Pith. sign in

REVIEW 4 major objections 6 minor 6 cited by

Large language models carry math difficulty as a high-dimensional linear feature that a probe can read from the final token's embedding.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-04 11:13 UTC pith:QCBKOH75

load-bearing objection The linear difficulty probe is plausible and worth taking seriously; the head-localization story is attractive but rests on a linearized final layer and a reused probe, so it needs stricter validation. the 4 major comments →

arxiv 2510.05969 v3 pith:QCBKOH75 submitted 2025-10-07 cs.CL cs.AI

Probing the Difficulty Perception Mechanism of Large Language Models

Sunbowen Lee , Qingyu Yin , Chak Tou Leong , Jialiang Zhang , Yicheng Gong , Shiwen Ni , Min Yang , Xiaoyu Shen This is my paper
classification cs.CL cs.AI
keywords difficulty perceptionlinear probingattention headsmechanistic interpretabilitymath reasoningDeepMathentropylarge language models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that a trained LLM does not merely answer math problems—it internally registers how hard each problem is, in a structured and manipulable way. Using a linear probe on the last-token embedding, the authors find that human-annotated difficulty levels can be recovered as a single high-dimensional direction. They then localize this perception to specific attention heads in the final transformer layer: heads 7, 8, 16, and 23 respond to hard problems, while heads 10–13 respond to easy ones in Qwen2.5-7B-Instruct, and scaling those heads up or down shifts the model's perceived difficulty. This matters because difficulty perception is a prerequisite for adaptive reasoning—models that know a problem is hard could budget more tokens and avoid collapsing on hard inputs—and for replacing costly human difficulty labeling with model-based annotation.

Core claim

The central claim is that difficulty perception in LLMs is real, linear, localizable, and causal. On DeepMath problems, a linear regressor trained on final-token embeddings predicts difficulty levels that match human labels and generalize out-of-distribution (GSM8K scores low). The authors then attribute each attention head's contribution by projecting the head-wise masked output through the output projection onto the probe direction, obtaining a differentiation score between hard and easy cohorts. For Qwen2.5-7B-Instruct, this isolates four difficulty-sensitive heads (7, 8, 16, 23) and four easy-sensitive heads (10, 11, 12, 13) in the final layer. Scaling these heads (easy heads ×0.1, hard

What carries the argument

The key machinery is a high-dimensional linear difficulty direction. A lightweight linear probe y = w^T h + b, trained by least squares on human-labeled DeepMath difficulties, establishes that the last-token embedding contains a difficulty axis. To locate heads, the paper models the final-layer output as a linear projection of concatenated attention heads, Z = Reshape(H) W_o^T, then zeroes all heads except one, projects the masked embedding onto the probe direction, and computes a differentiation score Δ = mean score(hard batch) − mean score(easy batch). This score ranks heads by sensitivity and yields the identified easy/hard head sets. The same machinery then becomes an intervention lever:

Load-bearing premise

The head-localization result rests on treating the final-layer representation as a pure linear projection of attention-head outputs, ignoring residual connections and normalization; if those nonlinearities substantially mix or redistribute head contributions, the identified difficulty heads might be artifacts of that linear model rather than the actual causal loci.

What would settle it

Run the same difficulty probe and head attribution using a true forward-pass ablation—zero or scale one attention head at a time and measure the model's own output difficulty—rather than the closed-form linear projection. If the full-model intervention on a non-identified head changes perceived difficulty as much as the identified heads do, or if zeroing identified heads does not change difficulty, the linear head attribution is wrong. A second check: train the probe on shuffled difficulty labels; if it still fits well or the head patterns persist, the linear signal may be an artifact of quest

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the linear probe generalizes, LLMs can serve as automatic difficulty annotators for benchmark construction and curriculum learning, reducing human labeling cost.
  • Because manipulating the identified heads shifts perceived difficulty, difficulty perception can in principle be steered, enabling adaptive token budgets for reasoning.
  • The finding that perceived difficulty resists entropy-based estimation implies entropy or perplexity is not a reliable token-level difficulty proxy; probing is a distinct signal.
  • Ablation results show the located heads are causally involved, so model behavior can be altered by head-wise scaling, not just observed.
  • Not all models show clear head patterns (Llama3.1-8B-Instruct is weak), so the phenomenon is model-dependent and affected by pre/post-training.

