ThinkLess: A Training-Free Inference-Efficient Method for Reducing Reasoning Redundancy

Gengyang Li; Yifeng Gao; Yuming Li; Yunfang Wu

REVIEW 5 major objections 6 minor 3 cited by

ThinkLess: A Training-Free Inference-Efficient Method for Reducing Reasoning Redundancy

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that visible chain-of-thought reasoning is largely redundant and can be skipped by inserting the reasoning terminator early, with accuracy preserved.

desk verdict A practical prompt trick for cutting CoT cost, but the mechanism is oversold and the evaluation needs more rigor before the claims are load-bearing. read the letter →

arxiv 2505.15684 v2 pith:OVWOYAU2 submitted 2025-05-21 cs.CL

Gengyang Li , Yifeng Gao , Yuming Li , Yunfang Wu This is my paper · ORCID

classification cs.CL

keywords chain-of-thoughtpromptinginferenceefficiencyearlyterminationreasoningcompressionattentionanalysisKVcachetraining-freemethodDeepSeek-R1distilledmodels

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Chain-of-thought (CoT) prompting makes an LLM write out intermediate reasoning before answering, and this paper sets out to show that most of that writing is redundant at inference time. Its central claim is that answer tokens chiefly rely on the reasoning terminator token, , rather than on the individual reasoning tokens, because causal masking pushes information toward the end of the reasoning span. ThinkLess therefore inserts immediately after , skips explicit CoT generation entirely, and appends a short task-specific instruction to keep the output well formatted. Across three DeepSeek-R1-distilled instruction-tuned models, the paper reports Top@1 accuracy within about one to two points of full-length CoT while using roughly 60–70% fewer tokens and about half the decoding time. If this is right, a large part of the inference cost of CoT-style models can be removed without fine-tuning, extra data, or any change to the model weights.

What carries the argument

The load-bearing object is the </think> terminator token, treated as a semantic anchor: a learned symbolic marker that, under causal masking, accumulates and compresses the reasoning state as it is propagated forward across tokens. The paper supports this with two analyses: attention heatmaps showing answer tokens concentrating on the </think> boundary in deeper layers, and pairwise cosine similarities of last-layer hidden states extracted at inserted </think> positions every 16 tokens, which are about 0.9 between adjacent segments. The second mechanism is a lightweight output-regulation instruction appended after early termination, which steers the model toward scorable answer formats and is credited with recovering most of the accuracy lost by premature truncation.

What would settle it

Measure the last-layer hidden-state cosine similarity between a </think> token generated immediately after <think> with no reasoning tokens and the final </think> state after full-length CoT on the same GSM8K and GPQA samples; then separately evaluate accuracy with the task instruction prompt but without any </think> insertion. If the zero-token state is not substantially similar to the final state, or if the instruction prompt alone reproduces the reported accuracy, the central semantic-anchor claim fails.

Watch

Extended reading notes

Core claim

The paper's discovery claim is that during autoregressive CoT decoding, information gradually migrates toward the end of the reasoning span, so the model's answer generation depends on the </think> token as a compressed representation rather than on each earlier reasoning token. ThinkLess operationalizes this by inserting </think> right after <think>, terminating reasoning before any explicit reasoning tokens are generated, and then using a tiny instruction prompt to restore answer formatting. On Qwen2.5-7B, Qwen2.5-14B, and LLaMA3.1-8B distilled with DeepSeek-R1 traces, the authors report average Top@1 accuracy of 62.91, 73.07, and 60.61, compared with 62.28, 74.82, and 60.28 for full CoT decoding, while reducing token use by 60–70% and inference time by roughly half. The paper interprets this as evidence that reasoning can be compressed into latent activations and that explicit CoT generation is not necessary for maintaining answer quality.

Load-bearing premise

The load-bearing premise is that the model really performs its reasoning internally when the </think> token is inserted right after <think>, even though no reasoning tokens are written out; the paper's similarity measurements only test </think> tokens placed after blocks of 16 reasoning tokens, not the zero-token condition the method actually uses.

