LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis
Haoxiang Cao, Yuxuan Zhang, Penghui Du, Bo Li, Jiajiong Cao, Yingying Fan, Zimeng Wu, Yifu Luo 等 19 位作者
AI 速览：作者提出 LongBanana，一个由专家设计并验证的长上下文多参考图像合成基准，并配套 LongBanana-Gen 智能体框架来评测和提升这类生成任务。
We introduce LongBanana, to our knowledge the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis in the age of agent-driven image generation. The task requires retrieving dispersed visual evidence, rejecting distractors, binding entities and relations, and composing one coherent final image from multiple references. LongBanana contains 90 carefully designed tasks organized by a 3-by-3 taxonomy; each task includes about 32 reference images, 20 constraints, and 50 evaluation items, yielding 4,670 expert-specified atomic evaluation items in total. Its graph-grounded evaluation design instantiates each task with reference images and an entity-relation graph, then converts node, edge, and whole-graph obligations into checklist items for Content Fidelity, Relational Binding, and Global Constraint Satisfaction. Reviewed expert labels include a 15% double-labeled subset with 89.4% exact agreement and 0.83 quadratic weighted kappa. We further introduce LongBanana-Gen, an agent-harness framework for long-horizon image generation; experiments show that minimal agent harnesses improve over direct generation, while a ReAct-Plan-Exec-Reflect (RPR) workflow yields harness-dependent gains by strengthening evidence grounding and repair. Together, LongBanana provides an auditable testbed for measuring dense visual evidence use and a shared target for developing stronger agent-driven image-generation systems. Code and data are available at the project page: https://long-banana.com/.
Hundreds of languages have written text but almost no translations, and whether a model learning one from untranslated text builds on a related language or familiar letters is unclear: they usually coincide. Here we show that familiar letters, not a related language, let an unseen language into a shared space of meaning. GlyphMind, a tokenizer-free vision transformer that reads text as pixels and lets any sentence be shown in any script, was pretrained to align translations across 27 writing systems. Adapted on untranslated Lao, a language and script it had never seen, it retrieved English translations seven times as often as the strongest simple baseline. Rewritten in unfamiliar letters, Lao did not enter although its close relative Thai was in pretraining; Santali, which had none, entered in Devanagari, familiar from Hindi, but not in its own script. Transliteration into familiar scripts may thus open multilingual models to languages without translations.
Effective LLM governance extends beyond detecting unsafe responses to guiding their correction through closed-loop assessment, revision, and verification. Learning-based guards provide strong semantic discrimination but often leave policy knowledge implicit in trained parameters, limiting its explicit reuse across stages. Programmable guardrails make policies explicit through prompts and workflows, but rely heavily on manual engineering to operationalize them. This motivates explicit policy representations that are both learnable and reusable across governance operations. To this end, we introduce PolicyMem, a geometric policy memory that compiles natural-language policies into learnable low-rank subspaces for reuse across governance stages. A policy assessor projects query-response representations onto these subspaces, producing evidence that directly mediates safety judgments and policy attribution. A response rewriter uses the implicated policies to revise unsafe responses, with the same assessor re-evaluating the revised responses. Beyond inference, the same memory constructs training data from model-generated candidates to supervise the rewriter without external teacher calls. Across five widely used benchmarks, PolicyMem achieves the best average unsafe-behavior detection performance among representative baselines while supporting effective policy attribution, rewriting, and post-intervention assessment through the shared policy memory.
Legal judgment prediction models may retain a guilty verdict after a legally decisive condition has been altered such that the specified charge is no longer established. We study this guilty bias via selective repair over matched charge-established and charge-not-established cases, where the unsupported verdict should change while its supported counterpart remains stable. We construct CAIL-MC, a legally reviewed dataset of 11,478 matched pairs covering 261 Chinese criminal charges. Held-out activation interventions show that contrasts learned from correctly resolved pairs transfer to false-guilty cases, but target-reaching interventions can also damage matched supported-guilty decisions. Motivated by this observation, we introduce PROSE, which reconstructs case-specific targets from a training-derived residual geometry, aligns them with a local margin-descent signal, attenuates supported-guilty-sensitive components, and realizes bounded targets through a preservation-regularized low-rank parameter update. Across Qwen3-8B and GLM4-9B, PROSE achieves stronger pairwise safe repair and evidence sensitivity, while preserving standard single-pass inference.
