pith. machine review for the scientific record. sign in

arxiv: 2605.12384 · v1 · submitted 2026-05-12 · 💻 cs.CL · cs.AI· cs.LG

Recognition: no theorem link

Scalable Token-Level Hallucination Detection in Large Language Models

Chao Du, Minhao Cheng, Rui Min, Tianyu Pang, Yi R. Fung

Pith reviewed 2026-05-13 05:44 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.LG
keywords hallucination detectiontoken-level analysislarge language modelsreasoning taskssynthetic data generationmodel scalingLLM evaluationerror detection
0
0 comments X

The pith

A training pipeline lets even small models detect token-level hallucinations in LLM reasoning better than much larger models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to build token-level detectors that can flag hallucinations directly in free-form text from large language models, especially in reasoning tasks where errors like logical flaws appear coherent. It does this through a data engine that generates large-scale annotations at scale and a training approach that weights important tokens. If the method works, it removes the need for manual step segmentation or text reformatting that limits current detectors. The experiments indicate performance rises steadily as the detector grows from 0.6B to 8B parameters, with the smallest trained version already beating much larger reasoning models.

Core claim

TokenHD supplies a complete pipeline that first uses a scalable data engine to synthesize hallucination annotations and then applies an importance-weighted training recipe, allowing the resulting detector to label errors token by token in unrestricted LLM output and to reach performance levels that increase reliably with detector size.

What carries the argument

The TokenHD pipeline, which combines a scalable synthesis engine for hallucination annotations with an importance-weighted training procedure to produce detectors that label individual tokens in free-form text without any step segmentation.

If this is right

  • Detectors can scan free-form text outputs without any predefined step breaks or reformatting.
  • Detection accuracy increases steadily as the detector model size grows from 0.6B to 8B parameters.
  • A 0.6B detector trained this way can exceed the hallucination detection performance of much larger reasoning models.
  • The detectors maintain effectiveness across a range of practical application scenarios.
  • Additional training adjustments can further improve performance on new domains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the synthesis method produces unbiased labels, the same engine could be adapted to flag other error types such as factual mistakes outside reasoning chains.
  • Placing the detector inside the generation loop could enable early correction of erroneous tokens before full outputs are produced.
  • The observed scaling suggests that detectors at 70B or beyond might reach levels of reliability sufficient for high-stakes verification tasks.
  • Combining token-level signals with existing sentence-level or external-knowledge checks could yield hybrid systems with higher overall coverage.

Load-bearing premise

The annotations created by the data engine accurately reflect genuine hallucinations that occur in real reasoning outputs rather than artifacts introduced during synthesis.

What would settle it

A fresh test set of human-annotated token errors drawn from actual LLM reasoning traces where the trained detector shows markedly lower agreement with the human labels than the paper reports would falsify the performance claim.

Figures

Figures reproduced from arXiv: 2605.12384 by Chao Du, Minhao Cheng, Rui Min, Tianyu Pang, Yi R. Fung.

Figure 1
Figure 1. Figure 1: An illustration of the token-level detection mecha [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: We report Sincor across three STEM benchmarks. Qwen3-1.7/8B are backbone models, GPT-4.1 and o4-mini are critic models, and TOKENHD-1.7/8B are our trained hallucination detectors. Generalization Across STEM Task Domains. While previous results show that our detector performs well on pure mathematical tasks, many real-world questions require reasoning in other STEM domains, such as science and finance. We t… view at source ↗
Figure 3
Figure 3. Figure 3: Detection performance under two ensemble [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Detection performance under two training [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Detection performance across diverse policy mod [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Human annotation quality assessment (1–5 scale). Annotators rated GT annotations on accuracy [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Detection performance across different training data scales on mathematical tasks. [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Prompts used for identifying hallucinations in mathematical and STEM tasks. [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Prompts used for identifying hallucinations in code generation tasks. [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Prompts used for restoring the identified hallucinated text to match the original response. [PITH_FULL_IMAGE:figures/full_fig_p028_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Prompt used for self-correction with token-level hints. Suspected error regions identified by [PITH_FULL_IMAGE:figures/full_fig_p029_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Average Sincor across seven benchmarks for TOKENHD-1.7B and TOKENHD-8B, grouped by response length. Samples are partitioned by absolute token count into three bins: <500, 500–1000, and >1000 [PITH_FULL_IMAGE:figures/full_fig_p029_12.png] view at source ↗
read the original abstract

