Proactive query recommendation aims to suggest relevant search queries to users on the homepage of local service apps. Unlike conventional recommendation over a fixed item set, this requires open-set generation from a vocabulary of millions, where outputs must simultaneously align with user preference, match fine-grained spatiotemporal context, cover long-tail intents, and provide diverse alternatives within a few candidate slots. Existing approaches address these axes in isolation: standard tokenization suppresses long-tail queries, coarse location encoding misses fine-grained geographic demand, generic preference optimization fails to align with production impression-click distributions, and standard beam search collapses to near-duplicate outputs. We present SPADE (Spatiotemporal Preference Alignment and Diversity Enhancement), an end-to-end generative framework that jointly optimizes the generation space, contextual representation, preference distribution, and output slate quality through four components: Objective-Aligned BPE (OA-BPE) aligns subword merging with business value signals to boost long-tail query generation probability; Spatio-Temporal Adaptive Representation (STAR) replaces coarse location tokens with continuous multi-scale encodings and soft gating for precise spatiotemporal grounding; MARS-DPO (Multi-dimensional Adaptive Reward Sampling DPO) combines cross-domain pre-training with preference alignment on production impression-click data via composite rewards and adaptive hard negative sampling; and Sigmoid Diversity Beam Search (SDBS) improves intent coverage without sacrificing recall. Offline experiments on large-scale production data show that SPADE outperforms the strongest generative baseline by +5.6% HR@5, -40.3% GeoC@5, and +1.9% Exec@5. A one-week online A/B test confirms production impact: +1.88% Module CTR, +0.177% Search Volume, and +0.105% Search DAU.
Tianqi Gao, Chenxu Wang, Wenxuan Yang, Yueqi Mou, Biao Tang, Ming Liu, Peng Yan
Recommender systems are widely used in e-commerce and content platforms, where improvement requires repeated model design, training optimization, and experimental validation. Large language model (LLM) agents increasingly automate this development cycle. Sustained autonomous improvement requires agents to improve both recommendation models and their development workflows. However, task-performance-driven approaches select updates mainly by candidates' final scores, without a separate optimization signal for experimentation quality. We therefore present DualEvolve, a dual-level framework that co-evolves recommendation mechanisms and agent orchestration toward recursive self-improvement. Inner evolution revises model and training code using task feedback to improve recommendation performance. Outer evolution scores the inner-loop experimental process, updates harness and skill rules using these scores and execution traces, and adjusts orchestration choices based on next-round scores to improve the quality of subsequent experiments. Experiments on MovieLens-1M and three Amazon Reviews 2023 subsets cover four recommendation methods and 16 settings. DualEvolve achieves higher test means than Base, OpenEvolve, and DeepEvolve in all 64 setting-metric comparisons, with macro-averaged relative gains over DeepEvolve of 7.6% (HR@5) and 9.0% (NDCG@5).
Jun Liu, Zhenglun Kong, Changdi Yang, Fan Yang, Tianqi Li, Peiyan Dong, Joannah Nanjekye, Hao Tang 等 15 位作者
Large language model (LLM) agents are increasingly organized into multi-agent systems for complex reasoning and decision-making tasks. However, scaling such systems can substantially increase inference cost, especially when multiple agents repeatedly process overlapping or role-irrelevant context. Many existing multi-agent frameworks focus on orchestration and role specialization, but still rely on full-context sharing or static context routing, leading to redundant token consumption, increased runtime, and inefficient memory utilization. We present \emph{Role-Conditioned Context Routing (RCR)}, an execution-time layer that allocates role-relevant shared-memory items to agents under token budgets. Given a structured shared memory and a set of role-specialized agents, RCR selectively routes task-relevant memory items to each agent according to its role, optional task-stage metadata, and token budget. RCR integrates role-conditioned relevance scoring, token-budgeted memory selection, and iterative memory updates to reduce redundant context exposure during multi-agent execution. Experiments on multi-hop QA, embodied reasoning (ALFWorld), and web-based decision-making (WebShop) show that RCR reduces token usage by 35\% on average and runtime by 38\% on average relative to Full-Context, while maintaining or improving task performance.
