跳到正文
研踪PaperTrail

论文库

检索计算机领域论文,按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+1199新增论文
+950其中 AI 相关
+132已有论文修订
+97新增 AI 导读

arXiv 最近完成扫描:2026-11-20 08:48 · 全站最近入库:2026-11-20 16:53各信源进度 →

统计口径与数据新鲜度

统计 2026-11-19 17:51 — 2026-11-20 17:51(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-20 08:48最近发现新增版本:2026-11-20 16:53库内论文最新更新时间:2026-11-20 02:59

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

28,609 已收录论文23,738 AI 相关8,785 有修订112 HF 社区发现

浏览论文库

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 17 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-20 08:46(北京时间)
版本更新 · v2arXiv:2611.24133

Evaluating LIM Trading Agents: A Critical Review of Backtesting Validity and Reporting Practices

Long Xu, Fuyi Wang, Xiaolu Hu, Chao Chen

原文摘要

Large language models (LLMs) are increasingly used in investment research, and many studies report promising historical returns. However, such results are only credible if researchers can demonstrate that the model used only information available at the time, account for how strategies and configurations were selected, and explain how reported returns were generated. We identify six common threats to validity in LLM investment evaluations and apply them to 18 published studies. Across 108 assessments, we find substantial gaps in reporting. No paper fully addresses information timing, model-memory exposure, universe selection, trading costs, or configuration-selection bias. Only four studies provide minimum evidence that results hold across multiple market periods. Our analysis reveals that commonly used safeguards often address different problems than researchers assume. For example, using dated inputs does not guarantee that a pretrained model lacked knowledge of future events, and repeated experimental runs do not correct for strategy-selection bias. We therefore propose an evaluation protocol for future LLM investment research. Rather than judging whether LLMs can generate profits, our study examines whether existing evidence is sufficient to support such claims. We conclude that current reporting practices are generally insufficient for that purpose.

cs.CR

新论文 · v1arXiv:2611.23673

Communicative On-Policy Distillation: Distilling Black-Box Teacher Capabilities through Communication for Long-Horizon Tasks

Lirui Luo, Kelong Mao, Xinwei Yang, Rongqing Li, Luyu Chen, Kieran Wong, Yudong Guo, Xinrui Wang 等 12 位作者

原文摘要

On-policy distillation (OPD) is a powerful paradigm for improving language-model agents on long-horizon tasks, yet its conventional formulation relies on teacher logits that text-only black- box models do not expose. To address this limitation, we introduce Communicative On-Policy Distillation (COPD), which uses natural-language communication as a medium for distilling a stronger model's capabilities into a smaller student. During task execution, the student au- tonomously decides when to consult a black-box teacher and formulates open-ended natural- language requests without predefined question templates or intervention schedules. Successful student-teacher interactions are retained as privileged references to supervise fresh adviser-free student rollouts through on-policy self-distillation, jointly with task-reward reinforcement learning. On CAMG-II, COPD achieves 2.46 times the gain of the strongest black-box OPD baseline. The capabilities it learns through communication generalize to Terminal-Bench 2.1 and SWE-Bench Pro V2, improving task success without adviser access.

cs.LG · cs.AI · cs.CL

版本更新 · v2arXiv:2605.05955

TableVista: Benchmarking Multimodal Table Reasoning under Visual and Structural Complexity

Zheyuan Yang, Liqiang Shang, Junjie Chen, Xun Yang, Chenglong Xu, Bo Yuan, Chenyuan Jiao, Yaoru Sun 等 9 位作者

原文摘要

We introduce TableVista, a comprehensive benchmark for evaluating foundation models in multimodal table reasoning under visual and structural complexity. TableVista consists of 3,000 high-quality table reasoning problems, where each instance is expanded into 10 distinct visual variants through our multi-style rendering and transformation pipeline. This process encompasses diverse scenario styles, robustness perturbations, and vision-only configurations, culminating in 30,000 multimodal samples for a multi-dimensional evaluation. We conduct an extensive evaluation of 29 state-of-the-art open-source and proprietary foundation models on TableVista. Through comprehensive quantitative and qualitative analysis, we find that while evaluated models remain largely stable across diverse rendering styles, they exhibit pronounced performance degradation on complex structural layouts and vision-only settings, revealing that current models struggle to maintain reasoning consistency when structural complexity combines with visually integrated presentations. These findings highlight critical gaps in current multimodal capabilities, providing insights for advancing more robust and reliable table understanding models.