Large language model (LLM) services face rapidly rising serving costs. To sustain free or affordable access for a broad user base, alternative monetization mechanisms are needed. Native advertising offers a promising approach, as it creates economic value by aligning user intent with merchant offerings. Yet how to integrate sponsored content naturally into LLM responses remains largely underexplored. Existing approaches primarily rely on prompt-level injection, multi-pass rewriting, or fine-tuning, which can compromise response quality or introduce additional computational costs. We propose AdSteer, a lightweight decoding-time advertising framework that adaptively adjusts token probabilities to integrate contextually appropriate advertisement into responses. AdSteer calibrates the strength and duration of advertising influence, making sponsored content a natural part of the response. Experiments on GEM-Bench across three application scenarios show that AdSteer achieves an average improvement of 14.1% in overall score over the strongest baselines and 22.5% fewer total tokens on average than rewrite-based methods. These results demonstrate that AdSteer provides an effective and practical solution for native advertising in LLMs.
Scaling depth is a key route to stronger large language models, but Pre-LN architectures often suffer from depth-wise information dilution, limiting the effective use of deep layers. While cross-layer connectivity has proven effective at mitigating this problem, existing designs usually introduce additional computation or memory access. We revisit cross-layer KV sharing as an efficient form of such connectivity. Beyond cache compression, each KV reuse edge creates a direct gradient path from deep layers to source KV projections, turning KV sharing into an implicit structural regularizer for KV representation geometry. We provide a systematic analysis of this mechanism, showing that proper sharing topologies sharpen key retrieval directions while enriching the value content subspace, whereas existing adjacent or global-anchor designs do not fully exploit this effect. Motivated by this, we propose DepthGraft, a hierarchical KV-sharing architecture that reconstructs deep-layer KV from stride-aligned shallow-middle source pairs. DepthGraft distributes non-local gradient flow across source layers, while preserving prefill early completion and reducing KV cache by roughly one third. Experiments on dense and MoE models from 0.5B to 16B parameters show consistent downstream improvements under both GQA and MLA. Notably, on the 16B MoE model, DepthGraft achieves a 7.0% absolute improvement on MMLU-Pro, along with gains of 6.6% on CMMLU and 6.4% on C3. These results validate the potential of DepthGraft as a scalable design for large-scale pretraining. The project page is available at https://github.com/jianke0604/DepthGraft.
Large language model agents automate web penetration testing through observation-action loops, enabling defenders to influence decisions with deceptive decoy observations. Advanced agents can recognize and bypass static decoys, then refocus on real targets. We propose AgentSnare, a trajectory-adaptive deception system that expands a decoy environment alongside exploration. For online deployment, ShadowOS combines an LLM-driven decoy constructor with a virtual kernel. Conditioned on exploration, it generates decoy artifacts that the kernel materializes and handles consistently across interactions, sustaining long-horizon exploration. Because deploying a powerful LLM for decoy construction is costly and slow, we train a small LLM using an offline adversarial interaction-simulation framework. It synthesizes multi-round trajectories by generating candidates and retaining those that sustain simulated interactions. Across 15 CVE-Bench web applications and four attacker models, AgentSnare absorbs 49.9% of tool calls to delay attacks, retains 57.2% of post-entry actions in the decoy to divert trajectories, and grounds 89.5% of completion reports in decoy evidence to defuse threats; across 60 attacker-CVE pairs, no real target is exploited at pass@3.
Spoken medical question answering can fail when automatic speech recognition (ASR) corrupts a clinically important term before retrieval. We present MedAstra, a framework that maps frozen WavLM speech representations through a lightweight adapter into a frozen E5 text embedding space. This enables a dense retrieval path that accepts a waveform without a query transcript. Retrieved passages and speech-linked textual metadata support answer generation with entity alternatives, diagnostic contrasts, relation heuristics, and bounded retrieval refinement. ASR-less refers to the initial speech retrieval interface; lexical retrieval, string-based entity correction, and the text generator require textual inputs. The deterministic speech embedding does not constitute a calibrated model of acoustic uncertainty. Reported retrieval results reach 87.6 Recall@10 for the speech-only retriever and 90.4 for the hybrid configuration, compared with 81.0 for transcript-first dense retrieval. Medical QA and ablations provide descriptive system comparisons, with input-provenance and evaluation limitations stated explicitly. We distinguish controlled synthetic acoustic tests from real clinical speech, qualify component attribution, and specify the remaining comparisons needed to assess speech-only QA and clinical generalization.
