Minjae Oh, Sangjun Song, Gyubin Choi, Yunho Choi, Yohan Jo
On-Policy Distillation (OPD) has emerged as a dominant post-training paradigm for Large Language Models, especially in reasoning domains. However, OPD remains unstable in practice due to the high gradient variance of its single-sample Monte Carlo estimator, and recipes for stable training are still immature. We propose vOPD (On-Policy Distillation with a control variate baseline), which casts OPD as policy-gradient RL and stabilizes it by introducing a control variate baseline---canonically a value function---from the RL literature. We show that the OPD value function admits a closed form: the per-token negative reverse KL divergence between student and teacher, available directly from the existing forward pass with no additional critic or inference. Existing stabilization methods either compute the full token-level reverse KL over the entire vocabulary, adding significant overhead, or restrict it to a top-k support, biasing the objective. vOPD instead preserves the lightweight single-sample estimator and subtracts the value function as a detached baseline to keep the gradient unbiased while reducing variance. Moreover, we show that a top-k approximation of the baseline further lowers cost without compromising performance. Across mathematics, science, and instruction-following benchmarks, vOPD consistently outperforms base OPD and matches the most expensive full-vocabulary method, offering an efficient stabilization of OPD through principled RL variance reduction.
Jun Nie, Yonggang Zhang, Jun Song, Qianshu Cai, Dahai Yu, Yike Guo, Xinmei Tian, Bo Han
The behavior of an LLM agent is determined not only by the underlying model, but also by its harness: the executable program that constructs context, invokes tools, verifies intermediate results, and recovers from failures. Existing approaches typically optimize harnesses on training or development data and freeze the resulting programs at test time, limiting adaptation to shifts in task distributions, failure modes, and tool interactions. We ask whether the harness can instead be optimized during evaluation using unlabeled execution traces. We introduce Test-Time Harness Evolution (TTHE), which treats the executable harness as the state of test-time adaptation. On each unlabeled test batch, TTHE iteratively evolves a population of candidate harnesses through agentic proposers that reason over execution traces and explore diverse edit directions. A judge selects a candidate using execution traces and execution-derived proxy signals. The selected harness determines the current batch's outputs and initializes adaptation on the next. Crucially, the solver, proposers, and judge operate through different harnesses around the same frozen LLM, without weight updates, access to final-answer labels, or a separately trained adaptation model. Across six execution-grounded benchmarks spanning text-to-SQL, competitive programming, software engineering, data-science coding, terminal-based tasks, and agentic tool use, TTHE improves fixed ReAct-style baseline harnesses through persistent, inspectable program adaptation. These results recast test-time adaptation for LLM agents as evolution over executable control programs and identify execution-derived proxy reliability as a central challenge for robust unsupervised agent improvement.
Hyeongyu Kim, Mingyeong Jang, Youngjun Song, GeonHui Han, Dosik Hwang
Test-time adaptation (TTA) updates a deployed model using only unlabeled target data, but standard entropy-based methods often rely on deterministic confidence and can reinforce overconfident errors under distribution shift. We propose MIRA, Mutual Information guided calibration for Reliable Test-Time Adaptation, a source-free online adaptation method that uses lightweight last-layer Gaussian perturbations to probe prediction reliability. For each target batch, MIRA reuses a single feature extraction pass and samples perturbed classifiers around the source classifier, forming an efficient local predictive distribution. Rather than fitting a Bayesian posterior, MIRA treats the perturbation scale as a local probing radius and selects a batch-specific scale using mutual information (MI) and prediction agreement. The resulting perturbation-sensitivity signal identifies predictions that are confident and locally stable under classifier perturbations, while suppressing brittle predictions with high stochastic disagreement. MIRA then performs adaptation with an MI-confidence weighted entropy loss, which down-weights perturbation-sensitive or low-confidence samples before they drive entropy minimization. This weighted loss reduces the influence of brittle predictions that can otherwise cause harmful adaptation under distribution shift. The method requires no source data, retraining, model ensembles, or offline posterior fitting, and can be applied as a plug-and-play modification to entropy-minimization TTA. Experiments on distribution-shift benchmarks show improvements in accuracy, calibration, and negative log-likelihood, suggesting that mutual-information-guided perturbation probing and weighted adaptation yield more robust online TTA than deterministic confidence.
LLM-based agentic systems are increasingly combined with CoT reasoning to solve tasks that require accurate reasoning, external knowledge, and multi-step deliberation. Yet, explicit CoT incurs substantial generation overhead, while implicit CoT, which replaces textual rationales with compact continuous tokens, does not naturally support the explicit communication interface required for agentic tool-use. We propose TAC-CoT, a tool-empowered adaptive implicit CoT framework that replaces explicit reasoning traces with compact latent reasoning segments while retaining textual tokens only for tool calls, observations, and final answers. This interleaved design enables agentic tool-use without verbalizing every intermediate reasoning step. Experiments on agentic math and QA benchmarks show that TAC-CoT improves accuracy by up to 8.8 points while reducing model forward passes by up to 6.8x. Furthermore, analyses show that TAC-CoT improves tool-use quality in both single- and multi-tool settings.
