AI 速览：作者提出首个面向多模态有状态交互的大规模多轮对话数据集 YANAMI-400K 与评测基准 YANAMI-Bench，并报告微调 BAGEL 后在图像编辑、生成与有状态交互上均有提升。
Unified Multimodal Models (UMMs) integrate visual understanding, generation, and editing within a single unified framework. However, existing unified multimodal datasets and benchmarks remain dominated by single-turn tasks. Even in recent multi-turn datasets, many turns can be solved independently without tracking information across turns. Consequently, they provide limited supervision and evaluation for multimodal stateful interaction, which requires models to understand, maintain, update, and faithfully realize evolving dialogue states across turns. To address this gap, we introduce YANAMI, the first large-scale multimodal dialogue dataset and benchmark for stateful interaction, comprising YANAMI-400K and YANAMI-Bench. YANAMI-400K contains 400K multi-turn multimodal dialogues, with more than 2M images and 2.6M dialogue turns. The dialogues span four representative scenarios: Multi-Image Fusion, Image Editing, World Modeling, and Visual Reasoning, and require models to continuously track, preserve, and revise evolving multimodal dialogue states across turns. YANAMI-Bench comprises 400 multi-turn dialogues for systematically evaluating multimodal stateful interaction. It jointly evaluates Dialogue State Understanding and Visual Realization, characterizing the gap between models' ability to understand evolving dialogue states and their ability to faithfully realize them in visual outputs. Experiments show that fine-tuning BAGEL on YANAMI-400K improves image editing, generation, and multimodal stateful interaction by 10.3%, 2.0%, and 13.23%, respectively, while preserving its general visual understanding capabilities. Evaluations on YANAMI-Bench further reveal that existing UMMs still struggle to maintain, update, and visually realize evolving multimodal dialogue states.
Junfeng Wang, Jiawei Liu, Yongchao Xu, Tao Jiang, Jiangbo Ai, Jin Zhang, Chong Wang, Qixing Zhang
Complex visual question answering often requires vision-language models to think with images, i.e., actively acquire local visual evidence by zooming into informative regions of high-resolution images. Our error analysis shows that many failures in this setting originate from incorrect observation actions, where the model fails to localize key regions or obtain crops that contain sufficient evidence for answering. We further observe that bounding-box tokens exhibit substantially higher generation entropy than ordinary language tokens, indicating high uncertainty when the model decides where to look. However, existing reinforcement learning methods based on final-answer rewards provide only trajectory-level feedback, making it difficult to assign precise credit to the observation actions. We propose Observation Policy Optimization (OPO), an RL algorithm for optimizing observation actions in tool-augmented thinking with images. OPO consists of two core modules. First, Uncertainty-Guided Observation Branching (UOB) treats each image zoom-in tool call as an observation event and selectively branches at high-uncertainty, low-redundancy events during rollout, enabling efficient exploration of alternative cropping decisions and their resulting visual evidence. Second, Evidence-Localized Advantage Attribution (ELAA) compares sibling branches using local evidence-sufficiency rewards and assigns credit only to the corresponding observation-action tokens. By decoupling observation-action optimization from trajectory-level outcome supervision, OPO improves visual evidence acquisition and further enhances visual reasoning performance. Experiments show that OPO consistently improves Qwen3-VL models across scales and achieves competitive performance against larger models and tool-augmented baselines. Code is publicly available.
Zhuyi Zhao, Shengjie Yuan, Shengyu Wang, Kang Wei, Wang Yang, Wenchao Xu, Fushuo Huo
Single-pass scoring models such as Jev give each code function a vulnerability probability, so that security teams can rank many functions for review without generating written explanations. Training an open-weight model for this task is limited by data. Benchmarks with both vulnerable and non-vulnerable labels keep few vulnerable functions once examples related to the test data are removed, while larger public vulnerability records mostly contain vulnerable functions only. Mixing the two lets the data source predict the label, and training on the labeled data alone leaves a model that often scores a fixed function above its vulnerable original. We present Cyber-JEV, a fine-tuning method that uses fix commits to turn these records into comparison groups. Each group holds a vulnerable function, its patched version and up to two unchanged functions from the same file, so the model sees the change that removed a recorded defect. One shared model learns two tasks: classifying functions from the labeled reference data, and ranking the vulnerable function above the other members of its group. At test time the input is one function and the output is a vulnerability score. On a test benchmark of 7,270 functions, an ensemble of five fine-tuned Qwen3.5-4B models reaches an area under the receiver operating characteristic curve of 0.814, compared with 0.743 for Jev 1.13. With the 4,031 alerts of Jev as the review budget, it misses 16 of 322 vulnerable functions instead of 39. Compared with training on the reference data alone, it ranks the vulnerable version above the patched version 14.8 to 29.2 percentage points more often on three sets of real fixes. Such scores can order the changed functions of a pull request or a release for review under a fixed budget.