Where Pith is reading between the lines

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

  • The same head-wise attribution could be run on mid-reasoning tokens to build a real-time difficulty meter that flags where a chain-of-thought starts to go wrong, not just the prompt's initial difficulty.
  • The reversal of easy/hard head patterns after DeepSeek-R1 distillation suggests that distillation can reassign, but not erase, the difficulty circuit; comparing head maps across checkpoints may expose when a model acquires or loses difficulty sensitivity.
  • If difficulty is a linear direction, adding or subtracting that direction from embeddings (rather than scaling heads) should reproduce the difficulty shift—an inexpensive testable extension.
  • The token-level divergence for numeric tokens implies a probe-based difficulty signal could be used to decide where to spend verification effort, complementing confidence and entropy in RL training.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper claims that LLMs encode mathematical problem difficulty as a high-dimensional linear feature in the final-token embedding, that this direction can be learned by a linear probe, that specific attention heads in the final transformer layer are functionally specialized for easy vs. hard problems, and that scaling those heads shifts the model's perceived difficulty. It also reports token-usage effects and a token-level divergence between difficulty perception and entropy. The headline results are that difficulty is linearly readable and that the relevant heads are localizable and manipulable.

Significance. If the linear-encoding and head-localization claims held, this would be a useful mechanistic account of difficulty perception and a practical route to automated difficulty annotation. Strengths include the use of a post-training human-annotated dataset (DeepMath), a sanity check of the label-quality/accuracy curve, an OOD test with GSM8K, a behavioral token-usage experiment independent of the probe, and released code. However, the absence of quantitative probe metrics and the simplified linear model behind head attribution leave the central causal claims under-supported as currently presented.

major comments (4)
  1. [§4.1, Figure 6] The central claim that difficulty is linearly encoded is supported only by a scatter plot and the qualitative statement that the probe 'accurately categorize[s]' the test set. No R², RMSE, Pearson/Spearman correlation, per-level calibration, or confidence intervals are reported, and the OOD claim for GSM8K rests on a visual comparison. Please add quantitative probe metrics on the DeepMath test split and OOD data, with baselines (e.g., token count, response length, length-based heuristics).
  2. [§3.2, Eqs. (3)–(7)] The head-attribution framework assumes Z = Reshape(H) W_o^T, i.e., the final-token representation is a pure linear projection of attention-head outputs. Qwen2.5's transformer layer includes a residual stream, RMSNorm, and MLP; the probe was trained on the actual final hidden state. Consequently, zeroing all heads except i and applying W_o^T does not correspond to removing heads in the live model, and Eq. (7) can rank a head high merely because its output projection aligns with the probe direction, even if the head has no causal role. Please derive the attribution on the true computational graph or validate the selected heads with actual forward-pass ablations that do not assume Eq. (3).
  3. [§5, Table 1] The ablation evidence is not independent of the selection procedure: the same fitted linear probe is used both to select S_easy/Shard and to measure the effect of scaling them, and scaling a head's output changes the projected representation in a mechanically predictable way. To support the causal localization claim, include control conditions (random head subsets, reverse scaling, scaling of low-ranked heads), report behavioral outcomes not measured by the probe (accuracy, response length, generation content), and give confidence intervals/error bars. Table 2 is a valuable independent behavioral signal, but it covers only the Increase condition and lacks a random-head control and uncertainty estimates.
  4. [§6, Figures 10–12] The token-level difficulty-vs-entropy analysis is presented as a general finding ('significant difference') but is based on one or a few qualitative examples. No quantitative measure of agreement/disagreement between difficulty scores and entropy, no aggregation over multiple problems, and no error bars are provided. Please either add a systematic quantitative analysis or explicitly frame these figures as illustrative case studies.
minor comments (6)
  1. [§3.2/§5] Specify the reshape ordering and head indexing in Eq. (3) and Eq. (8); 'headers' should be 'heads' in §4.2.
  2. [Appendix B.1] 'an good effect' should be 'a good effect'.
  3. [Appendix C] 'Referring to Figure 16' should likely be 'Figure 17'.
  4. [References] Several references have broken author formatting, e.g., 'Qwen, :' and 'YunXing, XingYu'; please fix.
  5. [Table 2] Report sample sizes and standard deviations; the header 'Avg. token used' should be 'Average tokens used'.
  6. [§4.1] State the train/validation/test split, number of examples per difficulty level, and probe hyperparameters (epochs, optimizer, seeds) to support reproducibility.