Edge SoCs combine CPUs, GPUs, and NPUs, yet runtimes for multimodal large language models (MLLMs) keep generation on one engine and leave others idle. Streaming workloads allow ingestion and answering to overlap, but backend-local KV state binds queries to the engine that built their history. Stage performance and co-running interference vary, so no single placement wins across SoCs and workloads. \name separates logical history from engine-specific KV state so any language engine can continue a stream. Stateful Streaming Generation overlaps ingestion and answering, while Concurrent Sequence Generation serves ready queries in parallel. Phase-level state continuation lets one engine prefill a query and another decode the answer, freeing the first to resume ingestion. State-Driven Execution Planning (SDEP) chooses paths by predicted completion time and sets each engine's share of queries using arrival rates and measured service times. Across M4 Max, RK3588, and QCS6490, \name beats the strongest single-engine baselines, even when they reuse context. On M4 Max, SDEP cuts live-stream p95 time to first token by $76\%$, with every first token within one second, and raises throughput under backlog by up to $35\%$ over Metal-only execution with context reuse. On two RK3588 streams, SDEP delivers $19.6\%$ higher request throughput than the best single engine with context reuse and $27.2\%$ more than parallel engines without shared state. Stateful Streaming Generation reaches up to $2.48\times$ serial throughput on the same engines. Every answer matches its serial reference.
Weihuang Lin, Xianing Chen, Zhibin Wang, Congzhi Zhang, Yiwei Ma, Jiayi Ji, Miao Chen, Jun Song 等 12 位作者
Despite significant progress, current models struggle to generate comprehensive video captions, constrained primarily by coarse-grained supervision, which often fails to capture spatial-temporal dynamics and the underlying logical intent. To address this issue, we introduce MVRCap, a novel framework that reformulates video captioning as a reinforcement learning task steered by structured verification. Central to our approach is a Fine-Grained Verifiable Reward mechanism, serving as dense supervision to enforce rigorous visual-semantic alignment. Specifically, we construct this reward system anchored by three complementary pillars: a Hierarchical Concept Coverage Reward utilizing our newly constructed MVRCap-QA dataset to ensure multi-dimensional completeness and semantic depth, a Factuality Verification Reward to enhance factual consistency, and a Conciseness Constraint to maximize semantic density. Furthermore, we curate the high-quality MVRCap-SFT dataset, providing structured annotations across six key dimensions, to serve as a robust foundational initialization for the RL policy in this training pipeline. Extensive experiments demonstrate that MVRCap achieves state-of-the-art performance on authoritative benchmarks including VDC, CAPability, and VC4VG.
Vision-language model-based mobile agents have gained the ability to understand complex instructions and mobile screenshots, benefiting from reinforcement learning paradigms like Group Relative Policy Optimization (GRPO). However, existing approaches centers on offline training or local action-level rewards often trap agents in local optima, hindering effective exploration and error correction with the environment. Crucially, we find that directly applying task-level rewards often leads to convergence difficulties due to the sparse nature of GUI interactions. To address these challenges, we present Mobile-R1, a systematic training recipe that bridges atomic action execution and strategic task completion. We propose a hierarchical curriculum consisting of three stages: (1) format alignment for reasoning structure, (2) on-policy exploration with verifiable action feedback to ground basic execution, and (3) multi-turn task-level training with realistic environment to unlock exploration and self-correction. This hierarchical strategy effectively bootstraps the agent, significantly enhancing its capability for exploration and self-correction (the “Eureka” moments). Furthermore, addressing the critical scarcity of diverse GUI data in non-English ecosystems, we contribute a comprehensive Chinese mobile dataset covering 28 applications with 24,521 high-quality manual annotations, and establish a rigorous benchmark with 500 trajectories. We will open source all resources, including the dataset, benchmark, model weight, and codes: https://mobile-r1.github.io/Mobile-R1/.