Large language models (LLMs) have demonstrated remarkable capabilities, but they still frequently produce hallucinations. These hallucinations are difficult to detect in reasoning-intensive tasks, where the content appears coherent but contains errors like logical flaws and unreliable intermediate results. While step-level analysis is commonly used to detect internal hallucinations, it suffers from limited granularity and poor scalability due to its reliance on step segmentation. To address these limitations, we propose TokenHD, a holistic pipeline for training token-level hallucination detectors. Specifically, TokenHD consists of a scalable data engine for synthesizing large-scale hallucination annotations along with a training recipe featuring an importance-weighted strategy for robust model training. To systematically assess the detection performance, we also provide a rigorous evaluation protocol. Through training within TokenHD, our detector operates directly on free-form text to identify hallucinations, eliminating the need for predefined step segmentation or additional text reformatting. Our experiments show that even a small detector (0.6B) achieves substantial performance gains after training, surpassing much larger reasoning models (e.g., QwQ-32B), and detection performance scales consistently with model size from 0.6B to 8B. Finally, we show that our detector can generalize well across diverse practical scenarios and explore strategies to further enhance its cross-domain generalization capability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces TokenHD, a pipeline for training token-level hallucination detectors on free-form LLM outputs. It consists of a scalable synthetic data engine to generate large-scale token-level hallucination annotations, combined with an importance-weighted training recipe, and a rigorous evaluation protocol. The central claims are that even a 0.6B-parameter detector trained via TokenHD substantially outperforms much larger reasoning models (e.g., QwQ-32B) on hallucination detection in reasoning-intensive tasks, that detection performance scales consistently with model size from 0.6B to 8B, and that the detector generalizes well across diverse practical scenarios.

Significance. If the synthetic annotations prove faithful to genuine logical and factual errors, the work would offer a scalable alternative to step-level detection methods, enabling fine-grained, segmentation-free hallucination detection. The reported outperformance by small models and consistent scaling would represent a notable empirical finding for LLM reliability, provided the evaluation protocol includes appropriate controls and the results hold beyond the synthetic distribution.

major comments (1)
  1. [§3] §3 (TokenHD Pipeline, data engine subsection): The synthesis process for token-level hallucination annotations is not validated against human-annotated reasoning errors (no reported inter-annotator agreement, human-synthetic label correlation, or ablation removing synthesis-specific features). This is load-bearing for the central claim because the headline result (0.6B detector surpassing QwQ-32B) assumes the labels reflect real errors rather than artifacts such as unnatural token distributions or stylistic cues introduced by the engine.
minor comments (1)
  1. [Abstract] Abstract: The performance claims would be strengthened by including at least one concrete metric (e.g., F1 or AUC) and the exact baselines used, rather than qualitative statements such as 'substantial performance gains'.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for highlighting the importance of validating the synthetic annotation process. We address the major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [§3] §3 (TokenHD Pipeline, data engine subsection): The synthesis process for token-level hallucination annotations is not validated against human-annotated reasoning errors (no reported inter-annotator agreement, human-synthetic label correlation, or ablation removing synthesis-specific features). This is load-bearing for the central claim because the headline result (0.6B detector surpassing QwQ-32B) assumes the labels reflect real errors rather than artifacts such as unnatural token distributions or stylistic cues introduced by the engine.

    Authors: We agree that the absence of direct human validation for the synthetic labels represents a gap in the current manuscript, and that this validation is important for supporting the claim that small detectors outperform larger reasoning models on genuine errors. In the revised version, we will expand §3 with a new human evaluation subsection. This will include: (1) sampling 500 synthetic instances across reasoning tasks, (2) recruiting multiple expert annotators to label token-level hallucinations using guidelines aligned with the synthesis engine, (3) reporting inter-annotator agreement via Fleiss' kappa, and (4) computing correlation metrics (e.g., token-level accuracy and Cohen's kappa) between human and synthetic labels. We will also add an ablation study training a detector variant on data stripped of synthesis-specific features (such as targeted error injection patterns) to quantify their contribution versus potential artifacts. These additions will directly address concerns about label fidelity and strengthen the empirical foundation for the scaling and outperformance results. revision: yes

