Hallucination remains a persistent challenge for Multimodal Large Language Models (MLLMs), severely limiting their reliability in high-stakes applications. Existing evaluations, predominantly based on static benchmarks, suffer from narrow taxonomical coverage and rapid performance saturation, failing to reflect model robustness in evolving real-world scenarios. To bridge this gap, we present a systematic evaluation framework integrating a comprehensive benchmark with self-evolving stress testing. First, we introduce UniHall, a fine-grained dataset grounded in a unified taxonomy spanning Object, Instruction, and Knowledge dimensions. Second, to address benchmark saturation, we propose Self-Adaptive Multimodal Fuzzing (SAMF), a self-adaptive framework that employs evolutionary mutation strategies to explore the boundaries of model hallucinations. Crucially, to ensure reliable assessment of dynamic inputs, SAMF incorporates a structured metric suite driven by an ensemble of multi-modal oracles. Our extensive experiments reveal that state-of-the-art MLLMs exhibit significant performance degradation under fuzzing compared to conventional settings, exposing a dissociation between reasoning capabilities and factual grounding. Furthermore, we identify a helpfulness-hallucination trade-off, where reinforcement learning alignment inadvertently exacerbates sycophancy in instruction-following tasks. The framework, code and benchmark are available at https://github.com/LanceZPF/EvalHall.
Re-ranking aims to present users with an ordered list that maximizes sequence-level utility. Industrial deployments widely adopt the Generator--Evaluator framework, where the generator proposes candidate sequences, and the evaluator scores them to select the best one for display. The fidelity of the evaluator's utility estimates thus directly governs recommendation quality. However, existing evaluators predominantly follow the \textbf{Independent Behavior Modeling (IBM)} paradigm, which simplifies the continuous user \emph{browsing journey} into independent per-item predictions (e.g., CTR). This oversimplification ignores two physical realities: (i) a click can occur only after the item has been reached and impressed, and (ii) items compete for finite user attention. As a result, IBM produces systematically biased evaluations and suboptimal sequence selection. To address this, we propose a \textbf{Joint Behavior Modeling (JBM)} perspective that explicitly models the \emph{display} $\rightarrow$ \emph{impression} $\rightarrow$ \emph{click} process to capture causal and competitive dependencies within a sequence. Building on JBM, we develop the \textbf{Impression-Click Evaluator (ICE)}, a plug-and-play framework that (i) predicts a browsing-depth distribution to estimate per-position impression probabilities, (ii) leverages predicted impressions to suppress spurious context from unexposed items during click modeling, and (iii) introduces a click distribution matching objective to capture intra-sequence dependencies. Extensive offline experiments across diverse backbones, together with a 14-day online A/B test in the Taobao App (+0.70\% clicks, +3.67\% revenue) and full-traffic deployment, validate the effectiveness and robustness of our approach.
Memory-augmented assistants retrieve evidence from stored conversations and documents to answer queries. Most embedding-based retrievers compare single-vector representations of queries and memories, even when relevance depends on a small part of the content. Such local evidence can be matched to queries using the contextual token representations already computed by the same embedding models. Our method uses sparse token matching to improve retrieval without additional training. Each query token retrieves a small set of matching tokens across candidate memories. The matching scores are aggregated for each memory and combined with whole-memory similarity to rank memories. An optional low-rank query adapter can be trained with supervision while keeping the encoder and stored memory representations fixed. Our method improves retrieval across eight benchmarks and multiple model families without additional training. Query adaptation further increases the gains to 13.3 and 14.3 percentage points in average Recall@10 and NDCG@10 on Qwen3-Embedding-4B. These retrieval gains translate into higher question-answering accuracy on long-term memory benchmarks.
Memory retrieval aims to identify evidence from stored records to answer a query. Stored records with different factual details may share topics, entities, or events. Such shared contents would make records without the required evidence more likely to receive high similarity scores and outrank the correct evidence. In this paper, we use the information shared by these records to construct local references for query-memory matching. We propose NEighborhood-Aware Residuals (NEAR), a training-free framework that uses these references to construct residual representations that complement the original embeddings. Specifically, each memory derives a reference direction from its neighbors, and each query obtains its reference from a precomputed table. Using these references, NEAR constructs projection residuals for both queries and memories. These residuals are combined with the original embeddings for a single approximate nearest-neighbor search over a fixed memory index. Experiments on eight long-context retrieval benchmarks show consistent improvements across nine embedding models. Downstream evaluations show that NEAR improves question-answering performance.