Editorial extensions

If this is right

For DeepSeek-R1-distilled models, visible CoT reasoning can be skipped with average Top@1 accuracy remaining within about one to two points of full-length CoT decoding.
Token usage drops by roughly 60–70% and inference time by about half across GSM8K, MMLU, GPQA, and BBH, making the trade-off practical for latency-sensitive deployment.
Under equal token budgets, generating several short candidate answers in parallel (Top@k) can beat the full CoT baseline's Top@1 accuracy while also lowering latency, because parallel decoding avoids sequential long generations.
The appended task-specific instruction is essential: removing it causes accuracy to drop, mostly because of malformed output formats rather than incorrect internal reasoning.
The method is training-free and model-agnostic at the inference level, so it can be applied to already-deployed instruction-tuned models without modifying weights or collecting auxiliary data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper's similarity evidence comes from inserting </think> after 16-token reasoning blocks, whereas the deployed method inserts it after zero reasoning tokens; the semantic-anchor interpretation for the zero-token case is therefore an extrapolation, and an equally simple explanation is that the appended instruction alone triggers the model to answer from familiar task formats.
A testable extension: if the anchor mechanism is real, the method should transfer to other reasoning-tag formats, such as [REASONING]...[/REASONING], provided the terminator token comes from the model's own training distribution.
The parallel Top@k variant suggests the real efficiency gain is not only fewer tokens but also parallelizability, so a dynamic per-question termination policy could push the accuracy-latency curve further than the fixed early insertion reported here.
Task-specific instruction design remains part of the deployment cost; the claimed savings are most concrete when prompt engineering is amortized over many queries per task.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Fairly simple paper with a practical result. The trick is to insert immediately after and append a task-specific instruction; on three R1-distilled models this gives about the same average Top-1 accuracy as full CoT while cutting tokens by 60-70% and time by half. I think the headline empirical result is probably real, and it's a useful plug-and-play trick for serving these models.

The specific recipe is not in the cited baselines (CoD, Sketch-of-Thought), and the paper is honest in its Limitations section that success depends on instruction quality and that dynamic truncation is not handled. The ablation (w/o Instruct) shows the instruction matters, and the failure examples in Appendix A are convincing that a big chunk of the loss is formatting.

The mechanistic story is the weakest part. The attention/similarity analysis inserts after every 16 tokens and only on six GSM8K samples; the deployed method inserts it with zero reasoning tokens, and no measurement covers that condition. The 'semantic anchor' claim reduces to the observation that the model answers correctly without visible reasoning, which any instruction-following model might do. The missing control is the same instruction without the pair; if that matches ThinkLess, the terminator token isn't doing what they say.

Also, Table 1 is single-run, and the small average gaps (1-2 points) are within noise on some datasets. The per-dataset picture is mixed: MMLU on Qwen-14B drops from 81.40 to 76.44, and BBH drops from 83.84 to 78.38. The Top@k comparison in Figure 5 is not apples-to-apples against Top@1. The 'first training-free' claim is contradicted by their own cited CoD and Sketch-of-Thought. The manual-correction recovery claim is unquantified. And the instructions in Appendix B tell the model to 'think step by step' and 'self-check', so 'without explicit CoT generation' is probably not literally true; the model may emit some reasoning after the instruction, though clearly much less.

All of that is fixable. The practical observation is useful enough that I'd send it to a serious referee, with the expectation that the authors either drop or properly test the mechanism and tighten the claims. It should be a conditional accept, not a reject, because the empirical finding has legs.

Referee Report

5 major / 6 minor

Summary. The paper proposes ThinkLess, a training-free inference-time method for reducing Chain-of-Thought (CoT) reasoning overhead. ThinkLess inserts the reasoning terminator token </think> immediately after <think>, thereby skipping explicit reasoning generation, and then applies a task-specific instruction prompt to regulate output format. The paper claims that this achieves accuracy comparable to full CoT decoding (Table 1: e.g., average Top@1 62.91 vs 62.28 for Qwen2.5-7B; 73.07 vs 74.82 for Qwen2.5-14B) while reducing token usage by roughly 60-70% and inference time by about half. The underlying mechanistic claim is that answer tokens attend primarily to the </think> token, which acts as a 'semantic anchor' encoding internally compressed reasoning. Experiments are conducted on GSM8K, MMLU, GPQA, and BBH with Qwen2.5-7B/14B and LLaMA3.1-8B.