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
Recent advances in Vision-Language Models (VLMs) have enabled mobile agents to perceive and interact with real-world mobile environments based on human instructions. However, the current fully autonomous paradigm poses potential safety risks when model understanding or reasoning capabilities are insufficient. To address this challenge, we first introduce InquireBench, a comprehensive benchmark specifically designed to evaluate mobile agents’ capabilities in safe interaction and proactive inquiry with users, encompassing 5 categories and 22 sub-categories, where most existing VLM-based agents demonstrate near-zero performance. In this paper, we aim to develop an interactive system that actively seeks human confirmation at critical decision points. To achieve this, we propose InquireMobile, a novel model inspired by reinforcement learning, featuring a two-stage training strategy and an interactive pre-action reasoning mechanism. Finally, our model achieves an 46.8% improvement in inquiry success rate and the best overall success rate among existing baselines on InquireBench. The project page is available at https://bit-aqh.github.io/InquireMobile/homepage/.
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
Joeun Kim, Seunghyouk Yoon, Xuan-Bach Le, Youngeun Nam University of Toronto, Doyoung Kim Korea Advanced Institute of Science and Technology, Hwanjun Song Korea Advanced Institute of Science and Technology, Jae-Gil Lee Korea Advanced Institute of Science and Technology
Retrieval-augmented generation(RAG) systems depend on retrieval modules to supply grounding evidence for large language models. While hybrid approaches combining sparse and dense retrievers improve performance, most rely on fixed weights that ignore query-specific and corpus-specific variation. Similarly, query expansion has long been used to enrich recall, but its integration with original queries is usually static and can introduce noise. We present QuDAR, a dual-perspective adaptive retrieval framework that adapts along two perspectives: retriever type (sparse vs. dense) and query format (original vs.expanded). Leveraging margin-derived confidence (e.g., top-1–top-2 score gaps) and blind LLM-based relevance scoring, QuDAR dynamically assigns query-specific weights, fusing lexical specificity with semantic breadth while mitigating noise. QuDAR is lightweight, retriever-agnostic, and broadly applicable. Experiments show consistent gains over static baselines, improving overall retrieval quality and yielding more stable performance across queries.
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
Zhuo Li The Chinese University of Hong Kong, Shenzhen, Yupeng Zhang Alibaba Group, Pengyu Cheng Alibaba Group, Jiajun Song Beijing Institute for General Artificial Intelligence, Mengyu Zhou Alibaba Group, Hao Li, Shujie Hu, Yu Qin 等 11 位作者
Hallucination remains a critical bottleneck for large language models (LLMs), undermining their reliability in real-world applications, especially in Retrieval-Augmented Generation (RAG) systems. While existing hallucination detection methods employ LLM-as-a-judge to verify LLM outputs against retrieved evidence, they suffer from inherent confirmation bias, where the verifier inadvertently reproduces the errors of the original generation. To address this, we introduce Multi-Agent Reinforced self-Check for Hallucination (MARCH), a framework that enforces rigorous factual alignment by leveraging deliberate information asymmetry. MARCH orchestrates a collaborative pipeline of three specialized agents: a Solver, a Proposer, and a Checker. The Solver generates an initial RAG response, which the Proposer decomposes into claim-level verifiable atomic propositions. Crucially, the Checker validates these propositions against retrieved evidence in isolation, deprived of the Solver’s original output. This well-crafted information asymmetry scheme breaks the cycle of self-confirmation bias. By training this pipeline with multi-agent reinforcement learning (MARL), we enable the agents to co-evolve and optimize factual adherence. Extensive experiments across hallucination benchmarks demonstrate that MARCH substantially reduces hallucination rates. Notably, an 8B-parameter LLM equipped with MARCH achieves performance competitive with powerful closed-source models. MARCH paves a scalable path for factual self-improvement of LLMs through co-evolution. The code is at https://github.com/Qwen-Applications/MARCH.
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Sashuai Zhou Zhejiang University, Qiang Zhou, Jijin Hu Alibaba Group, Hanqing Yang, Yue Cao, Junpeng Ma, Yinchao Ma, Jun Song Alibaba Group 等 12 位作者
Despite impressive progress in high-fidelity image synthesis, generative models still struggle with logic-intensive instruction following, exposing a persistent reasoning–execution gap. Meanwhile, closed-source systems (e.g., Nano Banana) have demonstrated strong reasoning-driven image generation, highlighting a substantial gap to current open-source models. We argue that closing this gap requires not merely better visual generators, but executable reasoning: decomposing high-level intents into grounded, verifiable plans that directly steer the generative process. To this end, we propose Unified Thinker, a task-agnostic reasoning architecture for general image generation, designed as a unified planning core that can plug into diverse generators and workflows. Unified Thinker decouples a dedicated Thinker from the image Generator, enabling modular upgrades of reasoning without retraining the entire generative model. We further introduce a two-stage training paradigm: we first build a structured planning interface for the Thinker, then apply reinforcement learning to ground its policy in pixel-level feedback, encouraging plans that optimize visual correctness over textual plausibility. Extensive experiments on text-to-image generation and image editing show that Unified Thinker substantially improves image reasoning and generation quality.
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