Scaling Mathematical Proof Supervision from Structured Textbooks
Youqing Fang, Zhouqi Hua, Yinhao Tang, Yuzhe Gu, Lingkai Kong, Bin Liu, Weiming Zhang, Kai Chen 等 10 位作者
AI 速览：将2,118本大学数学教科书转化为约38万条保留章节顺序与命题-证明结构的证明监督记录，并研究课程学习与规模规律。
Mathematical supervision for language models is typically flattened into isolated problem-solution pairs, discarding the curricular organization and evidential structure of the original sources. We introduce TOME-Math (Textbook-Oriented Mathematical Evidence), a large-scale corpus that converts 2,118 textbooks spanning the core areas of university-level mathematics subjects into approximately 380K structured proof-supervision records while preserving chapter order, proposition-proof links, dependency references, and source provenance. Leveraging these structural signals, we explore curriculum learning across multiple mathematical subjects and characterize scaling patterns with supervision volume and source breadth. At full scale, fine-tuning on TOME-Math raises the unweighted mean score across our held-out evaluations and established public benchmarks by 16.7 percentage points over the base model, while exceeding all three public instruction-data baselines evaluated in aggregate. Our findings highlight structured textbooks as both a scalable source of mathematical supervision and an empirical framework for studying the impact of data organization on mathematical reasoning. We will release the corpus to facilitate further research on structured mathematical supervision.
When we speak of recursive self-improvement (RSI), are we speaking of a phenomenon, a mechanism, or a prospect? Towards autonomous and evolving intelligence, RSI is being claimed at many scales, while no single framework that formally describes these emerging instances exists. Its counterpart in the classical realm, iterative policy improvement, is characterized by generalized policy iteration (GPI), a framework of broad applicability with well-understood theoretical properties, but only where the update principle and the evaluation base lie outside the agent. In this paper, we propose Generalized Agent Iteration (GAI), a formal framework that describes iterative policy improvement and RSI as two cases of a single learning paradigm. GAI defines the agent as a configuration of modifiable components within a system and models the learning process as a cycle of agent evaluation and agent improvement. Two pivotal dials then distinguish the instances, in particular the systems that self-identify as RSI in recent work: whether the improving mechanism is part of the agent and whether the standard it is measured against is grounded outside it. The former dial delineates the boundary between GPI and RSI, and the latter determines a system's polarity as anchored, goal drift, or fully self-referential. Moreover, we use these coordinates to place existing systems on the same two axes and make the defects of recursive self-improvement statable one condition at a time. We see this paper as a first step toward exploring a formal characterization of RSI that rests on the classical account, makes existing systems comparable, and provides a principled basis for analyzing and designing new ones.
Zero-shot object navigation requires an agent to reason about where a target may be found while retaining evidence from previous observations. We present HGWM, a hierarchical graph-guided world model that combines a category-conditioned room--anchor prior with persistent spatial-semantic memory. A coarse stage uses a vision-language model to score candidate headings, while a fine stage explicitly compares observed objects and relations with the goal prior. Phase-dependent fusion integrates these signals with a language-model recommendation to guide local navigation. HGWM achieves 62.6\% success rate (SR) and 31.5\% success weighted by path length (SPL) on HM3D v0.1, improving over WMNav by 4.5 and 0.3 percentage points, respectively. On MP3D, it achieves 46.8\% SR and 17.9\% SPL, with gains of 1.4 and 0.7 points. On HM3D v0.2, the full model reaches 74.4\% SR and 31.9\% SPL; removing individual components reduces SR by 0.9--2.3 points and SPL by 0.6--1.5 points. These results support the use of relational priors and persistent graph memory for target-directed search.
Zhefan Rao, Meng Chu, Xuan Billy Zhang, Jingfan Chen, Guohao Cai, Runyi Li, Shiyi Du, Xu Huang 等 92 位作者
World Models (WMs) can support Multimodal Agents (MAs) through simulated experience, while Agent trajectories can inform World Model refinement. Evaluating this interaction requires separate evidence about action selection and state evolution. We introduce Dual Exam, which tests Agents under fixed environment response mechanisms and World Models under fixed action interventions across physical, digital, social, and scientific domains. The Agent track compares 10 configurations on 1,088 items in 15 reporting categories. The World Model track contains 400 cases across eight scenarios; 15 configurations have full coverage and one covers six scenarios. GPT 6 Astra leads the case-weighted Agent Index at 31.3, and GPT-Image-2 leads the full-coverage WM Overall at 21.7. These scores use different tasks and metrics and are not directly comparable. A 150-case pilot of one coupled pairing records zero task successes and motivates component-level diagnosis. Results reveal uneven task performance and failures to preserve task-relevant state. Their value for predicting coupled success or sustained reciprocal learning remains to be established.