Elucidating the Design Space of LM Data Attribution
Steven Yu, Sheng-Yu Wang, Kimia Hamidieh, Antonio Torralba, David Bau
AI 速览：作者提出可控归因基准 CATT，通过合成虚构知识隔离数据集属性，并发现负对数几率梯度与仅聚合关键 token 能显著提升事实溯源归因效果。
Data attribution methods for Language Models (LMs) are typically evaluated using downstream tasks like fact retrieval. However, existing evaluations vary multiple dataset properties simultaneously, making it impossible to isolate exactly why an attribution algorithm fails. To diagnose these failures, we explore the design space of LLM data attribution, with a primary focus on evaluation. We introduce the Controlled Attribution Benchmark (CATT) for fact tracing, in which we fine-tune LLMs on entirely fictitious knowledge, providing unambiguous ground truth for attribution. This synthetic setup allows us to systematically alter one dataset configuration at a time, such as document redundancy and lexical distractors. With this configurable evaluation, we demonstrate that redundancy and distractors consistently degrade attribution performance. Rather than proposing a new estimator, we examine two design choices that are underexplored: the loss used to compute gradients and how per-token attribution scores are aggregated into a document score. We find that the true causal evidence within a training document is highly localized, but standard methods dilute this evidence into noise by averaging attribution scores across the entire document. We find that replacing standard cross-entropy loss gradients with the negative log-odds gradients surfaces the critical tokens from other uninformative ones. We can then aggregate only the top token scores to improve ground truth recall by 35 to 72 percentage points across multiple baselines on our diagnostic benchmark. The negative log-odds correction also generalizes to FTRACE-TREX, improving counterfactual retraining performance for every method we test, by up to 4.9x.
Proactive agents must decide whether, when, and how far to assist without an explicit request. Existing evaluations often rely on automated scores or binary verdicts, offering limited human diagnosis of evidence use and intervention quality. To address these gaps, we introduce HALO, Human-Grounded GrAnular EvaLuation of PrOactive Agents. HALO-Arena covers 4 proactive need types, Dependency, Blockage, Risk, and Opportunity, across 10 everyday domains and 50 personal environments. Scenarios require agents to retrieve evidence from structured databases or unstructured documents before judging whether assistance is warranted and choosing its scope. On HALO-Arena, we construct HALO-8K, a dataset of 8,000 responses from 10 models across 400 scenarios, annotated by 20 experts for action level, acceptability verdicts, and human diagnosis. For rejected responses, human diagnosis covers timing, intensity, and other issues. Recorded reasoning tokens and retrieval rounds measure the effort behind these responses. The best proprietary and open-source models achieve 72.88% and 56.63% acceptability, respectively. Our central finding is that over-proactivity is the dominant observed failure mode, with Premature timing or Excessive intensity accounting for 61.20% of rejected responses. Further analyses yield complementary findings on the balance between intervention and restraint, evidence acquisition, and the calibration of timing and intensity.
Preference modeling aims to capture heterogeneous human values and increasingly relies on learning latent components to represent this variation without fine-grained supervision. What these components actually represent, however, is rarely examined. We audit Mixture-of-Experts (MoE) reward models along three rungs---semantic coherence, functional correspondence, and composability---in controlled and real-world settings, holding the architecture and training data fixed so that differences isolate the effect of the training objective. Our diagnosis is that the components conventional MoE reward models learn are entangled, which limits not only interpretability but also the personalization such models are built for. The entanglement is not intrinsic to the architecture but a property of the training objective: off-the-shelf sparsity and diversity regularizers already substantially reduce it. The resulting configuration, which we call sparse MoE, recovers all underlying semantic categories without domain supervision in a controlled setting, reaching 84.8% expert purity against 60.8% for conventional MoE; on real-world data its experts admit human-interpretable task- and domain-level descriptions aligned with their functional strengths. Targeted interventions on routing weights produce predictable preference changes, showing that these interpretations are behaviorally grounded rather than merely descriptive. Crucially, the disentanglement induced by this objective also makes the experts composable: lightweight router updates yield a 25.81-point improvement in few-shot personalization with 50 examples, roughly 2.7 times the gain of the identical architecture and adaptation procedure applied to entangled experts. Together, these results show that MoE reward models can learn interpretable expert structure, and that the same structure can substantially improve personalization.