Significance. If the efficiency and accuracy results hold, ThinkLess offers a simple, plug-and-play way to reduce CoT inference cost without training or model modification, which is practically valuable. The paper reports concrete efficiency gains and provides attention visualizations that motivate the approach. However, the load-bearing mechanistic claim is not tested in the configuration actually deployed, a key control is missing, and several evaluation choices (single-run results, Top@k vs Top@1 comparison) weaken the empirical support. The paper would be a useful contribution if these issues are addressed; in its current form, the central claims are only partially supported.

major comments (5)

[Sec 3.3, Table 1, Table 5] The central claim that the </think> token acts as a 'semantic anchor' (Sec 3.3, Clarification) is not supported by the current experiments because no control condition removes the <think></think> pair while keeping the same task instruction prompt. Table 1 shows ThinkLess (with instruction) achieves accuracy comparable to Distill, and the ThinkLess w/o Instruct variant is worse, but this only shows the instruction prompt matters. Without a condition using just the instruction prompt and no think tokens, the comparable accuracy could be entirely a direct-answer prompt-formatting effect, making the attention-migration mechanism non-load-bearing. Please add this control and report its accuracy.
[Sec 3.2, Figure 3] The similarity analysis in Sec 3.2 inserts </think> after every 16 generated reasoning tokens, whereas ThinkLess inserts it immediately after <think> with zero intermediate reasoning tokens. Figure 3 shows that the first inserted state is only about 0.71-0.80 similar to the final state, and no measurement is reported for the zero-reasoning condition. Thus the evidence does not establish that the deployed configuration yields a hidden state that approximates the final reasoning state. Please measure the similarity of the </think> hidden state in the actual ThinkLess configuration (zero or very few preceding tokens) to the final-state representation, and report aggregated numbers across samples and models rather than six individual examples.
[Sec 3.3, Appendix A] The claim that the accuracy drop after early termination is a surface-level formatting artifact rests on a manual-correction recovery that is never quantified. Sec 3.3 states that after manually correcting malformed outputs 'a substantial recovery in overall accuracy' is observed, but no before/after numbers are given. Please report the accuracy before and after manual correction, specify the correction protocol, and ideally show that the recovery is consistent across datasets.
[Figure 5, Sec 4.2] The comparison in Figure 5 is not apples-to-apples because ThinkLess is evaluated with Top@k (k = Token Budget / 512, so k >= 2 for budgets above 512) while the Distill baseline is evaluated with Top@1. Top@k counts a response as correct if any of the k candidates is correct, which is a strictly more lenient metric even when the total token budgets are matched. To support the claim of 'Enhanced Accuracy under Comparable Token Budgets,' the Distill baseline should also be evaluated with Top@k under the same candidate-generation scheme, or ThinkLess should be compared at Top@1 with an equivalent token budget.
[Sec 4.2, Table 1] All results in Table 1 come from a single run, as stated in Sec 4.2, and no error bars or significance tests are provided. Since several per-dataset differences are large (e.g., MMLU drops of 3.8 points for Qwen2.5-14B and GPQA gains of 10.1 points for Qwen2.5-7B), the 'comparable accuracy' claim is not robustly established. Please report means and standard deviations across at least three runs, or provide statistical tests.

minor comments (6)