cs.CL · cs.CV

版本更新 · v2arXiv:2610.32664

Zoom Where the Answer Looks: Cross-Resolution Self-Distillation for MLLMs under Visual-Token Budgets

Zhensheng Hu, Guanglong Xu, Guanqiang Hu

原文摘要

Under the small per-view visual-token budgets of efficient deployment, the fine text of documents, charts and infographics becomes illegible to multimodal large language models (MLLMs). Zooming in recovers it, but the region must be chosen before the model answers. Existing methods decode an answer or a box first, at the cost of generation, or read the attention available before answering, which localizes the evidence only in top layers whose depth varies with model size. Learned region proposers instead distill the attention of the model's answers, but on the same budgeted view. Yet this attention is precise only on a view the model can read: on Qwen3-VL-4B, a window around its peak covers 83% of the evidence on a high-resolution view but 54% on the budgeted view. Since answers and high-resolution views are available during training, we introduce LookZoom, a zooming method that trains a dedicated look token by cross-resolution self-distillation of the answer's attention: with the MLLM frozen, the token learns to attend on the budgeted view, before any answer, where the model's own correct answers attend on a high-resolution view. The region around its attention peak is the window to zoom into, and reinforcement learning then trains the MLLM to answer from both views. With 98 visual tokens per view on five text-rich benchmarks, LookZoom surpasses attention-guided zooming by 20-24 points and grounding-based zooming and learned crop routing by 12-15 points on Qwen3-VL at three scales, and it outperforms learned region proposers at every budget from 98 to 392 tokens. On Qwen3-VL, the look token alone, with the MLLM frozen, already surpasses every baseline that leaves the language model unchanged, and the recipe carries over to InternVL3.5 and Qwen2.5-VL.

cs.CV

版本更新 · v2arXiv:2611.14009

Coding Agent Memory Post-Training II: Unlocking the Memory Potential of Pre-trained File Operations for Long-Horizon Tasks via On-Policy Distillation

Lirui Luo, Kelong Mao, Xinwei Yang, Rongqing Li, Luyu Chen, Kieran Wong, Yudong Guo, Xinrui Wang 等 13 位作者

原文摘要

Language-model agents increasingly tackle long-horizon tasks whose interaction histories exceed the model's active context. They must therefore decide what to write down and what to read back, and when to do either. Recent work has begun to use on-policy distillation (OPD) to make memory control part of the policy, relying on memory tools and compaction protocols that lie outside either model's pre-training. However, this weakens the distillation on both sides: the teacher cannot give useful guidance on an interface it has never used, and the student samples too few meaningful memory actions for that guidance to take effect. To study this problem, we introduce Coding Agent Memory Gym II (CAMG-II), a suite of long-horizon agentic-RL environments spanning Shop, Coding, AutoResearch, General, and Visual. Alongside each environment's native task interface, CAMG-II provides ordinary Bash access and an episode-persistent workspace, enabling the policy to create, revise, search, and reuse files as memory across context replacements. We also introduce Coding Agent Memory Post-Training II (CAMP-II), which trains a single policy jointly across all five environments with fully asynchronous OPD to learn this file-based memory behavior directly from a frozen teacher with a prior for file operations. On CAMG-II, CAMP-II obtains 2.97 times the gain of the strongest OPD baseline. The memory behavior CAMP-II learns generalizes to two external benchmarks, Terminal-Bench 2.1 and SWE-Bench Pro V2, with improved success rates.