Huihan Wang, Juncheng Wang, Yuxiang Feng, Wenlong Hou, Guoqi Yu, Chao Xu, Yang Liu, Baigui Sun 等 9 位作者
Missing modality imputation aims to synthesize unobserved modalities from observed ones. Existing methods treat each modality as living on its own latent manifold, forcing the generator to learn a long, unstructured jump between them. We present SphereFlow, which takes a fundamentally different view: by projecting all modalities into a shared hyperspherical latent space via a frozen self-supervised encoder, cross-modal imputation reduces to short-range geometric transport between nearby points on the sphere. Realizing this idea with flow matching, however, exposes two geometric conflicts: the standard Gaussian noise source is catastrophically mismatched with the unit-norm data manifold, and the linear interpolation path departs the sphere surface into a region where the decoder has no training signal. We resolve both with simple, geometry-aware modifications: replacing the noise source with the observed-modality latent so that the velocity field learns only the cross-modal displacement, and training the decoder with a time-weighted loss that extends its domain to the near-surface region traversed by the flow. We prove that this data-as-source formulation reduces the worst-case off-manifold deviation from a quantity that grows with the latent dimension to a small, dimension-independent bound governed by inter-modality similarity. Experiments on BraTS multi-modal MRI imputation and CT-MRI translation demonstrate state-of-the-art generation fidelity across all settings, while downstream brain tumor segmentation with our imputed data closes roughly eighty percent of the gap to the full-modality oracle. Code is available at https://github.com/WTHH031230/SphereFlow.
All languages are equal; when it comes to tokenization, some are more equal than others. Tokens are the hidden currency that dictates the cost and latency of contemporary LLM access. However, many languages written in non-Latin scripts observe a poor exchange rate: LLMs take several multiples of tokens to encode the same information in many languages as they do for English. We observe that this issue, known as token overfragmentation, persists in modern open-weight LLMs. The standard remedy is vocabulary expansion that adds target language items missing from the model's vocabulary. We focus on two core decisions in the vocabulary expansion process: What items should we add? and How should we initialize their corresponding input and output embeddings? We comprehensively study and advance the new research direction of interpretability-based vocabulary expansion along these two core decisions, making four key contributions: First, we question the conventional use of frequency-based methods to choose candidate vocabulary items to add (a decision long treated as settled), and show that interpretability-based methods offer a superior performance-token efficiency tradeoff. Second, we strengthen the case for interpretability-based embedding initialization by showing large gains (~20 pts) over baseline initialization methods for several languages written in non-Latin scripts. Third, we identify the phenomenon of "subword detokenization," where models progressively merge fragmented subword tokens into larger subwords across layers. Fourth, we propose FragMend, a method that leverages this phenomenon to push the efficiency ceiling of interpretability-based expansion, and validate its effectiveness.
A poor start lowers the ceiling, but does not negate the value of finishing well. In autoregressive video temporal grounding, models systematically exhibit an endpoint performance asymmetry, predicting ends significantly worse than starts. A committed flawed start acts as an immutable physical constraint, irreversibly capping the attainable overlap as a sunk loss. While models should minimize the remaining conditional regret, standard post-training paradigms fail to teach this adaptation. Specifically, teacher-forced SFT suffers from strict prefix exposure bias, and holistic trajectory evaluation in standard GRPO inadvertently penalizes conditionally optimal ends if they follow a poor start. To address this, we introduce FinishStrong, a novel two-stage post-training framework explicitly designed to teach optimal continuations. First, Online-start SFT exposes the end prediction to the model's own generated contexts. Second, Conditional-tree GRPO restructures the sampling topology, centering end rewards exclusively within the same-start branch to mathematically isolate the conditional end quality from the inherited start penalty. FinishStrong-9B achieves a state-of-the-art 61.7\% overall IoU on TimeLens-Bench, remarkably outperforming flagship closed-source models like Gemini-3.8-Flash and Qwen3.8-Max. Crucially, when forced to continue from imperfect starts, it reduces the subsequent end prediction error by 36.5\%, proving its exceptional ability to end well after flawed starts. Code will be open.