[Abstract, Sec 3.2] There are numerous typos and misspellings, including 'Atttention' in the abstract, 'discruption casued', 'analzye', 'migrration', and 'casued' throughout. A thorough proofreading pass is needed.
[Intro vs Sec 3.3] The placement of the regulation instruction is described inconsistently: the Introduction says 'appending a small instruction prompt after early termination,' while Sec 3.3 says 'we prepend a short instruction prompt.' Please clarify the actual placement (before the question or after </think>) and use consistent wording.
[Sec 4.4] The text claims the method works 'without any fine-tuning, distillation, or prompt engineering,' yet the method relies on hand-crafted task-specific instruction prompts in Tables 4 and 5. Please reconcile this claim or soften it, especially given the Limitations section acknowledges that designing effective instructions may require manual tuning.
[Table 1] Several numeric entries in Table 1 are merged or missing separators (e.g., '1817.8430.81' and '5523.1769.29' for Qwen2.5-7B Distill). These should be properly separated for readability.
[Figure 3] The caption states the hidden states are 'highly similar (0.9),' but the first-row similarity values are about 0.71-0.80. Please clarify the threshold for 'highly similar' and ensure the text accurately reflects the displayed values.
[Sec 3.2, Figure 2] The attention heatmaps in Figure 2 are from a single GSM8K sample, and the statement that 'similar observations can be found in other models and datasets' is not backed by quantitative evidence. Please provide aggregated attention statistics or additional figures to support the generality of the attention claim.

Circularity Check

1 steps flagged · score 4.0 of 10

One self-confirmatory loop: the semantic-anchor mechanism is supported only by the same accuracy result it is invoked to explain, while the zero-reasoning condition is never measured; the Table 1 accuracy itself is independent.

self definitional [Sec. 3.3, 'Clarification: ThinkLess Without Explicit Reasoning' paragraph]
"We contend, however, that the </think> token serves a deeper function than a mere delimiter. It acts as a semantic anchor —a learned symbolic abstraction that implicitly encodes a compressed representation of the reasoning process. ... This hypothesis is supported by our empirical observations: even when the reasoning trace is entirely omitted, the model frequently produces correct answers, indicating that the cognitive process of reasoning may have been executed internally and silently."

The only evidence offered for the semantic-anchor mechanism is the method's own empirical success: 'even when the reasoning trace is entirely omitted, the model frequently produces correct answers.' That correct-answer behavior is the very phenomenon the semantic-anchor construct was introduced to explain, so the observation cannot independently confirm the construct. No measurement isolates the </think> token's contribution in the deployed zero-reasoning condition: the similarity analysis in Sec. 3.2 inserts </think> only after 16-token segments, and the post-regulation instructions in App. B themselves say 'Solve the math problem step by step' and 'think step by step, self-check', providing an alternative explanation for the correct answers.

full rationale

Table 1's Top-1 accuracy numbers come from external benchmarks (GSM8K, MMLU, GPQA, BBH) and are not fitted to or algebraically implied by the input prompts, so the central efficiency/accuracy claim is self-contained against outside data. The only circular-adjacent step is the mechanistic story: the 'semantic anchor' construct in Sec. 3.3 is supported by the very correct-answer behavior it was invented to explain, and the supporting hidden-state similarity analysis (Sec. 3.2) measures a different condition (terminator inserted after 16-token segments) than the deployed condition (terminator immediately after <think>). The post-regulation instructions in Appendix B also independently prompt step-by-step reasoning and self-checking, so the observed accuracy does not uniquely confirm compression into </think>. This is a partial self-confirmatory loop in the explanation, not in the accuracy measurement itself. No load-bearing self-citation, uniqueness import, or fitted-input-as-prediction pattern is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The paper's core mechanism rests on an unmeasured construct: the semantic anchor at </think> when no reasoning has been generated. The similarity analysis measures a different condition, so the one empirical analysis does not directly support the deployed method. The append-instruction component is an acknowledged domain assumption. These are the main uncharged premises.

free parameters (4)

Analysis segment length = 16 tokens
Sec. 3.2 inserts </think> every 16 tokens for the similarity analysis; no sensitivity study is reported, and the similarity numbers depend on this choice.
Top-k budget divisor = 512
Figure 5 defines k = Token Budget / 512; this choice sets how many parallel candidates are allowed and is not justified or varied.
Termination position = immediately after <think>
The core method truncates at position zero of the reasoning span; the motivating analysis only tested terminator insertion after 16, 32, and later tokens, so the selected position is a design choice without direct measurement.
Maximum token budget = 8k
Table 1 fixes the maximum token budget at 8k; performance and latency comparisons may shift with this budget.