Circularity Check

1 steps flagged

Head localization and the ablation 'causal evidence' are measured with the same fitted probe that was used to select the heads, making part of the causal claim circular.

specific steps
  1. fitted input called prediction [Section 3.2 / Eq. (5)–(7); Section 5 / Table 1]
    "As shown in Table 1, the difficulty increasing setting leads to a consistent increase in the model’s estimated difficulty scores across all inputs, effectively biasing the model toward perceiving problems as more challenging. Conversely, difficulty decrease yields lower difficulty estimates, aligning with the hypothesis that these head groups encode complementary signals for problem complexity. These results provide causal evidence that specific attention heads are functionally specialized for perceiving inputs of different difficulty levels."

    The 'estimated difficulty scores' in Table 1 are computed with the same fitted linear probe vdiff that was used in Eq. (5)–(7) to define head-wise difficulty attribution and to select the easy- and hard-sensitive heads. Under the paper's own linear model, Eq. (3), scaling up heads whose projection onto vdiff is large and scaling down heads whose projection is small forces the probe score to shift in the advertised direction. Thus Table 1 is largely a restatement of the head-selection criterion, not independent confirmation of the localization. The only probe-free behavioral evidence is the token-use change in Table 2, which covers only the increasing-difficulty intervention and lacks a random-head control.

full rationale

The paper's main non-circular result is the linear probe itself: a linear regressor trained on DeepMath human difficulty labels predicts held-out DeepMath difficulty and produces the expected low scores on GSM8K. That is a genuine, self-contained empirical finding. The circularity appears in the attention-head localization and its validation. Head sensitivity is defined (Eq. 5–7) as the normalized projection of a masked, linearized final representation onto the already-fitted probe vector. The same probe vector is then used in the ablation to measure the resulting 'predicted difficulty' (Table 1). Since the heads were selected precisely for having large positive or negative projections, the ablation's probe-score shifts are mathematically expected, not evidence that the model's behavior or human-perceived difficulty is causally controlled by those heads. The token-usage result (Table 2) is a genuinely independent behavioral outcome and does not reduce to the probe parameters, which prevents the entire paper from being circular. However, the load-bearing claim that the specific heads {7,8,16,23} and {10,11,12,13} are causally specialized rests substantially on the same-probe evaluation. Additionally, Eq. (3) omits the residual stream and normalization that are present in actual Qwen2.5 blocks; while this is a correctness threat rather than a circularity, it reinforces that the head attribution is not independently secured. No load-bearing self-citation was found: the cited prior work by the same authors is background, not a uniqueness theorem or an ansatz on which the derivation depends. Overall, one central 'prediction' reduces by construction to the fitted probe, yielding a partial circularity score of 6.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or ontological entities. The 'difficulty perception direction' is the fitted probe weight vector, not an independently postulated mechanism. The main extra ingredients are the fitted probe, the hand-selected head sets, and the hand-chosen scaling factors.