cs.LG · cs.AI · cs.CL

版本更新 · v5arXiv:2608.28553

Four conditions for composing plugin-based agents without a host agents

Hanzhang Jia, Liheng Zeng, Hao Cheng, Yi Gao, Bo Ma, Guolong Xu

原文摘要

Large language model agents now run long-horizon, multi-tool, autonomous sessions, and the infrastructure must keep the current assembly of capabilities, the progress of execution, and the basis for recovery after a fault. A single-process host holds this information in its own running state, so its death takes the session record and the assembly with it. This paper gives four conditions under which a plugin-based agent needs no host process, one each on state placement, record carriers, recovery coordination, and provider resolution. The store, writer, and delivery disciplines keep the conditions in force, and each carries formal support from the spatiotemporal-composability calculus and from the proofs in the supplementary materials. An agent that meets them can be constructed with its assembly on a bus of peer processes and its record in an append-only carrier owned by no process, and removing any one part makes the assembly need a host again. The conditions themselves are the contribution, and every implementation that meets them inherits the same advantages. Such an agent keeps its session through the death of any process, wastes no calls after a fault, loses no records, recovers locally, changes its capability set while it runs, and carries its undo semantics across processes. This paper builds one SDK of the conditions, Logos, and measures whether it fulfills their promise against the single-process reference and a configuration over MCP under one fault in each of 200 tasks, realized in 197. The reference lost all 197 faulted sessions, the MCP configuration kept its sessions but issued 1,212 wasted calls, and the SDK kept every session, wasted no call, lost no record, and fulfilled the promise of the conditions in every faulted task, with task performance preserved. The conditions give builders a tested route from single-process hosts to peer processes.

cs.AI · cs.MA

版本更新 · v3arXiv:2610.18654

Kwai Summary Attention V2 Technical Report

OneRec Team, Chenglong Chu, Guohong Mu, Guowang Zhang, Hao Peng, Hui Wang, Jian Liang, Jiangxia Cao 等 33 位作者

原文摘要

Summary-token architectures such as Kwai Summary Attention (KSA) compress distant context into learnable Summary token per chunk, and hybrid variants interleave these layers with full attention to retain exact access to the original text. In such hybrid KSA-Full architecture, the interleaving inserted summary tokens are served as a streaming flow role across different layers. Although effective, however, it also introduce some bottlenecks: (1) the different layers summary compression chunk must be same, therefore hard to reach a flexible KV Cache compression, e.g., different layer set different summary compression chunk window. (2) in the full-attention layers, the summary token introduced additional computation cost, but such layer ignore them and still pay quadratic prefilling, therefore unfriendly to the computation side. Motivated by them, we developed the KSA V2, which not only achieves a more flexible KV cache memory compression but also reduces the computation cost in Full attention layer. Specifically, we found that the summary tokens can serve two roles: the KV memory and sparse indexer.

cs.LG · cs.AI · cs.IR

版本更新 · v2arXiv:2611.07644

Neural Image Compression with Internal-External Prior Fusion and Alignment

Tiancheng Si, Wenlong Xu, Dahua Gao

原文摘要

Existing context models for learned image com- pression rely primarily on internal priors from the current image, overlooking common patterns and typical textures shared across natural images. To address this limitation, we propose an Internal-External Prior Fusion-Alignment (IE-PFA) context model built upon quadtree spatial-channel joint modeling, where a learnable dictionary shared by the encoder and decoder serves as an external prior. Specifically, an RWKV-based Spatial- Channel Prior Fusion (RSCPF) module efficiently fuses inter- nal and external priors along spatial and channel dimensions with linear complexity, while a training-only Self-Distilled Prior Alignment (SDPA) mechanism aligns them without inference overhead. Moreover, our model supports variable-rate coding over 64 quality levels within a single network and is accelerated by custom CUDA operators. Experiments on seven benchmarks demonstrate state-of-the-art rate-distortion performance, achiev- ing an average BD-Rate saving of 21.65% over VTM-22.0-Intra, with only 29.9M parameters and the fastest coding speed among advanced neural codecs (61 ms encoding and 52 ms decoding). Code is available at https://github.com/stcnjtech/DCVC-X-Intra.