Circularity Check

0 steps flagged

No circularity in empirical pipeline or claims

full rationale

The paper describes an empirical pipeline (TokenHD) for synthesizing token-level hallucination annotations via a data engine, followed by model training with an importance-weighted strategy and evaluation under a stated protocol. No mathematical derivations, equations, or first-principles results appear that reduce performance claims to fitted parameters, self-definitions, or self-citation chains. The scaling results and comparisons (e.g., 0.6B detector vs. larger models) are presented as outcomes of training on the synthesized data and testing, not tautological. The work is self-contained against its own benchmarks and external evaluation protocol, with no load-bearing steps that collapse by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review limits visibility into exact parameters; the approach implicitly assumes synthesized annotations are faithful proxies for real hallucinations and that token-level signals are sufficient for detection.

axioms (1)
  • domain assumption Hallucinations in reasoning tasks can be reliably synthesized at token granularity to create training data
    The data engine component depends on this to generate large-scale annotations.
invented entities (1)
  • TokenHD pipeline no independent evidence
    purpose: Holistic training system for token-level hallucination detectors
    New method introduced by the paper combining data synthesis and weighted training.

pith-pipeline@v0.9.0 · 5537 in / 1228 out tokens · 109734 ms · 2026-05-13T05:44:14.271033+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 62 canonical work pages · 14 internal anchors

  1. [1]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024. 1

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  2. [2]

    Deepseek-r1 incentivizes reasoning in llms through reinforcement learning.Nature, 645(8081):633–638, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning.Nature, 645(8081):633–638, 2025. 1, 4.1

    work page 2025

  3. [3]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025. 1

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al. Qwen2. 5-coder technical report.arXiv preprint arXiv:2409.12186, 2024. 1

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    Claude sonnet 4.5, 2025

    Anthropic. Claude sonnet 4.5, 2025. 1, 5.1

    work page 2025

  6. [6]

    Survey of hallucination in natural language generation.ACM computing surveys, 55(12):1–38, 2023

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation.ACM computing surveys, 55(12):1–38, 2023. 1

    work page 2023

  7. [7]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM Transactions on Information Systems, 43(2):1–55, 2025

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM Transactions on Information Systems, 43(2):1–55, 2025. 1, 6

    work page 2025

  8. [8]

    Why Language Models Hallucinate

    Adam Tauman Kalai, Ofir Nachum, Santosh S Vempala, and Edwin Zhang. Why language models hallucinate.arXiv preprint arXiv:2509.04664, 2025. 1, 6

    work page internal anchor Pith review arXiv 2025

  9. [9]

    Hallucination detection: Robustly discerning reliable answers in large language models

    Yuyan Chen, Qiang Fu, Yichen Yuan, Zhihao Wen, Ge Fan, Dayiheng Liu, Dongmei Zhang, Zhixu Li, and Yanghua Xiao. Hallucination detection: Robustly discerning reliable answers in large language models. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 245–255, 2023. 1, 6 Scalable Token-Level Hallucination ...

    work page 2023

  10. [10]

    Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models

    Potsawee Manakul, Adian Liusie, and Mark Gales. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. InProceedings of the 2023 conference on empirical methods in natural language processing, pages 9004–9017, 2023. 1, 6

    work page 2023

  11. [11]

    Real-time detection of hallucinated entities in long-form generation.arXiv preprint arXiv:2509.03531,

    Oscar Obeso, Andy Arditi, Javier Ferrando, Joshua Freeman, Cameron Holmes, and Neel Nanda. Real-time detection of hallucinated entities in long-form generation.arXiv preprint arXiv:2509.03531,

    work page arXiv

  12. [12]

    Retrieval-Augmented Generation for Large Language Models: A Survey

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2(1), 2023. 1, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  13. [13]

    Search-o1: Agentic Search-Enhanced Large Reasoning Models

    Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Search-o1: Agentic search-enhanced large reasoning models.arXiv preprint arXiv:2501.05366,

    work page internal anchor Pith review arXiv

  14. [14]

    Let’s verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. InThe Twelfth International Conference on Learning Representations, 2023. 1, 6

    work page 2023

  15. [15]

    Math-shepherd: Verify and reinforce llms step-by-step without human annotations

    Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9426–9439, 2024. 1, 6

    work page 2024

  16. [16]

    The lessons of developing process reward models in mathematical reasoning.arXiv preprint arXiv:2501.07301, 2025

    Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. The lessons of developing process reward models in mathematical reasoning.arXiv preprint arXiv:2501.07301, 2025. 1, 6

    work page arXiv 2025

  17. [17]

    Lillicrap, Kenji Kawaguchi, and Michael Shieh

    Yuxi Xie, Anirudh Goyal, Wenyue Zheng, Min-Yen Kan, Timothy P Lillicrap, Kenji Kawaguchi, and Michael Shieh. Monte carlo tree search boosts reasoning via iterative preference learning.arXiv preprint arXiv:2405.00451, 2024. 1

    work page arXiv 2024

  18. [18]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025. 4.1

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    Measuring Mathematical Problem Solving With the MATH Dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset.arXiv preprint arXiv:2103.03874, 2021. 4.1, 4.1

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  20. [20]

    arXiv preprint arXiv:2505.16400 , year=

    Yang Chen, Zhuolin Yang, Zihan Liu, Chankyu Lee, Peng Xu, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Acereason-nemotron: Advancing math and code reasoning through reinforcement learning.arXiv preprint arXiv:2505.16400, 2025. 4.1

    work page arXiv 2025

  21. [21]

    Marthe Ballon, Brecht Verbeken, Vincent Ginis, and Andres Algaba

    Alon Albalak, Duy Phung, Nathan Lile, Rafael Rafailov, Kanishk Gandhi, Louis Castricato, Anikait Singh, Chase Blagden, Violet Xiang, Dakota Mahan, et al. Big-math: A large-scale, high-quality math dataset for reinforcement learning in language models.arXiv preprint arXiv:2502.17387, 2025. 4.1

    work page arXiv 2025

  22. [22]

    GPT-4o System Card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276,

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Qwq-32b: Embracing the power of reinforcement learning, 2025

    Qwen Team. Qwq-32b: Embracing the power of reinforcement learning, 2025. 4.1 Scalable Token-Level Hallucination Detection in LLMs13

    work page 2025

  24. [24]

    GPT-4 Technical Report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4.1

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  25. [25]

    Introducing openai o3 and o4-mini, 2025

    OpenAI Team. Introducing openai o3 and o4-mini, 2025. 4.1, 4.1

    work page 2025

  26. [26]

    Aime-2024, 2024

    AIME-2024. Aime-2024, 2024. 4.1

    work page 2024

  27. [27]

    Aime-2025, 2025

    AIME-2025. Aime-2025, 2025. 4.1

    work page 2025

  28. [28]

    Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems

    Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, et al. Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Paper...

    work page 2024

  29. [29]

    Gpqa: A graduate-level google-proof q&a benchmark

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. InFirst Conference on Language Modeling, 2024. 4.1

    work page 2024

  30. [30]

    Finqa: A dataset of numerical reasoning over financial data

    Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan R Routledge, et al. Finqa: A dataset of numerical reasoning over financial data. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3697–3711, 2021. 4.1

    work page 2021

  31. [31]

    Introducing gpt-5, 2025

    OpenAI Team. Introducing gpt-5, 2025. 4.1

    work page 2025

  32. [32]

    Introducing gemini 2.0: our new ai model for the agentic era, 2024

    Google. Introducing gemini 2.0: our new ai model for the agentic era, 2024. 5.1

    work page 2024

  33. [33]

    Claude 3.5 sonnet, 2024

    Anthropic. Claude 3.5 sonnet, 2024. 5.1

    work page 2024

  34. [34]

    CodeElo: Benchmarking competition-level code generation of LLMs with human-comparable elo ratings.arXiv preprint arXiv:2501.01257, 2025

    Shanghaoran Quan, Jiaxi Yang, Bowen Yu, Bo Zheng, Dayiheng Liu, An Yang, Xuancheng Ren, Bofei Gao, Yibo Miao, Yunlong Feng, et al. Codeelo: Benchmarking competition-level code generation of llms with human-comparable elo ratings.arXiv preprint arXiv:2501.01257, 2025. 5.1

    work page arXiv 2025

  35. [35]

    LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code.arXiv preprint arXiv:2403.07974, 2024. 5.1

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    Opencodereasoning: Advancing data distillation for competitive coding.arXiv preprint arXiv:2504.01943,

    Wasi Uddin Ahmad, Sean Narenthiran, Somshubra Majumdar, Aleksander Ficek, Siddhartha Jain, Jocelyn Huang, Vahid Noroozi, and Boris Ginsburg. Opencodereasoning: Advancing data distillation for competitive coding.arXiv preprint arXiv:2504.01943, 2025. 5.2

    work page arXiv 2025

  37. [37]