Bingchao Wang, Yongcheng Jing, Shunyu Liu, Chengyu Wang, Jun Huang, Jie Yang, Wei Liu, Dacheng Tao
Mobile GUI agents are typically fine-tuned 7B-scale vision-language models (VLMs) that generate each action as text. However, these VLMs decode every action token by token, and even agents that score candidate actions keep a large language model or VLM in the loop, making them slow and memory-hungry. To address these issues, we propose GUI-JEPA, to our knowledge the first decoder-free GUI agent policy built on a Joint-Embedding Predictive Architecture (JEPA), which selects actions in the latent space of a JEPA encoder pretrained on unlabelled GUI screenshots. Its action- and region-conditioned latent predictor turns the screen latent into one latent per candidate action, letting lightweight heads score all candidates against the instruction in a single batched pass, without decoding a token. We further show that the scorer relies on this latent pathway and that the predictor is trained for the decision itself: a randomly initialised predictor does just as well as one pretrained on screen dynamics. On AndroidControl, GUI-JEPA is 2.8-4.0x faster on the same GPU than a Qwen2.5-VL-7B agent we fine-tune with the same inputs, while reaching 86-92% of its in-distribution accuracy with 6% of its parameters. Code is available at https://github.com/bcwang-sjtu/GUI-JEPA
Bingchao Wang, Yongcheng Jing, Chengyu Wang, Jun Huang, Jie Yang, Wei Liu, Dacheng Tao
GUI grounding on high-resolution screens fails mostly on small targets, and zooming in is a common training-free remedy. However, most existing zooming pipelines do not match the zoom to the target's size, accept the zoomed answer unconditionally and never revisit a wrong answer. To address these issues, we propose AdaGround, a training-free framework with three components: (1) Token-Normalized Adaptive Zooming, which sets the magnification, field of view and zoom depth of each sample from the model's own estimate of the target size; (2) Verifier-in-the-Loop Refinement, which uses one marked-click Yes/No verifier to keep or reject each zoom, choose between zoom chains, trigger re-grounding and, with an existence query, decide when to abstain; and (3) Compositional Structure Grounding, which composes spreadsheet cells from their headers. On Qwen3-VL models, AdaGround surpasses state-of-the-art training-free zooming methods and many trained GUI models, improving accuracy over the backbone by 7.4 to 20.6 points on OSWorld-G and ScreenSpot-Pro, and its gains carry over to held-out UI-Vision. Code is available at https://github.com/bcwang-sjtu/AdaGround
Yao Huang, Zhiyuan Zeng, Yu Wang, Zhitong Zhu, Dingcheng Duan, Hao Li, Shangzhi Xu, Yifan Wang 等 17 位作者
As frontier language models grow increasingly capable and autonomous, independent third-party evaluations are critical for characterizing their systemic safety risks. While closed-source releases are routinely accompanied by comprehensive system cards, open-weight models receive less scrutiny regarding frontier risks. To address this gap, we present an independent safety evaluation of DeepSeek V4.1 Flash-1017, the latest version of V4.1 Flash that departs from V4 with a new architecture, scaled-up training, and strong agentic capabilities. Drawing on the EU General-Purpose AI Code of Practice, we evaluate the model across 34 benchmarks, including 5 in-house suites, spanning evaluation integrity, loss of control, and harmful manipulation. Conforming to the AEF-1 standard, we compare V4.1 Flash-1017 against V4.1 Flash, earlier DeepSeek V4 models, and leading closed-source systems (GPT-6 Astra, GPT-5.6 Sol, and Gemini 3.8 Flash). Overall, V4.1 Flash-1017 shows substantial safety improvements over the V4 generation and approaches or matches leading closed-source baselines on several dimensions, while still exhibiting notable behaviors that warrant continued attention. Building on our findings, we highlight three key observations and recommendations. First, because safety can shift and even regress across minor updates, developers should integrate safety evaluation into incremental releases. Second, as final-turn outputs and end-state checks alone cannot reliably detect loss-of-control behaviors, deployments require process-level observability across both reasoning traces and action logs. Third, because public benchmarks elicit strong evaluation awareness and face data contamination risks, developing private benchmarks is essential for trustworthy frontier evaluation.
Chain-of-Thought (CoT) reasoning has significantly enhanced the reasoning capabilities of large language models (LLMs). However, existing reinforcement learning methods primarily optimize outcome-level rewards, providing limited supervision over the reasoning process. As a result, models may produce shortcut-prone or unstable reasoning trajectories even when the final answer is correct. We propose GeoFaith, a spatio-temporal latent dynamics framework for learning faithful reasoning processes without process reward models. GeoFaith combines final-state latent uncertainty, which measures whether a completed trajectory lies in a well-supported manifold region, with predictive-entropy dynamics, which penalize abnormal uncertainty patterns such as premature collapse, oscillation, and prolonged uncertainty. These signals are incorporated as process regularization into GRPO, jointly optimizing answer correctness and reasoning quality without step-level process annotations. Experiments show that GeoFaith consistently improves reasoning accuracy, produces more efficient reasoning trajectories, and outperforms both outcome-only reinforcement learning and process-supervision baselines.