For a time-homogeneous Markov process, the fixed-lag solution operator \( (T_τf)(x)=\mathbb E[f(X_{t+τ})\mid X_t=x] \) of the Kolmogorov backward equation evolves observables linearly and deterministically, even when the dynamics are stochastic. We approximate this operator using a two-stage method where we first approximate this operator directly from data, without knowledge of the governing equation (e.g., Fourier Neural Operator), and then compress the Galerkin projection onto a finite-dimensional observable space that is fixed in advance or selected variationally. The error analysis traces the learning error of the estimator, the sampling error of the Galerkin inner products, and the invariance and reconstruction properties of the observable space to the computed spectra and forecasts. The sampling error enters only through a multiplicative factor. Near an isolated spectral cluster, the invariance defect of a subspace controls its distance from the target invariant subspace. The forecast error is controlled by the one-step error of the reduced model and the projection error of the observable; when the measure is invariant and the reduced model preserves constants and is stable on mean-zero observables, the bound is uniform in the horizon. For deterministic dynamics, a learning rate for Fourier neural operators (FNOs) gives one explicit bound for the learning error. Numerical experiments illustrate spectral approximation and forecasting, including long-rollout stability and a comparison with extended dynamic mode decomposition (EDMD) when fewer transition pairs than basis functions are observed.
Event cameras offering high dynamic range and low latency have emerged as disruptive technologies in imaging. Despite growing research on leveraging these benefits for different imaging tasks, a comprehensive study of recently advances and challenges are still lacking. This limits the broader understanding of how to utilize events in universal imaging applications. In this survey, we first introduce a physical model and the characteristics of different event sensors as the foundation. Following this, we highlight the advancement and interaction of image/video enhancement tasks with events. Additionally, we explore advanced tasks, which capture richer light information with events, \eg~light field estimation, multi-view generation, and photometric. Finally, we discuss new challenges and open questions offering a perspective for this rapidly evolving field. More continuously updated resources are at this link: https://github.com/yunfanLu/Awesome-Event-Imaging
Haocheng Mei, Yuchuan Tian, Ye Yuan, Chao Xu, Allen Chen, Kai Han, Yunhe Wang
Commercial language model APIs return only text, allowing a provider to cut costs by quietly routing a fraction of user requests to a cheaper substitute model while billing for the advertised one. Detecting such partial substitution is challenging because honest providers also alter the output distribution through undocumented decoding and serving choices, such as temperature scaling, sampling truncation, repetition and presence penalties, and default system prompts, forcing an audit to test against a composite null of legitimate configurations rather than a single reference. Our key insight is that these legitimate variations primarily shift response typicality across the entire batch, whereas partial substitution injects individual responses that remain anomalous under the claimed model. Building on this, we introduce API-TAIL, a typicality audit that first absorbs batch-level shifts by aligning a single effective temperature to center standardized surprisal scores across the audited batch, and then flags the endpoint if the number of remaining tail responses exceeds a threshold calibrated across legitimate configurations. Evaluated on real-world API responses and constructed substitution mixtures, API-TAIL controls false alarms across diverse honest providers and detects low-rate partial substitution with a small query budget. Under a common composite calibration, it outperforms the adapted rank- and MMD-based audit baselines.
Haocheng Mei, Yuchuan Tian, Ye Yuan, Enhua Wu, Chao Xu, Allen Chen, Kai Han, Yunhe Wang
JEV decision models pair a language model backbone with a lightweight scoring head to select among natural-language options. Direct selection compares candidate answers within one request, whereas entailment reranking evaluates each candidate in a separate natural language inference (NLI) request over fixed relation options. We examine whether the readout and adaptation choices used for direct selection remain effective for cross-request reranking. We combine a mean-span readout with selective backbone adaptation. Mean-span pooling replaces the option-ending pointer with an average of the contextualized states inside each NLI option, without adding head parameters. Selective adaptation retains low-rank updates in attention and gated DeltaNet projections while removing feed-forward adapters. On MC11, which averages accuracy over eleven reranking configurations from seven sources, selective adaptation improves NeoHorse-1-4B accuracy over the all-module baseline while cutting trainable parameters by more than half, and mean-span pooling yields further consistent gains at the same parameter count.