assumptions (4)

domain assumption Causal masking causes reasoning information to migrate toward later reasoning tokens, especially the terminator token.
Stated in Sec. 3.2 and borrowed from Lin et al. (2025); used as the basis for early termination, but only illustrated on a handful of samples.
ad hoc to paper The </think> token acts as a semantic anchor that encodes internalized reasoning even when inserted with no preceding reasoning content.
Introduced in Sec. 3.3 Clarification; no direct evidence is given for the zero-reasoning condition used by the actual method.
domain assumption Appending a short instruction after early termination leverages instruction-following without changing model behavior in other ways.
Core to the output regulation step; instruction effectiveness varies by task and is acknowledged as a limitation in Sec. 6.
domain assumption Benchmark accuracy from a single run without seeds is sufficient to compare methods.
All results are from a single run (Sec. 4.2); differences of 1-2 points may be noise.

invented entities (1)

Semantic anchor at the </think> token
purpose: Explains how skipping visible reasoning preserves accuracy.
Introduced in Sec. 3.3; no external prediction or measurement is given, and the measured zero-reasoning condition is absent.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ThinkLess: A Training-Free Inference-Efficient Method for Reducing Reasoning Redundancy." pith.science (2026). https://pith.science/paper/OVWOYAU2

@misc{pith2026250515684,
  author       = {Pith},
  title        = {Pith review of: ThinkLess: A Training-Free Inference-Efficient Method for Reducing Reasoning Redundancy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OVWOYAU2}},
  note         = {Machine review of arXiv:2505.15684}
}

read the original abstract

While Chain-of-Thought (CoT) prompting improves reasoning in large language models (LLMs), the excessive length of reasoning tokens increases latency and KV cache memory usage, and may even truncate final answers under context limits. We propose ThinkLess, an inference-efficient framework that terminates reasoning generation early and maintains output quality without modifying the model. Atttention analysis reveals that answer tokens focus minimally on earlier reasoning steps and primarily attend to the reasoning terminator token, due to information migration under causal masking. Building on this insight, ThinkLess inserts the terminator token at earlier positions to skip redundant reasoning while preserving the underlying knowledge transfer. To prevent format discruption casued by early termination, ThinkLess employs a lightweight post-regulation mechanism, relying on the model's natural instruction-following ability to produce well-structured answers. Without fine-tuning or auxiliary data, ThinkLess achieves comparable accuracy to full-length CoT decoding while greatly reducing decoding time and memory consumption.

Figures

Figures reproduced from arXiv: 2505.15684 by the authors.

**Figure 1.** GPQA (Rein et al., 2024) accuracy of DeepSeekR1-Distill-LLaMA-8B (Guo et al., 2025) under varying token budgets. Red: ThinkLess (compressed reasoning); Blue: full CoT reasoning.The left part of the legend illustrates the relationship between marker size and latency, the middle part denotes each methods, and the right part presents the maximum accuracy and corresponding latency of each method. long and autoregress… view at source ↗

**Figure 2.** Attention heatmaps across different layers of DeepSeek-R1-Distill-LLaMA-8B on a GSM8K sam [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** We insert a </think> token every 16 tokens in DeepSeek-R1-Distill-Qwen-7B and extract last-layer hidden states. These states are highly similar (0.9) across segments, showing that reasoning adds little new information. The final state is also similar to earlier ones, indicating early convergence and redundancy in later reasoning. Similar observations can be found across other models and datasets. Best view with zoom… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Accuracy of DeepSeek-R1-Distill-Qwen-7B vs. position where </think> is inserted. The benchmark is BBH dataset (Suzgun et al., 2022). ment introduces only marginal new information. Moreover, the similarity between the final </think> state and earlier ones increases st…

**Figure 5.** Figure 5: Top@k accuracy of ThinkLess vs. Top@1 accuracy of DeepSeek-distilled models across datasets and models. We set k = Token Budget 512 to match the token usage on par with distilled models. Legends follow [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Answer overlap between Distill and ThinkLess w/o Instruct. Each pie shows the proportion of “a, b” [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

Are Large Reasoning Models Interruptible?
cs.CL 2025-10 conditional novelty 6.0 of 10