Ge Fan, Kai Meng, Huiping Chu, Yuning Jiang, Han Zhu, Bo Zheng
Helping users discover new interests is vital to the long-term health of industrial recommender systems, yet the usual remedy, boosting exposure of under-explored (discovery) content, often backfires in production. We report URED (Unexposed Recycling for Efficient Discovery), a training-free recall-stage system serving Taobao feed recommendation at Alibaba for over six months to tens of millions of daily active users. It rests on a production-validated insight: naive boosting fails because it injects cold, information-poor candidates that rankers systematically over-estimate, buying no discovery gain while risking engagement. Items scored highly but never exposed (signals industrial pipelines routinely discard) form a pool that is both relevant and reliably predictable. Over a 30-day horizon URED aggregates them and re-injects at recall time through a lightweight pipeline spanning offline storage, distributed aggregation, and low-latency online serving, adding under 1% latency via a single O(1) lookup per request with no model retraining. Online A/B tests show +3.47% discovered categories and +8.47% content richness with engagement preserved: completion video views rise significantly (+2.62%) and total views move positive (+0.36%, not significant). Ablations confirm both design pillars are necessary: a one-day window collapses the discovery gain to +0.55% (not significant), and random injection without score filtering yields no gain while trending negative on engagement (-0.17%). Across production recall channels, URED reaches 13.9x the platform-average discovery rate (6.3x the strongest nearline baseline) at engagement quality on par with the platform average. We share deployment details and lessons for practitioners facing the exploration-exploitation dilemma under tight resource constraints.
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.
Jianghan Zhu, Cong Zhang, Ni Zhang, Yifan Yang, Yew Soon Ong, Zhiguang Cao
Automatic operator design (AOD) has increasingly adopted large language models (LLMs) to generate problem-specific search and variation operators for multi-objective combinatorial optimization (MOCO). Yet most methods evolve the operators while leaving the process that discovers them fixed. This limits adaptation as new Pareto trade-offs and operator interactions change where and how the search should proceed. We introduce HEMO, which makes the AOD harness itself a self-evolving optimization target. HEMO maintains a persistent harness state and updates one module at a time, revising the logic that determines which experience to retrieve, which operators to improve, which parents to select, and how to formulate the orchestration of search steps. Downstream solution quality and runtime guide these updates, while a discounted UCB scheduler directs adaptation toward promising modules as their utility changes. A persistent memory filesystem preserves candidate outcomes and diagnostic feedback across harness states, enabling successive states to build on accumulated optimization experience. Experiments across eight MOCO tasks and two structurally distinct evolutionary backbones demonstrate strong performance against conventional evolutionary, neural, and LLM-based AOD baselines, together with effective transfer to related tasks. HEMO extends automatic operator design beyond evolving individual operators to improving the search process that discovers them.
Rongguang Ye, Zhan Zhuang, Yichen Wu, Ming Tang, Kede Ma
Continually adapting large language models requires acquiring new knowledge while preserving previously learned capabilities. Although jointly adapting model parameters and task-specific soft prompts is promising, it has two key limitations: previously learned prompts may lose effectiveness as the model evolves, and relationships among prompts that could support cross-task transfer are not explicitly learned. We propose Learning an Anchored Prompt Space (LAPS), which preserves historical prompt effectiveness and learns relationships among task-specific soft prompts to facilitate positive transfer. Historical prompts are first aligned with the updated model through self-distillation. LAPS then constructs an anchored prompt space whose vertices correspond to learned task-specific soft prompts and whose intermediate geometry is shaped by learnable Bézier control prompts. Once this space is learned, LAPS identifies the best-performing prompt for each task on its validation set, allowing the selected prompt to draw on knowledge acquired from related tasks. Experiments on the TRACE benchmark across three Qwen3 model scales show that LAPS consistently outperforms existing distillation-based, prompt-based, and joint prompt-parameter adaptation methods, improving overall task performance while reducing forgetting.
Erhan Xu, Kai Ye, Hongyi Zhou, Luhan Zhu, Francesco Quinzan, Chengchun Shi
This paper studies reinforcement learning from human feedback (RLHF) for aligning large language models with human preferences. While RLHF has demonstrated promising results, many algorithms are highly sensitive to misspecifications in the underlying preference model (e.g., the Bradley-Terry model), the reference policy, or the reward function, resulting in undesirable fine-tuning. To address model misspecification, we propose a doubly robust preference optimization algorithm that remains consistent when either the preference model or the reference policy is correctly specified (without requiring both). Our proposal demonstrates superior and more robust performance than state-of-the-art algorithms, both in theory and in practice. The code is available at https://github.com/DRPO4LLM/DRPO4LLM
Skills provide agents with reusable experience to guide future decisions without modifying the underlying model. However, coding trajectories are often entangled with task- and repository-specific contexts, making it challenging to distill procedural skills that generalize across diverse coding environments. To address this challenge, we present Hierarchical Skill Evolution (HSE), which progressively distills procedural knowledge from coding trajectories through hierarchical evidence aggregation. HSE first extracts task-level evidence from individual trajectories, then aggregates recurring evidence within the same repository to form repository-specific skill candidates, and finally consolidates consistent patterns across repositories into transferable skills across repositories. The evolution is driven by an external writer-evaluator loop that generates and assesses skills while keeping the underlying coding model frozen. The loop externally updates the writer and evaluator policies to improve subsequent skill generation and assessment. This hierarchy progressively filters task-specific details while preserving procedures that generalize across diverse coding environments. Successful trajectories yield reusable guidance for future tasks, while failed trajectories reveal recurring failure modes that guide avoidance and recovery. We evaluate HSE on SWE-bench Verified and DeepSWE, using GLM-5.2 across Pi and Claude Code harnesses and GPT-5.5 and Opus 4.8 with Pi. Across both benchmarks, 10 skills evolved solely from SWE-Gym improve coding performance across all evaluated settings, demonstrating transfer across benchmarks, coding models, and agent harnesses. We further evolve the skill library from target-benchmark trajectories without verifier feedback, relying solely on the learned evaluator for candidate assessment and achieving additional gains through benchmark-tailored skills.