We introduce Nemotron 3.5 Super and Nemotron 3.5 Lightning, open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. Super is a 120B-parameter, 12B-active multimodal model, while Lightning is a smaller 30B-parameter, 3B-active text model; both designed for efficient deployment. Both models combine large-scale distillation, long-context training, and native speculative decoding, while Super further incorporates Indic continued pretraining and agentic software-engineering mid-training. We develop a comprehensive post-training pipeline spanning supervised fine-tuning and reinforcement learning (RLVR and RLHF), complemented in Super by specialized teacher training and multi- teacher on-policy distillation (MOPD). This transforms Super from a text-only base model into a multimodal agent capable of reasoning over text, images, video, and interactive tool environments. We further explore test-time scaling through parallel agent execution and self-verification. Across training and deployment, we jointly optimize capability and efficiency. Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points, while agentic mid-training strengthens software-engineering capabilities and robustness across agent harnesses. The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. Post-training MTP boosting improves speculative- decoding acceptance length by up to 21.8% under probabilistic draft sampling, while NVFP4 quantization achieves 99.62% median accuracy recovery relative to BF16. Inference optimizations enable up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user.
Vision-language models are shifting from offline passive understanding toward online proactive interaction. However, prevailing methods largely rely on explicit requests while overlooking implicit user intents (e.g., "where should I click next?" when a user is navigating a website) and fail to continuously update their understanding as user intents evolve. These limitations make them less suitable for intent-aware interaction in real-world scenarios. To address these challenges, we introduce StreamVibe-Bench, a new benchmark that serves as a testbed for user intent understanding and interaction, comprising 4.1K evaluation samples and 85K training samples spanning explicit, implicit, and dynamic intents. Building on this benchmark, we develop StreamVibe-VL, a streaming proactive model that infers implicit user intents from continuous observations and dynamically adapts its interactions as these intents evolve. We further introduce StreamVibe-Harness, which maintains user intents as persistent states with flexible lifecycles through hierarchical memory, enabling it to track the latest user intents and environmental states. Our experiments reveal three key findings: (i) StreamVibe-VL preserves strong general video understanding capabilities while improving response timing and answer quality by 10.7% over existing baselines; (ii) existing proactive models remain limited in understanding user intents and determining when to respond, especially under implicit and dynamic intents; and (iii) StreamVibe-Harness further improves response consistency and adaptation during sustained interactions, resulting in a 12.3% gain in proactive interaction performance.
We prove a complexity classification of counting planar graph homomorphisms with non-negative weights. For a real symmetric matrix $M$ with nonnegative entries, the problem $\GH(M)$ is either (1) P-time computable over all graphs, or (2) \#P-hard in general but P-time computable over planar graphs, or (3) \#P-hard over planar graphs. Furthermore, $\GH(M)$ in (2) consists of precisely those that involve the P-time FKT algorithm to count planar perfect matchings with a holographic transformation. The dichotomy for entrywise positive and full rank matrices lies at the core of the proof. In this setting, the main tool is a matrix algebra ${A}$ (a vector space with bilinear multiplication) formed by taking the real span of (centered) entrywise logarithms of all realizable matrices from $M$ using planar edge gadgets and polynomial interpolation. We then define a multiplicative homomorphism $F(\cdot)$ on the algebra ${A}$, which turns out to be a polynomial map. The exponents of prime factorization of component polynomials of $F$ induce graph distances for distance-regular graphs. These graphs are either the Hamming graphs or Doob graphs using a classification result for distance-regular graphs (if $\PlGH(M)$ is not \#P-hard), after obtaining their intersection numbers. The dichotomy is proved in either case. We then extend the complexity classification to all nonnegative matrices.
Zipeng Ling, Shuliang Liu, Seonil Son, Shenghong Fu, Yuehao Tang, Yao Wan, Xuming Hu
Large language models (LLMs) have become increasingly used for various tasks, often coupled with Chain-of-Thought (CoT) prompting to boost accuracy. Recent work has shown that high label-prediction accuracy does not guarantee correct intermediate reasoning, and the causes of *reasoning flaws* vary from sample to sample, yet existing remedies either focus on a single domain or assume that one flaw type applies uniformly across samples. A simple mitigation method is to provide the model with the correct answer, but we show that this yields no consistent improvement in reasoning quality. This indicates that the problem cannot be fixed by LLMs' awareness of answers, and must instead be addressed through the *structure* of reasoning. Motivated by this, we propose CRAFT (Consensus Reasoning knowledge graph Aggregation for Flaw-aware Trace synthesis), which aggregates the consensus components shared across multiple candidate reasoning traces to synthesize improved ones. CRAFT consistently improves label-prediction accuracy on both logical and mathematical reasoning benchmarks, outperforming most baselines, while its post-processed traces achieve higher quality under fine-grained benchmark evaluation.