Interrupting large reasoning models or changing the problem mid-thought sharply degrades accuracy, with up to 60% drops and distinct failure modes: reasoning leakage, panic, and self-doubt.
XRPO: Pushing the limits of GRPO with Targeted Exploration and Exploitation
cs.LG 2025-10 conditional novelty 5.0 of 10

XRPO extends GRPO with adaptive rollout allocation, in-context example seeding for unsolved prompts, and novelty-weighted advantages, reporting roughly 1-4% higher accuracy and faster convergence.
Strategic Reflectivism In Intelligent Systems
cs.AI 2025-05 conditional novelty 4.0 of 10

Strategic Reflectivism holds that intelligent systems should allocate reflective reasoning tactically, weighing its benefits against its costs.

Reference graph

Works this paper leans on

42 extracted references · 3 canonical work pages · cited by 3 Pith papers

[1]

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...
[2]

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...
[3]

Simon A Aytes, Jinheon Baek, and Sung Ju Hwang. 2025. Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching. arXiv preprint arXiv:2503.05179

arXiv 2025
[4]

Zhenni Bi, Kai Han, Chuanjian Liu, Yehui Tang, and Yunhe Wang. 2024. Forest-of-thought: Scaling test-time compute for enhancing llm reasoning. arXiv preprint arXiv:2412.09078

arXiv 2024
[5]

Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, and 1 others. 2024. Do not think that much for 2+ 3=? on the overthinking of o1-like llms. arXiv preprint arXiv:2412.21187

arXiv 2024
[6]

Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

arXiv 2021
[7]

Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang. 2023. Towards revealing the mystery behind chain of thought: a theoretical perspective. Advances in Neural Information Processing Systems, 36:70757--70798

work page 2023
[8]

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

arXiv 2024

Show all 42 references

[9]

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

2025 arXiv
[10]

Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. 2024. Token-budget-aware llm reasoning. arXiv preprint arXiv:2412.18547

2024 arXiv
[11]

Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Training large language models to reason in a continuous latent space. arXiv preprint arXiv:2412.06769

2024 arXiv
[12]

Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300

2020 arXiv
[13]

Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, and 1 others. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720

2024 arXiv
[14]

Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199--22213

2022
[15]

Zhihang Lin, Mingbao Lin, Luxi Lin, and Rongrong Ji. 2025. Boosting multimodal large language models with visual tokens withdrawal for rapid inference. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5334--5342

2025
[16]

Zichen Liu, Changyu Chen, Wenjun Li, Tianyu Pang, Chao Du, and Min Lin. 2025 a . There may not be aha moment in r1-zero-like training—a pilot study

2025
[17]

Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. 2025 b . Understanding r1-zero-like training: A critical perspective. arXiv preprint arXiv:2503.20783

2025 arXiv
[18]

Qing Lyu, Shreya Havaldar, Adam Stein, Li Zhang, Delip Rao, Eric Wong, Marianna Apidianaki, and Chris Callison-Burch. 2023. Faithful chain-of-thought reasoning. In The 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific...

2023
[19]

Shuofei Qiao, Yixin Ou, Ningyu Zhang, Xiang Chen, Yunzhi Yao, Shumin Deng, Chuanqi Tan, Fei Huang, and Huajun Chen. 2022. Reasoning with language model prompting: A survey. arXiv preprint arXiv:2212.09597

2022 arXiv
[20]

Ziran Qin, Yuchen Cao, Mingbao Lin, Wen Hu, Shixuan Fan, Ke Cheng, Weiyao Lin, and Jianguo Li. 2025. Cake: Cascading and adaptive kv cache eviction with layer preferences. arXiv preprint arXiv:2503.12491

2025
[21]

Xiaoye Qu, Yafu Li, Zhaochen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, and 1 others. 2025 a . A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond. arXiv preprint arXiv:2503.21614

2025
[22]

Yuxiao Qu, Matthew YR Yang, Amrith Setlur, Lewis Tunstall, Edward Emanuel Beeching, Ruslan Salakhutdinov, and Aviral Kumar. 2025 b . Optimizing test-time compute via meta reinforcement fine-tuning. arXiv preprint arXiv:2503.07572