Luozheng Qin, Yuhan Zhu, Changhai Zhou, Qian Qiao, Jun Gao, Yuhua Zhou, Harry Zhang, Xiangyu Zeng 等 13 位作者
Spatial intelligence lets an agent perceive and act in the physical world, and many of its judgments must be made at interaction rates: which spatial relation holds, how far an object is, or where to place it. Vision-language models (VLMs) produce these answers by free-form text generation, paying for autoregressive decoding and answer parsing even though the output space is compact: one of a few options, a scalar, or a 2D point. In light of this, we introduce MJev-Spatial, a System 1 multimodal decision model that tackles all three types of spatial understanding tasks within a unified classification framework. Specifically, MJev-Spatial treats each candidate as a class and scores it using the backbone's pretrained query and key projections, without adding parameters. The candidates are the given options for choice tasks, or value bins and image regions refined through coarse-to-fine classification for numerical and point prediction, respectively. At inference, each classification decision takes a single forward pass and generates no answer tokens. To train on heterogeneous spatial data efficiently, we supervise all coarse-to-fine stages jointly with teacher forcing in one forward pass, and combine this with length-aware global batching and cross-GPU load balancing, raising training throughput by 30.8% over the original batching strategy. Built on Qwen3.5-4B and trained on 12.3M question-answer pairs converted from existing spatial datasets, MJev-Spatial achieves the highest overall score (65.64%) among eight tested models, ranks first on seven of twelve benchmarks, and outperforms its unmodified autoregressive base model on all twelve, while running 2.5 times faster than the fastest autoregressive baseline and 5.6 times faster than the base model.
Han Zhu, Juntao Dai Peking University, Jiaming Ji, Haoran Li, Chengkun Cai University of Edinburgh, University of Edinburgh, Pengcheng Wen, Chi-Min Chan, Boyuan Chen 等 11 位作者
With the widespread use of multi-modal Large Language models (MLLMs), safety issues have become a growing concern. Multi-turn dialogues, which are more common in everyday interactions, pose a greater risk than single prompts; however, existing benchmarks do not adequately consider this situation. To encourage the community to focus on the safety issues of these models in multi-turn dialogues, we introduce SafeMT, a benchmark that features dialogues of varying lengths generated from harmful queries accompanied by images. This benchmark consists of 10,000 samples in total, encompassing 17 different scenarios and four jailbreak methods. Additionally, we propose Safety Index (SI) to evaluate the general safety of MLLMs during conversations. We assess the safety of 17 models using this benchmark and discover that the risk of successful attacks on these models increases as the number of turns in harmful dialogues rises. This observation indicates that the safety mechanisms of these models are inadequate for recognizing the hazard in dialogue interactions. We propose a dialogue safety moderator capable of detecting malicious intent concealed within conversations and providing MLLMs with relevant safety policies. Experimental results from several open-source models indicate that this moderator is more effective in reducing multi-turn Attack Success Rate (ASR) compared to existed guard models.
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
Chi-Min Chan, Han Zhu, Chunyang Jiang, Jiaming Ji, Juntao Dai, Wei Xue Hong Kong University of Science and Technology, Sirui Han The Hong Kong University of Science and Technology, Yike Guo Hong Kong University of Science and Technology and Imperial College London
Multimodal Process Reward Models (MPRMs) have emerged as a pivotal framework for enhancing the reasoning capabilities of Multimodal Large Language Models (MLLMs). However, the research community currently lacks a dedicated benchmark to rigorously assess the error discernment capabilities of these models.To address this gap, we introduce PRMBench-V, a novel benchmark specifically designed to evaluate MPRMs’ proficiency in detecting erroneous reasoning steps across diverse error categories. Leveraging a semi-automated annotation pipeline augmented with human verification, we construct a comprehensive dataset comprising 907 unique queries, each annotated with nine distinct error types, resulting in 8,163 test cases with fine-grained step-level error labels.Through extensive experiments involving over 15 open- and closed-source models, we uncover several key findings: (1) even the strongest existing MPRMs achieve only \textasciitilde30\% accuracy in error identification; (2) while partial error detection achieves moderate precision and recall (\textasciitilde60\%), overall accuracy remains low (\textasciitilde20\%); and (3) benchmark scores exhibit a strong correlation with downstream task performance gains (r=0.86). Furthermore, we demonstrate that PRMBench-V can inform the development of more robust MPRMs: by introducing the Bayesian Rater Reliability Process Reward Model (BR2-PRM), we achieve up to a 4.8% performance improvement through test-time scaling.We believe that PRMBench-V will serve as a valuable resource for advancing MPRM research, enabling more rigorous evaluation and fostering the development of models with fine-grained multimodal reasoning capabilities.
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