Linear attention and state space models (SSMs) offer efficient alternatives to softmax attention by maintaining a fixed-size recurrent state, but this design requires careful management to preserve useful historical associations and limit interference. The delta rule helps reduce interference through correction of the value associated with each key, but it does not explicitly account for how related value features are distributed across the state. We observe that a state $\boldsymbol{S}\in\mathbb{R}^{d_k\times d_v}$ can be viewed as a bank of $d_k$ addresses storing $d_v$-dimensional values. While a key-side read $\boldsymbol{S}^\top\boldsymbol{k}$ retrieves the value associated with $\boldsymbol{k}$, a value-side read $\boldsymbol{S}\boldsymbol{e}$ aggregates addresses whose stored value features match the selector $\boldsymbol{e}$. Based on this observation, we propose Dual DeltaNet (DDN), which extends Kimi Delta Attention (KDA) with a two-sided delta rule. The key-side delta rule reads with the current key to correct its associated value, while the complementary value-side delta rule reads with selected features of the current value to reorganize the addresses it occupies. We derive a chunkwise parallel training algorithm and a rank-2 recurrence that increases arithmetic intensity during inference. Compared with KDA, GDN-2, and Mamba-3, DDN achieves the strongest overall performance across language modeling, zero-shot downstream tasks, and retrieval, while incurring modest training and inference overhead relative to KDA. Code: https://github.com/ReichtumQian/DualDeltaNet
Rigidity and global rigidity are fundamental graph properties in discrete geometry, with applications to structural engineering and network localization. In two dimensions, both admit combinatorial characterizations in terms of the rigidity matroid and graph connectivity. We study the computational complexity of finding minimum-cost sets of edges or vertices whose deletion destroys either property in a finite simple graph. With unit costs, we give deterministic polynomial-time algorithms for edge deletion destroying global rigidity and vertex deletion destroying either property. We prove that edge deletion destroying rigidity is NP-hard and that, with positive integer costs, all four problems are NP-hard. These results resolve, in dimension two, the open questions on recognizing robustness under edge and vertex deletions posed by Jordán and Whiteley (2017), as well as the minimum edge-deletion problem for rigidity recorded in EGRES Open.
Automatic judges offer inexpensive information about model outputs, but an evaluation team must decide how to turn that information into fewer human annotations. Scores can guide sampling, correct an estimate, or support adaptive acquisition, and these uses may capture overlapping gains. We compare their incremental value on fixed evaluation pools from three preference datasets, and 19 attributed-question-answering systems. Matched comparisons use the same scores, account for pilot labels and budget rules, and separate acquisition from estimation. Score stratification is a strong reference: after ten-bin stratification, the best fixed local linear correction offers only 1.00--5.27% further MSE reduction in the primary configurations. Unequal allocation offers larger gains in some pools, but can increase error when confident scores conceal variable human outcomes. Comparison on AttributedQA, shrinking learned quotas toward score-only quotas lowers pooled MSE for 17 of 19 systems relative to the score-only procedure, but ordinary pooled 95% intervals cover only 88.09--93.54%. The same shrinkage improves on the score-only reference in only one of eight preference configurations. Ranking experiments further show that lower individual MSE need not improve system selection. These findings support a simple starting procedure, proportional sampling within score strata followed by a stratified human mean, and identify what additional evidence is needed to justify correction or adaptation for the intended evaluation output.
Bimanual manipulation with dexterous hands has attracted growing attention in research on robotic action models. However, few studies have developed benchmarks specifically tailored for dexterous manipulation, leaving the dexterous manipulation capabilities of existing models insufficiently characterized. In this paper, we present BiDex-100, a simulation benchmark for bimanual dexterous manipulation that combines broad task coverage, multi-dimensional evaluation, and high-DoF hand embodiments. Concretely, the benchmark comprises 106 tasks organized along five capability axes: generalization, instruction following, limited demonstrations, long horizon, and memory. Each task is paired with an out-of-domain variant featuring controlled visual and spatial changes, allowing systematic assessment of robustness. In addition, the benchmark supports two embodiments and provides 8,080 success-validated teleoperated demonstrations, totaling 6.25 million frames. We evaluate seven recent Vision-Language-Action models under a unified protocol, establishing baselines for capability-specific performance, robustness, and data efficiency. The website is available at https://robbyant-research.github.io/BiDex-100/.