    Unlocking efficient long-to-short llm reasoning with model merging.arXiv preprint arXiv:2503.20641,

    Han Wu, Yuxuan Yao, Shuqi Liu, Zehua Liu, Xiaojin Fu, Xiongwei Han, Xing Li, Hui-Ling Zhen, Tao Zhong, and Mingxuan Yuan. Unlocking efficient long-to-short llm reasoning with model merging. arXiv preprint arXiv:2503.20641, 2025. 5.2

    work page arXiv 2025

  38. [38]

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. InInternational conference on machine learning, pages 23965–23998. P...

    work page 2022

  39. [39]

    Editing Models with Task Arithmetic

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Han- naneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic.arXiv preprint arXiv:2212.04089,

    work page internal anchor Pith review arXiv

  40. [40]

    5.2 Scalable Token-Level Hallucination Detection in LLMs14

    work page

  41. [41]

    Ties-merging: Resolving interference when merging models.Advances in Neural Information Processing Systems, 36:7093–7115, 2023

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models.Advances in Neural Information Processing Systems, 36:7093–7115, 2023. 5.2

    work page 2023

  42. [42]

    Language models are super mario: Absorbing abilities from homologous models as a free lunch

    Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. InForty-first International Conference on Machine Learning, 2024. 5.2

    work page 2024

  43. [43]

    Mmboundary: Advancing mllm knowledge boundary awareness through reasoning step confidence calibration

    Zhitao He, Sandeep Polisetty, Zhiyuan Fan, Yuchen Huang, Shujin Wu, and Yi R Fung. Mmboundary: Advancing mllm knowledge boundary awareness through reasoning step confidence calibration. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16427–16444, 2025. 6

    work page 2025

  44. [44]

    Mitigating hallucinations in large vision-language models with instruction contrastive decoding.arXiv preprint arXiv:2403.18715, 2024

    Xintong Wang, Jingheng Pan, Liang Ding, and Chris Biemann. Mitigating hallucinations in large vision-language models with instruction contrastive decoding.arXiv preprint arXiv:2403.18715, 2024. 6

    work page arXiv 2024

  45. [45]

    Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474,

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474,

    work page

  46. [46]

    R-tuning: Instructing large language models to say ‘i don’t know’

    Hanning Zhang, Shizhe Diao, Yong Lin, Yi Fung, Qing Lian, Xingyao Wang, Yangyi Chen, Heng Ji, and Tong Zhang. R-tuning: Instructing large language models to say ‘i don’t know’. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7106...

    work page 2024

  47. [47]

    The law of knowledge overshadowing: Towards understanding, predicting, and preventing llm hallucination.arXiv preprint arXiv:2502.16143, 2025

    Yuji Zhang, Sha Li, Cheng Qian, Jiateng Liu, Pengfei Yu, Chi Han, Yi R Fung, Kathleen McKeown, Chengxiang Zhai, Manling Li, et al. The law of knowledge overshadowing: Towards understanding, predicting, and preventing llm hallucination.arXiv preprint arXiv:2502.16143, 2025. 6

    work page arXiv 2025

  48. [48]

    Empowering reliable visual-centric instruction following in mllms.arXiv preprint arXiv:2601.03198, 2026

    Weilei He, Feng Ju, Zhiyuan Fan, Rui Min, Minhao Cheng, and Yi R Fung. Empowering reliable visual-centric instruction following in mllms.arXiv preprint arXiv:2601.03198, 2026. 6

    work page arXiv 2026

  49. [49]

    Selfcheck: Using llms to zero-shot check their own step-by-step reasoning.arXiv preprint arXiv:2308.00436, 2023

    Ning Miao, Yee Whye Teh, and Tom Rainforth. Selfcheck: Using llms to zero-shot check their own step-by-step reasoning.arXiv preprint arXiv:2308.00436, 2023. 6

    work page arXiv 2023

  50. [50]

    Llm-check: Investigating detection of hallucinations in large language models

    Gaurang Sriramanan, Siddhant Bharti, Vinu Sankar Sadasivan, Shoumik Saha, Priyatham Kattakinda, and Soheil Feizi. Llm-check: Investigating detection of hallucinations in large language models. Advances in Neural Information Processing Systems, 37:34188–34216, 2024. 6

    work page 2024

  51. [51]