Yunqi Liu, Peiyu Wang, Bingli Wang, Leo Luo, Xuanwu Song, Xudong Ma, Hongxin Li, Yaxin Du 等 24 位作者
Command-line interface (CLI) agents increasingly execute longer-horizon tasks, yet existing benchmarks either assume that user instructions remain fixed throughout execution or evaluate task outcomes without capturing runtime collaboration quality. In this paper, we present REALCLI, a paired benchmark that unifies the evaluation of task execution and human-agent collaboration quality on realistic CLI tasks. To comprehensively measure collaboration, REALCLI builds a five-dimensional framework, operationalized into 22 sub-dimensions and 51 interaction scenario templates. REALCLI provides two modes for paired comparison: REALCLI-EXEC supplies complete specifications upfront, whereas REALCLI-COLLAB introduces a scripted user that delivers progressive information, corrections, constraint revisions, and authorization decisions. Evaluations reveal profound capability gaps in frontier agents: the best configuration attains an average score of only 49.2%, and when evaluated by the joint collaboration score ($J_{\mathrm{collab}}$), the peak performance reaches merely 37.6%. These results show that task effectiveness diverges sharply from collaboration quality, with models consistently struggling in steerability, trust calibration, and safety boundary management. REALCLI provides a rigorous, reproducible foundation for measuring and advancing collaborative CLI agents.
Mingyu Wang, Bin Yu, Wenjian Lu, Zhi Wang, Gao Kefeng, Cheng Wen Xi'an University of Electronic Science and Technology, Xu Lu Xidian University, Cong Tian
Linux kernel device drivers are tightly coupled with hardware, making them difficult to execute and test without physical devices. This heavily limits automated code analysis and vulnerability discovery. While manual modeling is unscalable, Large Language Models (LLMs) offer a new approach to scale virtual device construction across the Linux driver ecosystem. In this paper, we present DevGen, an LLM-powered tool that generates QEMU-based virtual devices directly from Linux driver source code. DevGen combines static analysis to gather necessary context, guides the LLM through step-by-step prompting, and uses an automated self-correction loop driven by compilation and execution feedback. To further reduce errors, similar fixes are retrieved from a library of common modeling failures and incorporated into the repair prompt, which supports more targeted corrections in later iterations. The generated devices finally integrate with QEMU and Syzkaller, enabling driver fuzzing without physical hardware. DevGen is evaluated on 50 PCI/PCIe drivers from Linux 6.18 using three mainstream LLMs, and successfully generates usable models for 44 drivers. In these drivers, 24% of them achieve significant improvements in fuzzing coverage, and 7 previously unknown crashes are triggered with 1 CVE assigned. These results demonstrate the practical capability of LLMs to automate complex, system-level code generation tasks.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) · 2026 · 正式发表
Generative AI has enabled the creation of photorealistic images and videos that are increasingly disseminated on social media, often used for spam, misinformation, manipulation, and fraud. Existing AI-generated content (AIGC) detection methods face challenges including poor generalization to new generation models, reliance on single modalities, and lack of interpretable explanations. We present our pipeline that mitigates these issues by continuously curating diverse multi-modal social media data and training a compact vision-language model for detection and explanation. Our model achieves state-of-the-art detection performance on public benchmarks and demonstrates robust detection and explanation capabilities on internal social media datasets across multiple platforms. We deployed our model for post recommendation on social media platforms and observed positive downstream impacts on user engagement, demonstrating that it is feasible to perform effective AIGC detection in dynamic, real-world social media environments.
