ThinkReset: Learnable Intermediate Interface Construction for Bounded-Context Long-Horizon Reasoning

Fei Ding; Runhao Liu; Yongkang Zhang; Yuhao Liao; Zijian Zeng

REVIEW 5 major objections 5 minor 26 references

Under a fixed context window, long-horizon reasoning is best understood as an interface-construction problem: the model must write a reusable intermediate state that replaces discarded history and still supports continued solving, and train

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 · deepseek-v4-flash

2026-08-03 00:49 UTC pith:FHD5AWF6

load-bearing objection A plausible and well-scoped empirical paper on training LLMs to write reset states by optimizing post-reset continuation success; the central idea is worth engaging, but the missing reset+RL control and the reward/eval identity keep me from fully trusting the causal story. the 5 major comments →

arxiv 2607.28642 v1 pith:FHD5AWF6 submitted 2026-05-26 cs.AI cs.CLcs.LG

ThinkReset: Learnable Intermediate Interface Construction for Bounded-Context Long-Horizon Reasoning

Fei Ding , Yongkang Zhang , Runhao Liu , Yuhao Liao , Zijian Zeng This is my paper

classification cs.AI cs.CLcs.LG

keywords bounded contextlong-horizon reasoningintermediate interfacechain-of-thoughtreinforcement learningwriteback and resetcontinuation successThinkReset

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 argues that when an LLM must solve a hard problem under a fixed context window, the bottleneck is not how much of the reasoning trace can be preserved, compressed, or scheduled, but whether the model can construct a reusable intermediate interface—a written state that replaces the discarded history and still lets solving continue. It claims that current outcome-reward RL for long chain-of-thought actively encourages premature guessing near the window limit, because the final-answer reward does not reward the ability to keep solving after a reset. To test this, the paper introduces ThinkReset, which triggers a text-space writeback at a context threshold, deletes the old trace, and trains the written state directly by the fraction of eight independent continuations that reach the correct answer after the reset. Across AIME 2024/2025, ZebraLogic, AutoLogi, and GPQA-Diamond, with Qwen3-8B/14B/32B under a 32k window, this consistently outperforms trajectory-compression and test-time-control baselines. If correct, the work reframes bounded-context long-horizon reasoning as an interface-learning problem rather than a memory-management problem.

Core claim

The paper's central claim: bounded-context long-horizon reasoning is an intermediate interface learning problem, and the core object is a reusable intermediate state s, written in text space at a context threshold, that replaces discarded history so later reasoning depends only on x⊕s. The training target Rsucc8 is the fraction of eight independent post-reset continuations reaching the correct answer; optimizing this functional criterion makes the state an interface rather than a summary. The paper also identifies a failure mode of outcome-reward long-chain RL: near window exhaustion, unsolved problems reward guessing, because no term rewards continued solving after reset. ThinkReset's three

What carries the argument

The reusable intermediate interface s: a natural-language state generated from the trigger prefix hi = (x, r1:i, ptrig) when context usage reaches αC, after which the original trace is discarded and all subsequent reasoning sees x⊕s. Its training signal is Rsucc8, the average success of 8 independent continuations sampled from the written-back state; this is a functional reward for 'can solving continue?', not a fidelity or length proxy. The three-stage pipeline (cold-start SFT on 500 examples; RLOO on trajectories that exceed the threshold and end incorrectly, with leave-one-out baseline; second-reset training on samples whose 8 continuations all fail) instantiates the objective and prevent

Load-bearing premise

The training reward Rsucc8 requires an automatic verifier that can decide whether each post-reset continuation reaches the correct answer; all tested benchmarks have checkable answers, and the paper offers no evidence that the method trains without such a verifier.

What would settle it

Run the identical three-stage pipeline on a long-horizon reasoning benchmark with no automatic answer checker, so Rsucc8 cannot be computed; if ThinkReset no longer beats the free-writeback baseline, the method's dependence on verifiable answers is confirmed as load-bearing.