Multimodal Emotion Recognition (MER) systems often suffer from missing modalities in real-world scenarios. Existing methods usually generate, align, or distill missing modalities as a whole, overlooking the heterogeneous nature of the information carried by each modality. Such holistic treatment mixes inferable shared semantics with uncertain modality-specific details, yielding unstable representations and degrading robustness. To address this issue, we propose the Primitive Memory Distillation (PriMD) framework. Unlike existing methods, PriMD takes an intra-modal perspective and focuses on how different types of information within a modality differ in recoverability. PriMD first disentangles cross-modal shared semantics from modality-specific representations, and then discretizes the latter into learnable semantic primitives to construct modality-specific memory banks. When modalities are missing, PriMD is a teacher-student framework that the student model uses the shared semantics of available modalities as queries to dynamically retrieve primitives. It compensates for missing modality-specific information within a constrained memory space and aligns with the teacher model. Extensive experiments on IEMOCAP, CMU-MOSI, and CMU-MOSEI demonstrate that PriMD achieves state-of-the-art performance and consistently stronger robustness across a wide range of missing-modality settings, while mitigating the instability caused by holistic feature inference. Our code and project website are available at https://www.jiaqizhang-sengoku.com/Code/PriMD and https://www.jiaqizhang-sengoku.com/PriMD/ , respectively.
In September 2026, a new large language model was released. Soon after, a new way of making explainer videos became very popular: people asked the model to write a program, and the program produced the video. "One-click video" content then appeared in large amounts on many platforms. On the surface, this is a rise in creative power at the level of the individual. We argue that, at a deeper level, it changes the structure of the content industry. This judgment rests on two observations. First, this wave of content is service-type content. People watch it to meet a need, not to enjoy the content itself. The other kind is work-type content. AI makes both kinds faster to produce. But anyone can come in and make service-type content, while work-type content is limited by the number of people who have something to say. So the ratio between the two slowly becomes unbalanced. Second, free platforms choose content through layers of trial viewing, and this uses up user attention. The total amount of attention is fixed. The unbalanced ratio also dilutes good content. A simplified mathematical model shows that, as long as these conditions hold, the recommendation pool mechanism will fail once new content grows past a certain multiple. Based on this, we discuss one possible way out. Machines and entry rules judge whether service-type content passes the bar. Trusted people judge work-type content. We call this the BoLe system. At the end, we briefly imagine how the two kinds of content may develop from here.
Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, Linke Ouyang, Zhiyuan Zhao, Tao Chu, Tianyao He 等 59 位作者
We introduce MinerU2.5, a 1.2B-parameter document parsing vision-language model that achieves state-of-the-art recognition accuracy while maintaining exceptional computational efficiency. Our approach employs a coarse-to-fine, two-stage parsing strategy that decouples global layout analysis from local content recognition. In the first stage, the model performs efficient layout analysis on downsampled images to identify structural elements, circumventing the computational overhead of processing high-resolution inputs. In the second stage, guided by the global layout, it performs targeted content recognition on native-resolution crops extracted from the original image, preserving fine-grained details in dense text, complex formulas, and tables. To support this strategy, we developed a comprehensive data engine that generates diverse, large-scale training corpora for both pretraining and fine-tuning. Ultimately, MinerU2.5 demonstrates strong document parsing ability, achieving state-of-the-art performance on multiple benchmarks, surpassing both general-purpose and domain-specific models across various recognition tasks, while maintaining significantly lower computational overhead.
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 6: Industry Track) · 2026 · 正式发表
Embodied agents in open-ended environments such as Minecraft increasingly adopt planner–controller architectures, with large language models acting as high-level planners. While planning has advanced rapidly, control remains underexplored. Existing systems commonly rely on a monolithic policy to execute subgoals across varying contexts, forcing incompatible behaviors into a shared parameter space and causing interference that scaling only partially mitigates. To address this, we propose MoEC, a Memory-Routed Mixture-of-Experts Controller for Adaptive Minecraft Control. MoEC routes via a subgoal-indexed, non-parametric expert memory and regulates capacity through failure-triggered expert growth and redundancy-aware consolidation. This design enables continual adaptation without full retraining, while maintaining parameter efficiency and with bounded inference cost. We evaluate MoEC on diverse and compositional Minecraft tasks, demonstrating significant gains in adaptability, robustness, and execution consistency over strong baselines, yielding a scalable and efficient alternative for open-ended control.