free parameters (4)
  • Linear probe weights w and bias b = not reported (released in code repo, no commit hash)
    Fitted to DeepMath human difficulty labels with MSE (Eq. 2); every downstream difficulty score is computed by projecting hidden states onto this fitted direction.
  • Attention head index sets Seasy / Shard = Seasy={10,11,12,13}; Shard={7,8,16,23} for Qwen2.5-7B
    Selected from Figure 7 heatmap as the extreme heads; used as ablation targets in Eq. 8, so the intervention targets are chosen from the same data they are tested on.
  • Head scaling factors alpha_reduce, alpha_increase = 0.1 and 2.0
    Chosen by hand (Eq. 8); no sweep or sensitivity analysis is reported, so the quantitative shift in Table 1 may depend on these values.
  • Difficulty contrast levels for delta = levels 9 and 3
    The head differentiation score uses only the two extreme difficulty cohorts; other pairs might yield different head rankings.
axioms (4)
  • domain assumption Linear probing of last-token embeddings is a valid method for reading out internal representations.
    Section 3.1 adopts probing as a standard interpretability tool, but no control tasks (e.g., random-label probes) are run, so the probe's success is not compared against what a high-dimensional linear model can fit trivially.
  • domain assumption DeepMath human difficulty labels are reliable ground truth.
    Section 4.1 validates labels only by the aggregate trend that model accuracy falls as difficulty rises (Figure 4); this does not establish per-item reliability or that human difficulty equals model difficulty.
  • ad hoc to paper The final-layer representation is a linear projection of attention-head outputs (Eq. 3), ignoring residual connections and LayerNorm.
    Eq. 3-5 assume Z=Reshape(H)W_o^T with no residual stream or normalization; actual Transformer blocks include both, and the head attribution depends on this simplification.
  • domain assumption Probe trained on prompt-final tokens transfers to every token during generation.
    Section 6 applies the probe to token-level states inside the reasoning trace, assuming the same linear direction measures difficulty there.

pith-pipeline@v1.3.0-alltime-deepseek · 10807 in / 16182 out tokens · 108670 ms · 2026-08-04T11:13:57.296088+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) are increasingly deployed on complex reasoning tasks, yet little is known about their ability to internally evaluate problem difficulty, which is an essential capability for adaptive reasoning and efficient resource allocation. In this work, we investigate whether LLMs implicitly encode problem difficulty in their internal representations. Using a linear probe on the final-token representations of LLMs, we demonstrate that the difficulty level of math problems can be linearly modeled. We further locate the specific attention heads of the final Transformer layer: these attention heads have opposite activation patterns for simple and difficult problems, thus achieving perception of difficulty. Our ablation experiments prove the accuracy of the location. Crucially, our experiments provide practical support for using LLMs as automatic difficulty annotators, potentially substantially reducing reliance on costly human labeling in benchmark construction and curriculum learning. We also uncover that there is a significant difference in entropy and difficulty perception at the token level. Our study reveals that difficulty perception in LLMs is not only present but also structurally organized, offering new theoretical insights and practical directions for future research. Our code is available at https://github.com/Aegis1863/Difficulty-Perception-of-LLMs.

Figures

Figures reproduced from arXiv: 2510.05969 by Chak Tou Leong, Jialiang Zhang, Min Yang, Qingyu Yin, Shiwen Ni, Sunbowen Lee, Xiaoyu Shen, Yicheng Gong.

Figure 1
Figure 1. Figure 1: The LLM’s perception of problem difficulty [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Probe training and attention heads pattern recognition. On the left we demonstrate how to train a difficulty [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qwen2.5-7B-Instruct’s low-dimensional rep [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Probe training and validation loss. Training [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Probe test result. Among them, GSM8K is out [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Different size model attention head pattern recognition results. We show the directions under significant [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The attention head patterns of Qwen2.5-7B [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Utilize Probe for difficulty prediction during [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: In the inference of a difficult math question [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Token-level entropy during inference. To [PITH_FULL_IMAGE:figures/full_fig_p008_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: DeepSeek-R1-Distill-Qwen-7B’s low￾dimensional representation for DeepMath problems and difficulty is meticulously annotated by humans [PITH_FULL_IMAGE:figures/full_fig_p011_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Tokenskip’s low-dimensional representation [PITH_FULL_IMAGE:figures/full_fig_p011_14.png] view at source ↗
Figure 17
Figure 17. Figure 17: Reasoning Collapse Token Reduction. As the [PITH_FULL_IMAGE:figures/full_fig_p011_17.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. The Shape of Addition: Geometric Structures of Arithmetic in Large Language Models

    cs.LG 2026-05 unverdicted novelty 7.0

    LLM residual streams during addition form an Iso-Raw-Sum Trajectory anchored by digit semantics and modulated by continuous carry signals, with errors arising as geometric slippages across quantization thresholds in a...

  2. Cognitive offloading and the speedup illusion in human-AI interaction

    cs.CY 2026-05 unverdicted novelty 6.0

    Preregistered behavioral study identifies a speedup illusion where users overestimate time savings from AI assistance on cognitive tasks despite no actual difference in completion times.