Large Language Models (LLMs) are increasingly trained to abstain from answering questions they are unsure about. However, this ability is often misapplied: in real-world applications, user prompts sometimes contain elements of uncertainty, which lead LLMs to abstain even on problems they are capable of solving. We argue that LLM abstention is not only an expression of genuine uncertainty; it can also be an artifact largely shaped by prompts. We name this phenomenon *Abstention Inflation*. We add "Unknown" as an extra option for LLMs to choose from; experiments show serious accuracy drops on True/False Questions (TFQs). Replacing "Unknown" with an unrelated random word produces a similar effect. We argue that LLMs are trained to imitate the surface pattern of abstention, rather than to express genuine uncertainty. Based on ten experimental settings, we support four claims that form a progressive argument: **(C1)** *Abstention Inflation* can be triggered by the presence of an extra option, not by genuine uncertainty; **(C2)** it makes the models deny they can answer, even when they can; **(C3)** it is a later-layer output override, as the reasoning traces and mid-layer representations preserve correct answers; **(C4)** it is not stochastic noise: it results from various factors, emerges through instruction tuning, is boosted by problems' higher difficulty, and can be mitigated at larger model sizes.
Chenglong Ma, Yuanfeng Ji, Junzhi Ning, Jiyao Liu, Yujie Wei, Shaohao Rui, Chaoyang Zhang, Wenjie Li 等 21 位作者
Multimodal medical AI requires a common visual interface for encoding heterogeneous imaging evidence for patient condition modeling. However, existing medical visual tokenizers are often tied to a specific dimension and imaging modality, forcing multimodal systems to rely on fragmented and poorly reusable representations. We introduce MedVTok, a general-purpose visual tokenizer that maps 2D images and 3D volumes from diverse imaging modalities into a unified token space. Building such a tokenizer poses two key challenges: preserving anatomical structure across dimensions and capturing clinical semantics across modalities. To this end, MedVTok introduces (i) slice differential regularization, which explicitly models inter-slice anatomical coherence missing in previous per-pixel/per-voxel supervision, and (ii) multi-expert representation alignment, which integrates knowledge from multiple medical expert encoders while retaining modality-specific complementary knowledge. Trained on 28M images and 70K volumes across 9 imaging modalities, MedVTok supports classification, retrieval, segmentation, synthesis, visual question answering, and medical report generation through a single tokenizer. Across 24 benchmarks, MedVTok achieves state-of-the-art performance, demonstrating a scalable token interface that unifies medical image dimension, modality, and task usage.
Yinghao Tang State key laboratory of CAD&CG, Zhejiang University, Yupeng Xie, Yingchaojie Feng, Tingfeng Lan, Wei Chen Zhejiang University
Interactive articles help readers engage with complex ideas through exploration, yet creating them remains costly, requiring both domain expertise and web development skills. Recent LLM-based agents can automate content creation, but naively applying them yields uncontrollable and unverifiable outputs. We present Vividoc, a human-agent collaborative system that generates interactive educational documents from a single topic input. Vividoc introduces a multi-agent pipeline (Planner, Executor, Evaluator) and the Document Specification (DocSpec), a human-readable intermediate representation that decomposes each interactive visualization into State, Render, Transition, and Constraint components. The DocSpec enables educators to review and refine generation plans before code is produced, bridging the gap between pedagogical intent and executable output. We collect a dataset of 101 real-world interactive documents across 11 domains and conduct a user study showing that ViviDoc produces documents comparable in quality to human-authored ones. Our demo is available at https://vividoc.vercel.app/ and a video demonstration at https://www.youtube.com/watch?v=rJrnPJLyHUI.