cs.CV · cs.AI

版本更新 · v3arXiv:2606.09864

Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation

Bruce Changlong Xu, Adarsh Kumarappan, Mu Zhou

原文摘要

Key-value (KV) cache quantization is widely used to reduce LLM inference memory, yet existing evaluations focus only on perplexity and accuracy without assessing the safety impact. In this study, we explore alignment preservation under KV cache quantization. Across eleven instruction-tuned models (3.8B-72B) and five benchmarks (1,894 prompts), we find that low-bit quantization can silently destroy safety alignment: Mistral-7B loses 15.2% of its refusals at only 1.03x perplexity, and no universal safe bit-width exists, with sharp model-specific phase transitions invisible to standard metrics. We identify that the root cause is geometric: safety features occupy a low-dimensional activation subspace $10^2$-$10^3\times$ more vulnerable to quantization noise than the full representation space perplexity averages over. We propose Per-Channel Reduction (PCR), a diagnostic that classifies each model into one of three mechanistic failure modes: outlier-crushes-safety, where safety lives in non-outlier channels collaterally damaged by outlier-driven scale factors; outlier-as-safety, where safety overlaps outlier channels and finer granularity cannot rescue it; and multi-layer dilution, where safety is distributed across many layers and per-layer fixes fail. PCR predicts the correct mitigation direction on all nine primary models and one held-out model from an independent family using 20 calibration prompts. PCR generalizes across unseen prompts, models, and quantizer families, including production quantizers (KIVI: up to 97.2% recovery) and rotation-based quantization (QuaRot-style Hadamard), succeeding where attention-based allocation methods fail. The resulting training-free protocol, requiring ~35 GPU-minutes, recovers up to 97% of lost alignment at under 1% latency and minimal memory overhead, addressing vulnerabilities confirmed in production vLLM serving with FP8 KV cache on NVIDIA GPUs.

cs.LG · cs.AI · cs.ET

版本更新 · v3arXiv:2601.01558

Utilizing Earth Foundation Models to Enhance the Simulation Performance of Hydrological Models with AlphaEarth Embeddings

Pengfei Qu, Lei Ye, Wenyu Ouyang, Chi Zhang, Yikai Chai, Shuolong Xu, Yongri Piao, Miao Zhang 等 9 位作者

原文摘要

Accurate prediction in ungauged basins (PUB) requires effective transfer across hydrologically diverse regions. While deep learning has advanced large-sample hydrological modeling, its cross-regional generalization depends in part on the static basin descriptors used to guide model behavior. Earth Foundation Models (EFMs) provide a promising alternative by learning environmental representations from multi-modal satellite data. In this study, we evaluate the 64-dimensional AlphaEarth Foundations (AEF) embeddings as a new form of static basin representation and assess their value for streamflow prediction under both in-sample (IS) and out-of-sample (OOS) settings using the CAMELS dataset. Results show that AEF embeddings offer higher OOS robustness than CAMELS attributes, despite nearly identical IS performance. Mutual-information analysis reveals statistical associations between AEF embeddings and conventional catchment attributes related to terrain, vegetation, and rooting characteristics. We further investigate similarity-based donor-basin selection for PUB using CAMELS attributes and AEF embeddings. Across 100 target basins, AEF static inputs generally improved PUB performance. With AEF inputs, similarity-based donor selection outperformed random selection for smaller donor pools, but its advantage diminished as pool size increased. Adding more donor basins did not consistently improve performance. Leave-one-cluster-out experiments showed AEF input benefits depended on basin partitioning, suggesting that transferability remains sensitive to the catchment conditions represented during training. Overall, this study demonstrates that Earth Foundation Model embeddings provide a transferable representation for PUB prediction and underscores the importance of basin representation quality and donor pool design, motivating future research toward unified hydrological foundation models.

cs.LG · cs.AI

版本更新 · v2arXiv:2610.33020

Mine or Yours? Self-Other Confusion in Latent LLM Communication

Yunlong Xu

原文摘要

If two minds were linked, would each still know which thoughts are its own? Language models are increasingly linked in this way: instead of exchanging text, one model reads another's key-value cache. Systems that join a foreign cache to the receiver's own use one of two layouts: the cache lies over the receiver's own context, at the same token positions, or in front of it. We give two copies of one model different private priorities and link them in both layouts. With the cache over its own context, the receiver gives a fact found only in its partner's cache in every answer. Yet in about half of its answers it gives the partner's priority as its own, in both Qwen3-4B and Llama-3.1-8B: as often as a receiver that ignored whose priority was whose, although the text in the cache names its owner. With the same cache moved in front, this falls to 5% in Qwen but only to 43% in Llama, although both models nearly stop naming the partner's priority as the one they are using. Where the cache sits matters, yet neither layout in use protects both models. Fewer such errors do not mean better attribution: once the cache is moved, the receiver names the partner's priority as the partner's less often, not more, and its report of an inserted memory cannot tell the partner's cache from a copy of its own. In short, shared caches deliver content but not ownership.

cs.CL · cs.AI

版本更新 · v3arXiv:2603.01563

Efficient Token-Level Policy Optimization for Masked Diffusion Models via Velocity Alignment

Chenxing Wei, Jiazhen Kang, Hong Wang, Jianqing Zhang, Hao Jiang, Xiaolong Xu, Ningyuan Sun, Ying He 等 11 位作者