  3. Reasoning Models Don't Just Think Longer, They Move Differently

    cs.CL 2026-05 unverdicted novelty 6.0

    After length correction, reasoning-trained language models exhibit distinct hidden-state trajectory geometries on harder problems compared to instruction-tuned baselines, with the strongest effect in code domains.

  4. DARE: Difficulty-Adaptive Reinforcement Learning with Co-Evolved Difficulty Estimation

    cs.LG 2026-05 unverdicted novelty 6.0

    DARE co-evolves difficulty estimation and policy in RL for LLMs to improve training efficiency, final performance, and inference speed by using tailored strategies for different difficulty levels.

  5. RouteLMT: Learned Sample Routing for Hybrid LLM Translation Deployment

    cs.CL 2026-04 unverdicted novelty 6.0

    RouteLMT learns to route MT requests to large or small LLMs by predicting marginal quality gain from small-model token representations, yielding a better quality-budget Pareto frontier than baselines.

  6. Reasoning Models Don't Just Think Longer, They Move Differently

    cs.CL 2026-05 unverdicted novelty 5.0

    After length-correcting hidden-state trajectories during chain-of-thought, reasoning models show systematically different geometry on harder problems than baselines, strongest in competitive programming.

Reference graph

Works this paper leans on

38 extracted references · 1 canonical work pages · cited by 5 Pith papers

  1. [1]

    Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. 2024. Refusal in language models is mediated by a single direction. Advances in Neural Information Processing Systems, 37:136037--136083

    2024

  2. [2]

    Seyedarmin Azizi, Erfan Baghaei Potraghloo, and Massoud Pedram. 2025. https://arxiv.org/abs/2507.04742 Activation steering for chain-of-thought compression . Preprint, arXiv:2507.04742

    Pith/arXiv arXiv 2025

  3. [3]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. https://arxiv.org/abs/2502.13923 Qwen2.5-vl technical report . Preprint, arXiv:2502.13923

    Pith/arXiv arXiv 2025

  4. [4]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

    Pith/arXiv arXiv 2021

  5. [5]

    Yingqian Cui, Pengfei He, Jingying Zeng, Hui Liu, Xianfeng Tang, Zhenwei Dai, Yan Han, Chen Luo, Jing Huang, Zhen Li, Suhang Wang, Yue Xing, Jiliang Tang, and Qi He. 2025. https://doi.org/10.18653/v1/2025.findings-acl.956 Stepwise perplexity-guided refinement for efficient chain-of-thought reasoning in large language models . In Findings of the Associatio...

    doi:10.18653/v1/2025.findings-acl.956 2025

  6. [6]

    Gongfan Fang, Xinyin Ma, and Xinchao Wang. 2025. https://arxiv.org/abs/2505.13379 Thinkless: Llm learns when to think . Preprint, arXiv:2505.13379

    Pith/arXiv arXiv 2025

  7. [7]

    Mills, Baochun Li, and Di Niu

    Amirhosein Ghasemabadi, Keith G. Mills, Baochun Li, and Di Niu. 2025. https://arxiv.org/abs/2505.20325 Guided by gut: Efficient test-time scaling with reinforced intrinsic confidence . Preprint, arXiv:2505.20325

    Pith/arXiv arXiv 2025

  8. [8]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, and 175 others. 2025. https://doi.org/10.1038/s41586-025-09422-z Deepseek-r1 incentivizes reasoning in llms through reinforcement lear...

    doi:10.1038/s41586-025-09422-z 2025

  9. [9]

    Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. 2025. https://arxiv.org/abs/2412.18547 Token-budget-aware llm reasoning . Preprint, arXiv:2412.18547

    Pith/arXiv arXiv 2025

  10. [10]

    Zhiwei He, Tian Liang, Jiahao Xu, Qiuzhi Liu, Xingyu Chen, Yue Wang, Linfeng Song, Dian Yu, Zhenwen Liang, Wenxuan Wang, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. 2025. https://arxiv.org/abs/2504.11456 Deepmath-103k: A large-scale, challenging, decontaminated, and verifiable mathematical dataset for advancing reasoning . Preprint, ar...