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
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
Yuanjie Lyu, Chengyu Wang, Haonan Zheng, Yuanhao Yue, Junbing Yan, Ming Wang, Jun Huang
Modern industrial applications increasingly demand language models that act as agents, capable of multi-step reasoning and tool use in real-world settings. These tasks are typically performed under strict cost and latency constraints, making small agentic models highly desirable. In this paper, we introduce the AgenticQwen family of models, trained via multi-round reinforcement learning (RL) on synthetic data and a limited amount of open-source data. Our training framework combines reasoning RL and agentic RL with dual data flywheels that automatically generate increasingly challenging tasks. The reasoning flywheel increases task difficulty by learning from errors, while the agentic flywheel expands linear workflows into multi-branch behavior trees that better reflect the decision complexity of real-world applications. We validate AgenticQwen on public benchmarks and in an industrial agent system. The models achieve strong performance on multiple agentic benchmarks, and in our industrial agent system, close the gap with much larger models on search and data analysis tasks.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track) · 2026 · 正式发表
Minda Zhao, Yilun Du Google, Mengyu Wang Harvard University
As large language models (LLMs) transition from chat interfaces to integral components of stochastic pipelines and systems approaching general intelligence, the ability to faithfully sample from specified probability distributions has become a functional requirement rather than a theoretical curiosity. We present the first large-scale, statistically powered audit of native probabilistic sampling in frontier LLMs, benchmarking 11 models across 15 distributions. To disentangle failure modes, we employ a dual-protocol design: Batch Generation, where a model produces N{=}1000 samples within one response, and Independent Requests, comprising N{=}1000 stateless calls. We observe a sharp protocol asymmetry: batch generation achieves only modest statistical validity, with a 7% median pass rate, while independent requests collapse almost entirely, with 10 of 11 models passing none of the distributions. Beyond this asymmetry, we reveal that sampling fidelity degrades monotonically with distributional complexity and aggravates as the sampling horizon N increases. Finally, we demonstrate how the propagation of these failures into downstream real-world application tasks introduces systematic biases: models fail to enforce uniform answer-position constraints in Multiple Choice Question generation and systematically violate demographic targets in attribute-constrained text-to-image prompt synthesis. These findings indicate that current LLMs lack a functional internal sampler, necessitating external tools for applications requiring statistical guarantees.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Hao Gu The Hong Kong University of Science and Technology, Lujun Li, Hao Wang, Lei Wang, Zheyu Wang, Bei Liu The Hong Kong University of Science and Technology, Jiacheng Liu, Qiyuan Zhu 等 10 位作者
Binary quantization represents the most extreme form of compression, reducing weights to \pm1 for maximal memory and computational efficiency. While recent sparsity-aware binarization achieves sub-1-bit compression via weight pruning, it faces critical challenger: performance degradation, mask-management overhead, and limited hardware compatibility. In this paper, we present BTC-LLM, a novel sub-1-bit LLM quantization framework that leverages binary pattern clustering and weight transformation to overcome these limitations. Our approach incorporates two key innovations: (1) a Binary Codebook that clusters recurring vectors into compact indices using custom distance metrics and sign-based updates; (2) a Learnable Transformation that reduces outliers and promotes shared sign patterns among binary weights. This eliminates sparse masks, enabling efficient inference on standard hardware. Extensive evaluations across LLaMA, Qwen, and FBI-LLM families demonstrate that BTC-LLM achieves state-of-the-art results in extreme compression (1.11–0.7 bits). Notably, BTC-LLM compressed to 0.8 bits on LLaMA-2-13B maintains high performance—with only a 3.1% accuracy drop in zero-shot benchmarks—while delivering a 1.6\times speedup over FP16.
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
Rongchuan Mu, Zexin Wang Harbin Institute of Technology, Qianyu Wang, MingHua Ma, Zekun Wang Alibaba Group, Ming Liu Harbin Institute of Technology, Bing Qin Harbin Institute of Technology
Large Reasoning Models (LRMs) excel at complex problem-solving but frequently overlook specific instruction constraints. Existing alignment methods struggle to balance general reasoning with instruction-following (IF), hindered by dependency on teacher models, reward hacking, and reasoning-answer inconsistencies. We propose PARIF, a two-stage curriculum learning framework based on Reinforcement Learning from Verifiable Rewards (RLVR) to enhance both IF and general reasoning capabilities. The framework employs a correctness proxy across different stages to mitigate reward hacking. Stage I employs a dynamic weighting strategy simultaneously to optimize the model’s reasoning paradigm regarding constraints. Stage II introduces Decoupled-GRPO, which builds upon the first stage to enhance the logical consistency between the reasoning process and the final answer, enabling the model to better leverage its optimized reasoning paradigm. To support the framework, we curate 26,000 high-quality instructions featuring diverse constraints. Extensive experiments demonstrate PARIF’s effectiveness: our 7B model achieves a remarkable 21.25% relative average improvement to the original model across six representative IF tasks, while our 8B model outperforms leading models like DeepSeek-V3 on these IF tasks, effectively pushing the Pareto frontier of instruction following and reasoning for models of comparable scale. We open-source our code and models to facilitate future research.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Lingyue Fu, Hao Guan, Bolun Zhang, Haowei Yuan, Yaoming Zhu, Lin Qiu, ZongYu Wang Meituan, Xuezhi Cao Meituan 等 12 位作者