原文摘要

Reinforcement Learning with Verifiable Rewards (RLVR) successfully aligns autoregressive models in rigorous domains like mathematical reasoning and code generation. However, applying RLVR to Diffusion Large Language Models (dLLMs) is hindered by intractable sequence-level likelihoods, forcing existing methods to rely on unstable, high-variance approximations. We propose Sequence Likelihood-Free Policy Optimization (LFPO), an efficient framework that circumvents this bottleneck by formulating policy alignment as discrete velocity alignment. Inspired by continuous flow matching, LFPO eschews trajectory-level estimation and directly optimizes denoising logits via contrastive token-level updates, providing a stable, low-variance surrogate. Furthermore, predicting terminal solutions from intermediate states implicitly straightens the generation trajectory, enabling high-quality sampling with significantly fewer denoising iterations. Extensive experiments on reasoning and coding benchmarks demonstrate that LFPO matches or exceeds the accuracy of likelihood-based baselines while substantially accelerating inference.

cs.LG · cs.AI

新论文 · v1arXiv:2610.35001

More Memory, Better Videos? A Controlled Study of Visual Memory in Long-Video Generation

Yuzhuo Chen, Huafeng Shi, Xinyu Wang, Yulong Xu, Yucheng Wang, Haoqin Hong, Guoxin Zhang, Zehua Ma 等 9 位作者

原文摘要

Long-video generators increasingly retrieve past visual observations to prevent identity drift. The intuition is compelling: more access to history should make future video more consistent. But an image can be faithful to the past and wrong for the present. Does visual memory always improve the generated story? We investigate this question through VMem-Bench, a two-track diagnostic study. Track-A measures which references are selected from 92 real films (16,548 segments), before generation and without generator confounds. Track-B compares four retrieval policies with no-memory and gold-informed last-seen controls on 30 causally generated stories (2,824 segments) using a shared generator. Retrieval increases entity recall from 0.559 to 0.628-0.693 and improves similarity to past appearances, but decreases precision from 0.977 to 0.915-0.938 and forbidden-entity avoidance from 0.968 to 0.845-0.898. It can also reinforce obsolete appearances after the story changes them. Conversely, high long-gap recall can arise without memory when the prompt itself supplies the appearance; an appearance-stripping analysis exposes this confound. These results distinguish historical fidelity from current-story correctness and show why retrieval relevance or final-video consistency alone does not establish memory utility.

cs.CV

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.254ACL Anthology · ACL 2026

Incentivizing Parametric Knowledge via Reinforcement Learning with Verifiable Rewards for Cross-Cultural Entity Translation

Jiang Zhou, Xiaohu Zhao, Xinwei Wu, Tianyu Dong, Hao Wang, Yangyang Liu, Heng Liu, Linlong Xu 等 11 位作者

原文摘要

Cross-cultural entity translation remains challenging for large language models (LLMs) as literal or phonetic renderings are usually yielded instead of culturally appropriate translations in context. However, relevant knowledge may already be encoded in model parameters during large-scale pre-training. To incentivize the effective use of parametric knowledge, we propose EA-RLVR (Entity-Anchored Reinforcement Learning with Verifiable Rewards), a training framework that optimizes cross-cultural entity translation without relying on external knowledge bases. EA-RLVR anchors supervision on a verifiable, entity-level reward signal and incorporates lightweight structural gates to stabilize optimization. This design steers the model toward learning a robust reasoning process rather than merely imitating reference translations. We evaluate EA-RLVR on XC-Translate and observe consistent improvements in both entity translation accuracy and out-of-domain generalization. Specifically, training on merely 7k samples boosts Qwen3-14B’s entity translation accuracy from 23.66% to 31.87% on a 50k test set comprising entirely unseen entities. The learned entity translation ability also transfers to general translation, yielding +1.35 XCOMET on WMT24pp, which scales to +1.59 with extended optimization. Extensive analyses of pass@k dynamics and reward formulations attribute these gains to superior sampling efficiency and a stable optimization landscape.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.283ACL Anthology · ACL 2026

From Insight to Action: A Novel Framework for Interpretability-Guided Data Selection in Large Language Models

Ling Shi, Xinwei Wu, Xiaohu Zhao, Hao Wang, Heng Liu, Yangyang Liu, Linlong Xu, Longyue Wang Alibaba Group 等 10 位作者