2025 arXiv
[23]

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

2024
[24]

Yi Shen, Jian Zhang, Jieyun Huang, Shuming Shi, Wenjing Zhang, Jiangze Yan, Ning Wang, Kai Wang, and Shiguo Lian. 2025. Dast: Difficulty-adaptive slow-thinking for large reasoning models. arXiv preprint arXiv:2503.04472

2025
[25]

Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36:8634--8652

2023
[26]

Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Hanjie Chen, Xia Hu, and 1 others. 2025. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419

2025 arXiv
[27]

Mirac Suzgun, Nathan Scales, Nathanael Sch \"a rli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, and 1 others. 2022. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261

2022 arXiv
[28]

Qwen Team. 2025. https://qwenlm.github.io/blog/qwq-32b/ Qwq-32b: Embracing the power of reinforcement learning

2025
[29]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30

2017
[30]

Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171

2022 arXiv
[31]

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

2022
[32]

Siwei Wu, Zhongyuan Peng, Xinrun Du, Tuney Zheng, Minghao Liu, Jialong Wu, Jiachen Ma, Yizhi Li, Jian Yang, Wangchunshu Zhou, and 1 others. 2024. A comparative study on reasoning patterns of openai's o1 model. arXiv preprint arXiv:2410.13639

2024 arXiv
[33]

Heming Xia, Yongqi Li, Chak Tou Leong, Wenjie Wang, and Wenjie Li. 2025. Tokenskip: Controllable chain-of-thought compression in llms. arXiv preprint arXiv:2502.12067

2025
[34]

Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. 2025. Chain of draft: Thinking faster by writing less. arXiv preprint arXiv:2502.18600

2025 arXiv
[35]

An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

2024 arXiv
[36]

Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023 a . Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809--11822

2023
[37]

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023 b . React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR)

2023
[38]

Hai Ye, Mingbao Lin, Hwee Tou Ng, and Shuicheng Yan. 2024. Multi-agent sampling: Scaling inference compute for data synthesis with tree search-based agentic collaboration. arXiv preprint arXiv:2412.17061

2024 arXiv
[39]

Jintian Zhang, Yuqi Zhu, Mengshu Sun, Yujie Luo, Shuofei Qiao, Lun Du, Da Zheng, Huajun Chen, and Ningyu Zhang. 2025 a . Lightthinker: Thinking step-by-step compression. arXiv preprint arXiv:2502.15589

2025
[40]

Wenqi Zhang, Yongliang Shen, Linjuan Wu, Qiuying Peng, Jun Wang, Yueting Zhuang, and Weiming Lu. 2024. Self-contrast: Better reflection through inconsistent solving perspectives. arXiv preprint arXiv:2401.02009

2024 arXiv
[41]

Xiaoying Zhang, Da Peng, Yipeng Zhang, Zonghao Guo, Chengyue Wu, Chi Chen, Wei Ke, Helen Meng, and Maosong Sun. 2025 b . Will pre-training ever end? a first step toward next-generation foundation mllms via self-improving systematic cognition. arXiv preprint arXiv:2503.12303

2025 arXiv
[42]

Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. 2022. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493

2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

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...

[2] [2]

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...

[3] [3]

Simon A Aytes, Jinheon Baek, and Sung Ju Hwang. 2025. Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching. arXiv preprint arXiv:2503.05179

arXiv 2025

[4] [4]

Zhenni Bi, Kai Han, Chuanjian Liu, Yehui Tang, and Yunhe Wang. 2024. Forest-of-thought: Scaling test-time compute for enhancing llm reasoning. arXiv preprint arXiv:2412.09078

arXiv 2024

[5] [5]

Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, and 1 others. 2024. Do not think that much for 2+ 3=? on the overthinking of o1-like llms. arXiv preprint arXiv:2412.21187

arXiv 2024

[6] [6]

Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

arXiv 2021

[7] [7]

Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang. 2023. Towards revealing the mystery behind chain of thought: a theoretical perspective. Advances in Neural Information Processing Systems, 36:70757--70798

work page 2023

[8] [8]

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

arXiv 2024