Zhaokun Wang, Wenyi Li, Wenhong Tian
Many-Shot In-Context Learning (ICL) has emerged as a promising paradigm, leveraging extensive examples to unlock the reasoning potential of Large Language Models (LLMs). However, existing methods typically rely on a predetermined, fixed number of shots. This static approach often fails to adapt to the varying difficulty of different queries, leading to either insufficient context or interference from noise. Furthermore, the prohibitive computational and memory costs of long contexts severely limit Many-Shot’s feasibility. To address the above limitations, we propose AdapShot, which dynamically optimizes shot counts and leverages KV cache reuse for efficient inference. Specifically, we design a probe-based evaluation mechanism that utilizes output entropy to determine the optimal number of shots. To bypass the redundant prefilling computation during both the probing and inference phases, we incorporate a semantics-aware KV cache reuse strategy. Within this reuse strategy, to address positional encoding incompatibilities, we introduce a decoupling and re-encoding method that enables the flexible reordering of cached key-value pairs. Extensive experiments demonstrate that AdapShot achieves an average performance gain of \sim10% and a 4.64\times speedup compared to state-of-the-art DBSA.
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
Yu Guo University of Science and Technology of China, Shenghao Ye University of Science and Technology of China, Shuangwu Chen, Zijian Wen, Tao Zhang University of Science and Technology of China, Bai Qirui, Dong Jin, Yunpeng Hou 等 11 位作者
Table Question Answering (TableQA) benefits significantly from table pruning, which extracts compact sub-tables by eliminating redundant cells to streamline downstream reasoning. However, existing pruning methods typically rely on sequential revisions driven by unreliable critique signals, often failing to detect the loss of answer-critical data. To address this limitation, we propose TabTrim, a novel table pruning framework which transforms table pruning from sequential revisions to gold trajectory-supervised parallel search. TabTrim derives a gold pruning trajectory using the intermediate sub-tables in the execution process of gold SQL queries, and trains a pruner and a verifier to make the step-wise pruning result align with the gold pruning trajectory. During inference, TabTrim performs parallel search to explore multiple candidate pruning trajectories and identify the optimal sub-table. Extensive experiments demonstrate that TabTrim achieves state-of-the-art performance across diverse tabular reasoning tasks: TabTrim-8B reaches 73.5% average accuracy, outperforming the strongest baseline by 3.2%, including 79.4% on WikiTQ and 61.2% on TableBench.
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
Xiangfeng Wang University of Science and Technology of China, Hangyu Guo Stephens College, Yanlin Lai Tsinghua University, Mitt Huang Stepfun, Liang Zhao Megvii Technology Inc., Chengyuan Yao StepFun Inc, Yinmin Zhang StepFun, Qi Han 等 14 位作者
While model-based verifiers are essential for scaling Reinforcement Learning with Verifiable Rewards (RLVR), current outcome-centric verification paradigms primarily focus on the consistency between the final result and the ground truth, often neglecting potential errors in the derivation process. This leads to assigning positive rewards to correct answers produced from incorrect derivations. To bridge this gap, we introduce PRIME, a benchmark for evaluating verifiers on PRocess-outcome alignment verification In Mathematics and Engineering. Curated from a comprehensive collection of college-level STEM problems, PRIME comprises 2,530 high-difficulty samples through a consistency-based filtering pipeline. Through extensive evaluation, we find that current verifiers frequently fail to detect derivation flaws. Furthermore, we propose a process-aware RLVR training paradigm utilizing verifiers selected via PRIME. This approach substantially outperforms the outcome-only verification baseline, achieving absolute performance gains of 8.29%, 9.12%, and 7.31% on AIME24, AIME25, and Beyond-AIME, respectively, for the Qwen3-14B-Base model. Finally, we demonstrate a strong linear correlation (R^2 > 0.92) between verifier accuracy on PRIME and RLVR training effectiveness, validating PRIME as a reliable predictor for verifier selection.
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
Zijian Wen, Tao Zhang University of Science and Technology of China, Shuangwu Chen, Shenghao Ye University of Science and Technology of China, Yu Guo University of Science and Technology of China, Qirui Chen University of Science and Technology of China, Jingxian Shuai, Yunpeng Hou 等 10 位作者
Retrieval-Augmented Generation (RAG) significantly enhances LLMs but faces high prefill latency during long-context processing. While KV cache reuse can mitigate this, current methods relying on shallow features or static heuristics often fail to identify critical tokens for recomputation, resulting in generation quality degradation.We have an insight that KV deviations are more pronounced in deep layers.However, directly extracting deep-layer features from the target model is computationally prohibitive. Crucially, we find that the deep-layer features of a lightweight speculative model exhibit strong consistency with the target model in the selection of critical tokens for recomputation.In light of these insights, we propose SpecCache, which employs deep-layer hidden-state norms from a speculative model as a proxy to guide the critical token selection for target large model.Experiments demonstrate that SpecCache outperforms state-of-the-art (SOTA) baselines. Compared to full KV recomputation, it reduces time-to-first-token (TTFT) by 2.17-3.95\times and increases inference throughput by 2.7-5.2\times, with negligible degradation in generation quality relative to full recomputation.
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