    Pith/arXiv arXiv 2025

  11. [11]

    John Hewitt and Percy Liang. 2019. https://doi.org/10.18653/v1/D19-1275 Designing and interpreting probes with control tasks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2733--2743, Hong Kong, China. Association fo...

    doi:10.18653/v1/d19-1275 2019

  12. [12]

    Leyang Hu and Boran Wang. 2024. https://arxiv.org/abs/2411.09125 Droj: A prompt-driven attack against large language models . Preprint, arXiv:2411.09125

    Pith/arXiv arXiv 2024

  13. [13]

    Chengyu Huang, Zhengxin Zhang, and Claire Cardie. 2025. https://arxiv.org/abs/2505.11225 Hapo: Training language models to reason concisely via history-aware policy optimization . Preprint, arXiv:2505.11225

    arXiv 2025

  14. [14]

    Sunbowen Lee, Shiwen Ni, Chi Wei, Shuaimin Li, Liyang Fan, Ahmadreza Argha, Hamid Alinejad-Rokny, Ruifeng Xu, Yicheng Gong, and Min Yang. 2025. https://arxiv.org/abs/2501.16727 xjailbreak: Representation space guided reinforcement learning for interpretable llm jailbreaking . Preprint, arXiv:2501.16727

    Pith/arXiv arXiv 2025

  15. [15]

    Haotian Luo, Li Shen, Haiying He, Yibo Wang, Shiwei Liu, Wei Li, Naiqiang Tan, Xiaochun Cao, and Dacheng Tao. 2025. https://arxiv.org/abs/2501.12570 O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning . Preprint, arXiv:2501.12570

    Pith/arXiv arXiv 2025

  16. [16]

    OpenAI. 2025. https://arxiv.org/abs/2508.10925 gpt-oss-120b & gpt-oss-20b model card . Preprint, arXiv:2508.10925

    Pith/arXiv arXiv 2025

  17. [17]

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, and 25 others. 2025. https://arxiv.org/abs/2412.15115 Qwen2.5 technical report . Preprint, arXiv:2412.15115

    Pith/arXiv arXiv 2025

  18. [18]

    Yi Shen, Jian Zhang, Jieyun Huang, Shuming Shi, Wenjing Zhang, Jiangze Yan, Ning Wang, Kai Wang, Zhaoxiang Liu, and Shiguo Lian. 2025. https://arxiv.org/abs/2503.04472 Dast: Difficulty-adaptive slow-thinking for large reasoning models . Preprint, arXiv:2503.04472

    arXiv 2025

  19. [19]

    Leheng Sheng, An Zhang, Zijian Wu, Weixiang Zhao, Changshuo Shen, Yi Zhang, Xiang Wang, and Tat - Seng Chua. 2025. On reasoning strength planning in large reasoning models. CoRR, abs/2506.08390

    arXiv 2025

  20. [20]

    Taiwei Shi, Yiyang Wu, Linxin Song, Tianyi Zhou, and Jieyu Zhao. 2025. https://arxiv.org/abs/2504.05520 Efficient reinforcement finetuning via adaptive curriculum learning . Preprint, arXiv:2504.05520

    Pith/arXiv arXiv 2025

  21. [21]

    Parshin Shojaee, Iman Mirzadeh, Keivan Alizadeh, Maxwell Horton, Samy Bengio, and Mehrdad Farajtabar. 2025. https://arxiv.org/abs/2506.06941 The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity . Preprint, arXiv:2506.06941

    Pith/arXiv arXiv 2025

  22. [22]

    Zineddine Tighidet, Jiali Mei, Benjamin Piwowarski, and Patrick Gallinari. 2024. https://doi.org/10.18653/v1/2024.blackboxnlp-1.35 Probing language models on their knowledge source . In Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pages 604--614, Miami, Florida, US. Association for Computational Linguistics

    work page doi:10.18653/v1/2024.blackboxnlp-1.35 2024

  23. [23]

    Shenzhi Wang, Le Yu, Chang Gao, Chujie Zheng, Shixuan Liu, Rui Lu, Kai Dang, Xionghui Chen, Jianxin Yang, Zhenru Zhang, Yuqiong Liu, An Yang, Andrew Zhao, Yang Yue, Shiji Song, Bowen Yu, Gao Huang, and Junyang Lin. 2025. https://arxiv.org/abs/2506.01939 Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforcement learning for llm reas...