The evaluation of Large Language Models (LLMs) for software engineering has shifted towards complex, repository-level tasks. However, existing benchmarks predominantly rely on coarse-grained pass rates that treat programming proficiency as a monolithic capability, obscuring specific cognitive bottlenecks. Furthermore, the static nature of these benchmarks renders them vulnerable to data contamination and performance saturation. To address these limitations, we introduce CoreCodeBench, a configurable repository-level benchmark designed to dissect coding capabilities through atomized tasks. Leveraging our automated framework, CorePipe, we extract and transform Python repositories into a comprehensive suite of tasks that isolate distinct cognitive demands within identical code contexts. Unlike static evaluations, CoreCodeBench supports controllable difficulty scaling to prevent saturation and ensures superior data quality. It achieves a 78.55% validity yield, significantly surpassing the 31.7% retention rate of SWE-bench-Verified. Extensive experiments with state-of-the-art LLMs reveal a significant capability misalignment, evidenced by distinct ranking shifts across cognitive dimensions. This indicates that coding proficiency is non-monolithic, as strength in one aspect does not necessarily translate to others. These findings underscore the necessity of our fine-grained taxonomy in diagnosing model deficiencies and offer a sustainable, rigorous framework for evolving code intelligence. Code of CorePipe framework and data of CoreCodeBench are available in https://github.com/AGI-Eval-Official/CoreCodeBench and https://huggingface.co/collections/tubehhh/corecodebench.
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
Hongcheng Liu, Yuhao Wang, Zhe Chen, Pingjie Wang, Zhiyuan Zhu, Yixuan Hou Shanghai Jiaotong University, Yanfeng Wang Shanghai Jiao Tong University, Yu Wang Shanghai Jiao Tong University
Omni Large Language Models (Omni-LLMs) have demonstrated impressive capabilities in holistic multi-modal perception, yet they consistently falter in complex scenarios requiring synergistic omni-modal reasoning. Beyond understanding global multimodal context, effective reasoning also hinges on fine-grained cross-modal alignment, especially identifying shared referents across modalities, yet this aspect has been largely overlooked. To bridge this gap, we formalize the challenge as a cross-modal coreference problem, where a model must localize a referent in a source modality and re-identify it in a target modality. Building on this paradigm, we introduce CrossOmni, a dataset comprising nine tasks equipped with human-designed reasoning rationales to evaluate and enhance this capability. Experiments on 13 Omni-LLMs reveal systematic weaknesses in cross-modal coreference, which we attribute to the absence of coreference-aware thinking patterns. To address this, we enhance cross-modal alignment via two strategies: a training-free In-Context Learning method and a training-based SFT+GRPO framework designed to induce such thinking patterns. Both approaches yield substantial performance gains and generalize effectively to collaborative reasoning tasks. Overall, our findings highlight cross-modal coreference as a crucial missing piece for advancing robust omni-modal reasoning.
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
Xiaojian Li, Rongwu Xu, Tianyun Zhang, Yue Wang, Shuo Chen, Qiner Lyu Shanghai Qi Zhi Institute, Briana Zhang, Peiran Yang Tsinghua University 等 12 位作者
As language models (LMs) exhibit increasingly consciousness-like behaviors, evaluating their cognitive abilities becomes essential. We introduce AwarenessBench, the first comprehensive benchmark for assessing the cognitive abilities of LMs in four dimensions: metacognition, self-awareness, social awareness, and situational awareness, covering 15 cognitive functions and 14,381 samples. Evaluating 18 state-of-the-art LMs, we find that all consistently surpass random baselines, with more advanced models performing better. We further compare LMs with human performance across three demographic groups, where the best-performing model surpasses human averages overall, but most still fall markedly short in metacognition and self-awareness. Finally, we show that awareness is a distinct capability: progress in language modeling or reasoning does not necessarily translate into improved cognition.
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
Hongcheng Liu, Pingjie Wang, Yuhao Wang, Siqu Ou Shanghai Jiaotong University, Yanfeng Wang Shanghai Jiao Tong University, Yu Wang Shanghai Jiao Tong University
Multimodal large language models (MLLMs) have shown strong capabilities across a broad range of benchmarks. However, most existing evaluations focus on passive inference, where models perform step-by-step reasoning under complete information. This setup is misaligned with real-world use, where seeing is not enough. This raises a fundamental question: Can MLLMs actively acquire missing evidence under incomplete information? To bridge this gap, we require the MLLMs to actively acquire missing evidence and iteratively refine decisions under incomplete information, by selecting a target image from a candidate pool without task-specific priors. To support systematic study, we propose GuessBench, a benchmark with both perception-oriented and knowledge-oriented images for evaluating active reasoning in MLLMs. We evaluate 20 superior MLLMs and find that performance on active reasoning lags far behind it on passive settings, indicating substantial room for improvement. Further analysis identifies fine-grained perception and timely decision-making as key challenges. Ablation studies show that perceptual enhancements benefit smaller models, whereas thinking-oriented methods provide consistent gains across model sizes. These results suggest promising directions for future research on multimodal active reasoning.