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
Chao Xue, Yao Wang, Mengqiao Liu, Di Liang, Xingsheng Han, Peiyang Liu, Xianjie Wu Beijing Information Science & Technology University, Chenyao Lu Tencent AI Lab 等 14 位作者
Supervised Fine-Tuning (SFT) is the standard approach for adapting large language models (LLMs) to downstream tasks. However, we observe a persistent failure mode: even after convergence, models often fail to correctly reproduce a subset of their own supervised training data. We refer to this behavior as the Incomplete Learning Phenomenon (ILP). This paper presents the first systematic study of ILP in LLM fine-tuning. We formalize ILP as post-training failure to internalize supervised instances and demonstrate its prevalence across multiple model families, domains, and datasets. Through controlled analyses, we identify five recurrent sources of incomplete learning: (1) missing prerequisite knowledge in the pre-trained model, (2) conflicts between SFT supervision and pre-training knowledge, (3) internal inconsistencies within SFT data, (4) left-side forgetting during sequential fine-tuning, and (5) insufficient optimization for rare or complex patterns. We introduce a diagnostic-first framework that maps unlearned samples to these causes using observable training and inference signals, and study several targeted mitigation strategies as causal interventions. Experiments on Qwen, LLaMA, and OLMo2 show that incomplete learning is widespread and heterogeneous, and that improvements in aggregate metrics can mask persistent unlearned subsets. The findings highlight the need for fine-grained diagnosis of what supervised fine-tuning fails to learn, and why.
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
Zekai Lin University of Glasgow, Chao Xue, Di Liang, Xingsheng Han, Peiyang Liu, Xianjie Wu Beijing Information Science & Technology University, Lei Jiang Tencent AI Lab, Yu Lu 等 11 位作者
Supervised Fine-Tuning (SFT) of large language models often suffers from task interference and catastrophic forgetting. Recent approaches alleviate this issue by isolating task-critical parameters during training. However, these methods represent a static solution to a dynamic problem, assuming that parameter importance remains fixed once identified. In this work, we empirically demonstrate that parameter importance exhibits temporal drift over the course of training. To address this, we propose Evolving Parameter Isolation (EPI), a fine-tuning framework that adapts isolation decisions based on online estimates of parameter importance. Instead of freezing a fixed subset of parameters, EPI periodically updates isolation masks using gradient-based signals, enabling the model to protect emerging task-critical parameters while releasing outdated ones to recover plasticity. Experiments on diverse multi-task benchmarks demonstrate that EPI consistently reduces interference and forgetting compared to static isolation and standard fine-tuning, while improving overall generalization. Our analysis highlights the necessity of synchronizing isolation mechanisms with the evolving dynamics of learning diverse abilities.
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
Jiongxiao Wang University of Wisconsin - Madison, Qiaojing Yan Amazon, Yawei Wang Amazon, Yijun Tian, Soumya Smruti Mishra Amazon, Zhichao Xu Amazon, Megha Gandhi Amazon, Panpan Xu Amazon 等 9 位作者
Large Language Model (LLM)-based agents have demonstrated remarkable capabilities in complex reasoning and multi-turn interactions but struggle to continuously improve and adapt when deployed in new environments. One promising approach is implementing skill libraries that allow agents to learn, validate, and apply new skills. However, current skill library approaches rely primarily on LLM prompting, making consistent skill library implementation challenging. To overcome these challenges, we propose a Reinforcement Learning (RL)-based approach to enhance agents’ self-improvement capabilities with a skill library. Specifically, we introduce Skill Augmented GRPO for self-Evolution (SAGE), a novel RL framework that systematically incorporates skills into learning. The framework’s key component, Sequential Rollout, iteratively deploys agents across a chain of similar tasks for each rollout. As agents navigate through the task chain, skills generated from previous tasks accumulate in the library and become available for subsequent tasks. Additionally, the framework enhances skill generation and utilization through a Skill-integrated Reward that complements the original outcome-based rewards. Experimental results on AppWorld demonstrate that SAGE, when applied to supervised-finetuned model with expert experience, achieves 8.9% higher Scenario Goal Completion while requiring 26% fewer interaction steps and generating 59% fewer tokens, substantially outperforming existing approaches in both accuracy and efficiency. Our code is available at https://github.com/amazon-science/SAGE.
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