    Pith/arXiv arXiv 2025

  24. [24]

    Siye Wu, Jian Xie, Yikai Zhang, Aili Chen, Kai Zhang, Yu Su, and Yanghua Xiao. 2025. https://arxiv.org/abs/2505.20258 Arm: Adaptive reasoning model . Preprint, arXiv:2505.20258

    arXiv 2025

  25. [25]

    Heming Xia, Chak Tou Leong, Wenjie Wang, Yongqi Li, and Wenjie Li. 2025. https://arxiv.org/abs/2502.12067 Tokenskip: Controllable chain-of-thought compression in llms . Preprint, arXiv:2502.12067

    arXiv 2025

  26. [26]

    Zhihao Xu, Ruixuan Huang, Changyu Chen, and Xiting Wang. 2024. https://arxiv.org/abs/2404.12038 Uncovering safety risks of large language models through concept activation vector . Preprint, arXiv:2404.12038

    Pith/arXiv arXiv 2024

  27. [27]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025 a . https://arxiv.org/abs/2505.09388 Qwen3 technical report . Preprint, arXiv:2505.09388

    Pith/arXiv arXiv 2025

  28. [28]

    Junjie Yang, Ke Lin, and Xing Yu. 2025 b . https://arxiv.org/abs/2504.03234 Think when you need: Self-adaptive chain-of-thought learning . Preprint, arXiv:2504.03234

    Pith/arXiv arXiv 2025

  29. [29]

    Wenkai Yang, Shuming Ma, Yankai Lin, and Furu Wei. 2025 c . https://arxiv.org/abs/2502.18080 Towards thinking-optimal scaling of test-time compute for llm reasoning . Preprint, arXiv:2502.18080

    arXiv 2025

  30. [30]

    Qingyu Yin, Chak Tou Leong, Linyi Yang, Wenxuan Huang, Wenjie Li, Xiting Wang, Jaehong Yoon, YunXing, XingYu, and Jinjin Gu. 2025. https://arxiv.org/abs/2510.06036 Refusal falls off a cliff: How safety alignment fails in reasoning? Preprint, arXiv:2510.06036

    arXiv 2025

  31. [31]

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, and 16 others. 2025. https://arxiv.org/abs/2503.14476 Dapo: An open-source llm reinforcement learning system at scale . Preprin...

    Pith/arXiv arXiv 2025

  32. [32]

    Mohammad Zbeeb, Hasan Abed Al Kader Hammoud, and Bernard Ghanem. 2025. https://arxiv.org/abs/2509.01363 Reasoning vectors: Transferring chain-of-thought capabilities via task arithmetic . Preprint, arXiv:2509.01363

    Pith/arXiv arXiv 2025

  33. [33]

    Chujie Zheng, Shixuan Liu, Mingze Li, Xiong-Hui Chen, Bowen Yu, Chang Gao, Kai Dang, Yuqiong Liu, Rui Men, An Yang, Jingren Zhou, and Junyang Lin. 2025. https://arxiv.org/abs/2507.18071 Group sequence policy optimization . Preprint, arXiv:2507.18071

    Pith/arXiv arXiv 2025

  34. [34]

    Chujie Zheng, Fan Yin, Hao Zhou, Fandong Meng, Jie Zhou, Kai-Wei Chang, Minlie Huang, and Nanyun Peng. 2024. https://arxiv.org/abs/2401.18018 On prompt-driven safeguarding for large language models . Preprint, arXiv:2401.18018

    Pith/arXiv arXiv 2024

  35. [35]

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Xuehui Wang, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, and 32 others. 2025 a . https://arxiv.org/abs/2504.10479 Internvl3: Exploring advanced training and test-time recipes for open...

    Pith/arXiv arXiv 2025

  36. [36]

    Yubo Zhu, Dongrui Liu, Zecheng Lin, Wei Tong, Sheng Zhong, and Jing Shao. 2025 b . https://arxiv.org/abs/2509.12886 The llm already knows: Estimating llm-perceived question difficulty via hidden representations . Preprint, arXiv:2509.12886

    arXiv 2025

  37. [37]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  38. [38]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.