    Starling-7b: Improving helpfulness and harmlessness with rlaif

    Banghua Zhu, Evan Frick, Tianhao Wu, Hanlin Zhu, Karthik Ganesan, Wei-Lin Chiang, Jian Zhang, and Jiantao Jiao. Starling-7b: Improving helpfulness and harmlessness with rlaif. InFirst Conference on Language Modeling, 2024. 6

    work page 2024

  52. [52]

    Skywork-Reward: Bag of tricks for reward modeling in llms

    Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Jujie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. Skywork-reward: Bag of tricks for reward modeling in llms.arXiv preprint arXiv:2410.18451, 2024. 6

    work page arXiv 2024

  53. [53]

    A comprehensive survey of reward models: Taxonomy, applications, challenges, and future.arXiv preprint arXiv:2504.12328, 2025

    Jialun Zhong, Wei Shen, Yanzeng Li, Songyang Gao, Hua Lu, Yicheng Chen, Yang Zhang, Wei Zhou, Jinjie Gu, and Lei Zou. A comprehensive survey of reward models: Taxonomy, applications, challenges, and future.arXiv preprint arXiv:2504.12328, 2025. 6 Scalable Token-Level Hallucination Detection in LLMs15

    work page arXiv 2025

  54. [54]

    Rm-bench: Benchmarking reward models of language models with subtlety and style.arXiv preprint arXiv:2410.16184, 2024

    Yantao Liu, Zijun Yao, Rui Min, Yixin Cao, Lei Hou, and Juanzi Li. Rm-bench: Benchmarking reward models of language models with subtlety and style.arXiv preprint arXiv:2410.16184, 2024. 6

    work page arXiv 2024

  55. [55]

    (2025 b ), Pairjudge RM : Perform best-of- N sampling with knockout tournament, arXiv preprint arXiv:2501.13007

    Yantao Liu, Zijun Yao, Rui Min, Yixin Cao, Lei Hou, and Juanzi Li. Pairjudge rm: Perform best-of-n sampling with knockout tournament.arXiv preprint arXiv:2501.13007, 2025. 6

    work page arXiv 2025

  56. [56]

    Prmbench: A fine-grained and challenging benchmark for process-level reward models.arXiv preprint arXiv:2501.03124, 2025

    Mingyang Song, Zhaochen Su, Xiaoye Qu, Jiawei Zhou, and Yu Cheng. Prmbench: A fine-grained and challenging benchmark for process-level reward models.arXiv preprint arXiv:2501.03124, 2025. 6

    work page arXiv 2025

  57. [57]

    Processbench: Identifying process errors in mathematical reasoning

    Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. Processbench: Identifying process errors in mathematical reasoning. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1009–1024, 2025. 6

    work page 2025

  58. [58]

    A survey of monte carlo tree search methods.IEEE Transactions on Computational Intelligence and AI in games, 4(1):1–43, 2012

    Cameron B Browne, Edward Powley, Daniel Whitehouse, Simon M Lucas, Peter I Cowling, Philipp Rohlfshagen, Stephen Tavener, Diego Perez, Spyridon Samothrakis, and Simon Colton. A survey of monte carlo tree search methods.IEEE Transactions on Computational Intelligence and AI in games, 4(1):1–43, 2012. 6

    work page 2012

  59. [59]

    Adam: A Method for Stochastic Optimization

    Diederik P Kingma. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    work page internal anchor Pith review Pith/arXiv arXiv

  60. [60]

    Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvement

    An Yang, Beichen Zhang, Binyuan Chen, et al. Qwen2.5-math technical report: Toward mathematical expert model via self-improvement.arXiv preprint arXiv:2409.12122, 2024. F Scalable Token-Level Hallucination Detection in LLMs16 A Broader Impacts This paper introduces TOKENHD, a framework designed for token-level hallucination detection in LLM- generated res...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  61. [61]

    **Original Text **: A long document or passage

    work page

  62. [62]

    NO_MATCH_FOUND

    **Extracted Text **: A segment supposedly extracted from the original text, but may contain slight variations, errors, or modifications **Your Task **: For each extracted text segment, locate the corresponding section in the original text that best matches it, and output the exact original text segment with appropriate tags. **Instructions**: - The extrac...

    work page