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 3: System Demonstrations) · 2026 · 正式发表
Zihao Tang, Xin Yu, Ziyu Xiao, Zengxuan Wen Microsoft, Zelin Li Microsoft, Jiaxi Zhou, Hualei Wang Center for Excellence in Brain Science and Intelligence Technology (Institute of Neuroscience), Chinese Academy of Sciences and University of Chinese Academy of Sciences, Haohua Wang 等 12 位作者
AI Memory, specifically how models organizes and retrieves historical messages, becomes increasingly valuable to Large Language Models (LLMs), yet existing methods (RAG and Graph-RAG) primarily retrieve memory through similarity-based mechanisms. While efficient, such System-1-style retrieval struggles with scenarios that require global reasoning or comprehensive coverage of all relevant information. In this work, We propose Mnemis, a novel memory framework that integrates System-1 similarity search with a complementary System-2 mechanism, termed Global Selection. Mnemis organizes memory into a base graph for similarity retrieval and a hierarchical graph that enables top-down, deliberate traversal over semantic hierarchies. By combining the complementary strength from both retrieval routes, Mnemis retrieves memory items that are both semantically and structurally relevant. Mnemis achieves state-of-the-art performance across all compared methods on long-term memory benchmarks, scoring 93.9 on LoCoMo and 91.6 on LongMemEval-S using GPT-4.1-mini.
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
Yinghao Tang State key laboratory of CAD&CG, Zhejiang University, Xueding Liu University of Electronic Science and Technology of China, Boyuan Zhang, Tingfeng Lan, Yupeng Xie, Jiale Lao, Yiyao Wang, Haoxuan Li Zhejiang University 等 16 位作者
Infographics are composite visual artifacts that combine data visualizations with textual and illustrative elements to communicate information. While recent text-to-image (T2I) models can generate aesthetically appealing images, their reliability in generating infographics remains unclear. Generated infographics may appear correct at first glance but contain easily overlooked issues, such as distorted data encoding or incorrect textual content. We present IGenBench, the first benchmark for evaluating the reliability of text-to-infographic generation, comprising 600 curated test cases spanning 30 infographic types. We design an automated evaluation framework that decomposes reliability verification into atomic yes/no questions based on a taxonomy of 10 question types. We employ multimodal large language models (MLLMs) to verify each question, yielding question-level accuracy (Q-ACC) and infographic-level accuracy (I-ACC). We comprehensively evaluate 10 state-of-the-art T2I models on IGenBench. Our systematic analysis reveals key insights for future model development: (i) a three-tier performance hierarchy with the top model achieving Q-ACC of 0.90 but I-ACC of only 0.49; (ii) data-related dimensions emerging as universal bottlenecks (e.g., Data Completeness: 0.21); and (iii) the challenge of achieving end-to-end correctness across all 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
Zhong-Zhi Li, Xiao Liang, Zihao Tang, Lei Ji, Peijie Wang, Haotian Xu, Xing W, Haizhen Huang 等 14 位作者
Large Language Models (LLMs) have recently achieved remarkable progress on complex reasoning tasks by leveraging extended Chain-of-Thought (CoT) techniques. These reasoning processes can be roughly categorized into System-1 (fast and intuitive) and System-2 (slow and deliberate) paradigms. However, excessive reliance on lengthy System-2-style reasoning during inference can produce extremely long outputs, thereby reducing efficiency. In this work, we propose Thinking Length Data Re-weighting (TLDR), that does not rely on sophisticated data annotations or interpolation between multiple models. We continuously balance the weights between the model’s System-1 and System-2 data to eliminate redundant reasoning processes while preserving the model’s reasoning capability. We validate our method across multiple base models, including Deepseek-R1-Distilled Qwen models, as well as on a diverse benchmarks with varying difficulty levels. Our method significantly reduces the number of output tokens by nearly 40% while maintaining the accuracy of the reasoning.
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
Dezhao Tang, Meihan Liu China University of Mining Technology - Xuzhou, Yulai Tong China University of Mining Technology - Xuzhou, Guan Yuan, Qiuyan Yan China University of Mining Technology - Xuzhou
Dynamic reasoning trees can help large language models solve complex tasks by explicitly structuring intermediate decisions.However, existing approaches often rely on manually specified subproblems or predefined decomposition patterns, which limits the effectiveness of reasoning and generalization.To solve this problem, we propose SyRA, a hierarchical reasoning framework based on MFR theory that supports the construction of adaptive reasoning trees and reliable error correction within a single LLM. Specifically, SyRA focuses on reasoning-tree construction, dynamically controlling branching and expansion using MFR principles to enable informative, non-redundant subproblem decomposition. In addition, it introduces a residual backtracking mechanism for adaptive cross-layer error correction, allowing the model to revise earlier reasoning decisions based on downstream feedback.Across eight reasoning benchmarks, SyRA significantly reduces logical errors and improves reasoning accuracy, while achieving a better balance between accuracy and reasoning time than the Chain-of-Thought, Decompose–Analyze–Rethink and Tree-of-Thought. Our code and dataset are available at https://github.com/Tim798-art/SyRA/tree/main/SyRA.