原文摘要

While mechanistic interpretability tools like Sparse Autoencoders (SAEs) can uncover meaningful features within Large Language Models (LLMs), a critical gap remains in transforming these insights into practical actions for model optimization. We bridge this gap with the hypothesis that data selection guided by a model’s internal task features is a effective training strategy. Inspired by this, we propose Interpretability-Guided Data Selection (IGDS), a framework that first identifies these causal task features through frequency recall and interventional filtering, then selects “Feature-Resonant Data” that maximally activates task features for fine-tuning. We validate IGDS on mathematical reasoning, summarization, and translation tasks within Gemma-2, LLaMA-3.1, and Qwen3 models. Our experiments demonstrate exceptional data efficiency: on the Math task, IGDS surpasses full-dataset fine-tuning by a remarkable 17.4% on Gemma-2-2B while using only 50% of the data, and outperforms established baselines focused on data quality and diversity. Analysis confirms a strong positive correlation between feature amplification and task performance improvement. IGDS thus provides a direct and effective framework to enhance LLMs by leveraging their internal mechanisms, validating our core hypothesis.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.469ACL Anthology · ACL 2026

M^2PO: Multi-Perspective Multi-Pair Preference Optimization for Machine Translation

Hao Wang, Linlong Xu, Heng Liu, Yangyang Liu, Xiaohu Zhao, Bo Zeng, Liangying Shao, Yichen Dong 等 14 位作者

原文摘要

Aligning Large Language Models (LLMs) to human preferences is pivotal for Machine Translation (MT), yet current approaches are often hindered by misleading reward signals. Our analysis reveals that prevailing Quality Estimation (QE) models exhibit a systematic blind spot towards partial errors—specifically partial hallucinations and omissions—often favoring superficially fluent but unfaithful translations. To address this, we propose M^2PO (Multi-Perspective Multi-Pair Preference Optimization), a data-centric framework for preference optimization in machine translation. First, to correct the bias towards fluency, M^2PO uses a multi-perspective alignment mechanism that decouples semantic fidelity from fluency, prioritizing faithfulness via a curriculum strategy. Second, with the bias corrected, partial errors fall between perfect and severely incorrect translations, making them inefficient to learn via standard best-versus-worst comparisons. We thus introduce a multi-pair objective that leverages the full candidate list to capture these fine-grained error signals. Experiments on WMT23, WMT24, and FLORES-200 show that M^2PO enables a 9B model to outperform leading open-source baselines and achieve parity with proprietary models like GPT-4o and Gemini-2.0-Flash, demonstrating significant potential for efficient, high-fidelity LLM-based translation.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.844ACL Anthology · ACL 2026

Curing Miracle Steps in LLM Mathematical Reasoning with Rubric Rewards

Youliang Yuan The Chinese University of Hong Kong-Shenzhen, Qiuyang Mang, Jingbang Chen The Chinese University of Hong Kong, Hong Wan, Xiaoyuan Liu Xiaohongshu and The Chinese University of Hong Kong, Shenzhen, Junjielong Xu Tencent and The Chinese University of Hong Kong, Shenzhen, Jen-tse Huang, Wenxuan Wang Renmin University of China 等 10 位作者

原文摘要

In this paper, we observe that current models are susceptible to reward hacking, leading to a substantial overestimation of a model’s reasoning ability. This is evidenced by a high incidence of “false positives”—solutions that reach the correct answer through an unsound process.Through a systematic analysis with human verification, we establish a taxonomy of these failure modes, identifying patterns like Miracle Steps—abrupt jumps to a correct output without a valid preceding derivation. Probing experiments suggest that these Miracle Steps are linked to answer-recall shortcuts, including memorization from pretraining, where the model accesses the correct answer independently of its reasoning chain.To mitigate this systemic issue, we introduce the Rubric Reward Model (RRM), a process-oriented reward function that evaluates the entire reasoning trajectory against problem-specific rubrics.The RRM explicitly penalizes logical flaws and encourages rigorous deduction.When integrated into an RL pipeline, RRM-based training consistently outperforms outcome-only supervision across four math benchmarks.Notably, it boosts Verified Pass@1024 on AIME2024 from 26.7% to 62.6% and reduces the incidence of Miracle Steps by 71%.Our work demonstrates that rewarding the solution process is crucial for building accurate and reliable models.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表