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
Ao Sun, Xiaoyu Wang, Zhe Tan, Yu Li, Jiachen Zhu ByteDance Inc., Yuheng Jia Southeast University, Shu Su Southeast University
As Large Language Models (LLMs) serve a global audience, alignment must transition from enforcing universal consensus to respecting cultural pluralism. We demonstrate that dense models, when forced to fit conflicting value distributions, suffer from Mean Collapse, converging to a generic average that fails to represent diverse groups. We attribute this to Cultural Sparsity, where gradient interference prevents dense parameters from spanning distinct cultural modes. To resolve this, we propose CuMA (Cultural Mixture of Adapters), a framework that frames alignment as a conditional capacity separation problem. By incorporating demographic-aware routing, CuMA internalizes a Latent Cultural Topology to explicitly disentangle conflicting gradients into specialized expert subspaces. Extensive evaluations on WorldValuesBench, Community Alignment, and PRISM demonstrate that CuMA achieves competitive performance, outperforming both dense baselines and semantic-only MoEs. Our analysis confirms that CuMA effectively mitigates mean collapse and preserves cultural diversity. Our code is available at https://github.com/Throll/CuMA.
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
Xue Jiang, Ge Li Peking University, Jiaru Qian, Xianjie Shi, Chenjie Li Peking University, Hao Zhu, Ziyu Wang Peking University, Jielun Zhang Peking University 等 14 位作者
Large language models (LLMs) excel at general programming but struggle with domain-specific software development. This gap motivates research into domain specialization methods that enable LLMs to learn and utilize domain knowledge and data. However, existing domain-specific code benchmarks focus on assessing what knowledge LLMs possess rather than how they acquire and apply new knowledge, lacking explicit knowledge corpora for developing domain specialization methods. To this end, we present KOCO-bench, a novel benchmark designed for evaluating domain specialization methods in real-world software development. KOCO-bench contains 6 emerging domains with 11 software frameworks and 25 projects, featuring curated knowledge corpora alongside multi-granularity evaluation tasks including domain code generation (from function-level to project-level with rigorous test suites) and domain knowledge understanding (via multiple-choice Q A). Unlike previous benchmarks that only provide test sets for direct evaluation, KOCO-bench requires acquiring and applying diverse domain knowledge (APIs, rules, constraints, etc.) from the corpora to solve evaluation tasks. Our evaluations reveal that KOCO-bench poses significant challenges to state-of-the-art LLMs. Even with domain specialization methods (e.g., SFT, RAG, kNN-LM) applied, improvements remain marginal. Best-performing coding agent, Claude Code, achieves only 34.2%, highlighting the urgent need for more effective domain specialization methods. We release KOCO-bench, evaluation code, and baselines to advance further research at https://github.com/jiangxxxue/KOCO-bench.
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
Yubo Shan, Kun Zhang Tencent, Qiming Xu Beijing Technology and Business University, Liping Cao, Yingying Cao, Jian Zhang, Yu Wang, Jingyuan Li 等 9 位作者
Interleaved multimodal understanding and generation—where models can interactively comprehend and produce images and text in arbitrary orders—has emerged as a key research direction in generative Multimodal Large Language Models(MLLMs). Such interleaved image–text content plays an increasingly important role in information dissemination. However, the compounded persuasive power of multimodal narratives also raises the risk of factual misinformation. Despite this, existing benchmarks lack effective mechanisms to evaluate factual consistency in interleaved image–text content. To bridge this gap, we introduce FactVerse, a benchmark dedicated to evaluating factual consistency in interleaved image-text generation. FactVerse comprises 3,000 human-verified instances across four categories and 50 domains, supporting both English and Chinese. We also establish a multi-dimensional evaluation framework designed to rigorously assess factual consistency. Experiments demonstrate that our framework achieves high alignment with human judgments, significantly outperforming existing evaluation methods. Furthermore, our analysis reveals systematic deficiencies in current models, offering critical insights for future design.
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