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
Jiahao Tang Central South University, Henry Hengyuan Zhao, Lijian Wu, Zijian Zhang, Yifei Tao, Dongxing Mao Central South University, Yang Wan, Jingru Tan Central South University 等 11 位作者
We introduce Chart2Code, a new benchmark for evaluating the natural language to chart code generation capabilities of large multimodal models (LMMs). Chart2Code is explicitly designed from a user-driven perspective, capturing diverse real-world scenarios and progressively increasing task difficulty. It consists of three levels: Level 1 (Chart Reproduction) reproduces charts from a reference figure and user query; Level 2 (Chart Editing) involves complex modifications such as changing chart types or adding elements; and Level 3 (Long-Table to Chart Generation) requires models to transform long, unprocessed tables into faithful charts following user instructions. To our knowledge, this is the first hierarchical benchmark that reflects practical chart2code usage while systematically scaling task complexity. In total, Chart2Code contains 2,186 tasks across 22 chart types, paired with multi-level evaluation metrics that assess both code correctness and the visual fidelity of rendered charts. We benchmark 29 state-of-the-art (SoTA) LMMs, including both proprietary and the latest open-source models such as GPT-5.2, Qwen3-VL, InternVL3/3.5, MiMo-VL, and Seed-1.6-VL. Experimental results demonstrate that even the SoTA model GPT-5.2 averages 72.21 on code-based evaluation and only 33.41 on chart-quality assessment across the editing tasks, underscoring the difficulty of Chart2Code. We anticipate this benchmark will drive advances in multimodal reasoning and foster the development of more robust and general-purpose LMMs.
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
Lingxiao Tang, He Ye University College London, University of London, Zhaoyang Chu, Muyang Ye, Zhongxin Liu Zhejiang University, Xiaoxue Ren Zhejiang University, Lingfeng Bao
Code LLMs still struggle with code execution reasoning, especially in smaller models. Existing methods rely on supervised fine-tuning (SFT) with teacher-generated explanations, primarily in two forms: (1) input–output (I/O) prediction chains and (2) natural-language descriptions of execution traces. However, intermediate execution steps cannot be explicitly verified during SFT, so the training objective can reduce to merely matching teacher explanations. Moreover, training data is typically collected without explicit control over task difficulty. We introduce ExecVerify, which goes beyond text imitation by incorporating verifiable white-box rewards derived from execution traces, including next-statement prediction and variable value/type prediction. Our work first builds a dataset with multiple difficulty levels via constraint-based program synthesis. Then, we apply reinforcement learning (RL) to reward correct answers about both intermediate execution steps and final outputs, aligning the training objective with semantic correctness at each execution step. Finally, we adopt a two-stage training pipeline that first enhances execution reasoning and then transfers to code generation. Experiments demonstrate that a 7B model trained with ExecVerify achieves performance comparable to 32B models on code reasoning benchmarks and improves pass@1 by up to 5.9% on code generation tasks over strong post-training baselines.
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
MEMENTO: Teaching LLMs to Manage Their Own Context
Vasilis Kontonis, Yuchen Zeng, Shivam Garg, Lingjiao Chen, Hao Tang, Ziyan Wang, Ahmed Awadallah, Eric Horvitz 等 10 位作者
AI 速览：通过将推理过程分块并压缩为稠密摘要（memento），让模型仅依赖摘要向前推理，从而降低上下文、KV 缓存与计算开销。
Reasoning models think in long, unstructured streams with no mechanism for compressing or organizing their own intermediate state. We introduce MEMENTO: a method that teaches models to segment reasoning into blocks, compress each block into a memento, i.e., a dense state summary, and reason forward by attending only to mementos, reducing context, KV cache, and compute. To train MEMENTO models, we release OpenMementos, a public dataset of 228K reasoning traces derived from OpenThoughts-v3, segmented and annotated with intermediate summaries. We show that a two-stage SFT recipe on OpenMementos is effective across different model families (Qwen3, Phi-4, Olmo 3) and scales (8B--32B parameters). Trained models maintain strong accuracy on math, science, and coding benchmarks while achieving ${\sim}2.5\times$ peak KV cache reduction. We extend vLLM to support our inference method, achieving ${\sim}1.75\times$ throughput improvement while also enabling us to perform RL and further improve accuracy. Finally, we identify a dual information stream: information from each reasoning block is carried both by the memento text and by the corresponding KV states, which retain implicit information from the original block. Removing this channel drops accuracy by 15\,pp on AIME24.
Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent 等 1137 位作者
In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.