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

If this is right

Under a fixed 32k context window, training for post-reset continuation success consistently beats trajectory compression, length-penalized RL, and test-time entropy control across math and logic benchmarks (AIME 2024/2025, ZebraLogic, AutoLogi, GPQA-Diamond).
The written-back state learns to correct and discard erroneous paths rather than faithfully preserve them, which is the opposite of learned summarization behavior.
The same objective supports hierarchical interfaces: a second reset on failed samples further improves performance on the hardest instances.
Because the method works entirely in text space, it transfers to any autoregressive LLM without latent heads or compression modules.
The gain over a same-trigger free-writeback baseline shows the improvement comes from training the interface, not from the reset action itself.

Where Pith is reading between the lines

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

The interface-construction objective could be applied to agentic and tool-using settings where hand-designed state summarization is common; the paper leaves that extension untested.
The paper's own failed-proxy results (CRP, length rewards, forward consistency, frozen judge) form a spectrum that implies any fidelity-based objective will fail on this problem—a testable prediction for other compression methods.
If the thesis holds, the same Rsucc8-style reward could be defined in latent space, offering a bridge to latent-state methods like continuous thought; the paper does not claim this.
The verifier requirement boundaries the method: on open-ended tasks without checkable answers, a learned or heuristic verifier would be needed, and the paper provides no evidence for that regime.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A plausible and well-scoped empirical paper on training LLMs to write reset states by optimizing post-reset continuation success; the central idea is worth engaging, but the missing reset+RL control and the reward/eval identity keep me from fully trusting the causal story.

read the letter

This is a solid empirical paper with a clear central claim: bounded-context long-horizon reasoning is better framed as learning a reusable intermediate interface than as compressing or scheduling the trajectory. The method is concrete—writeback at a fixed context ratio, discard history, continue from the new text state—and the training objective is genuinely different from prior compression or summarization work: it directly optimizes the fraction of 8 post-reset continuations that reach the correct answer. That is not the same as reconstruction fidelity or length control, and the three-stage pipeline with cold-start SFT and easy-sample filtering is a plausible recipe.

What the paper does well: the main comparison against the shared-trigger free-writeback baseline is clean. Both use the same trigger, prompt, stopping rule, and context window; the only difference is whether the written-back state is trained with RL. The gains are consistent across 8B, 14B, and 32B models and across five benchmarks, with bootstrap CIs and multiple seeds. The appendix also documents six failed alternative objectives honestly, which is more than most papers do.

The soft spots are real but not fatal. The biggest one is the missing control cell: reset plus RL with an alternative reward. The paper compares reset without RL and RL without reset, but never holds reset+RL fixed while swapping Rsucc8 for, say, a length penalty or a fidelity proxy. That means the improvement could come from the combination of reset and any RL objective, not specifically from "continuation success." The appendix failures are self-reported and not run under the same evaluation protocol or with confidence intervals, so they don't close this gap. This is the load-bearing weakness in the causal story.

Second, Rsucc8 is identical in form to the evaluation metric Avg@8. That is not circular—you need a training signal—but it means the model is directly selected on the exact metric being reported, and it requires an automatic verifier for every benchmark. The verifier assumption is structural: no verifier, no reward, no method. Third, the trigger threshold alpha is tuned on AIME 2025, which is also in the main table; that is a minor selection issue but should be flagged. Finally, no code or data are released, and the appendix config omits several training details, so reproduction is currently impossible.

Overall: the central thesis is plausible, the method is sensible, and the empirical pattern is encouraging, but the missing reset+RL control means the paper does not yet prove that continuation success—rather than reset-plus-any-reinforcement—is what drives the gains. That is fixable. I would send this to a serious referee, but with a clear request for that control cell and for artifacts before accepting.

Referee Report

5 major / 5 minor

Summary. The paper argues that bounded-context long-horizon reasoning should be viewed not as a trajectory-compression or test-time-control problem, but as an intermediate-interface construction problem: the model must learn to write a reusable text-space state that can replace the discarded reasoning history and still support continued solving. The proposed method, ThinkReset, triggers a writeback at a fixed context threshold (αC), replaces the trace with the written state, and trains the writeback with RLOO using Rsucc8 (Eq. 2), the success rate over 8 independent continuations after reset. Training proceeds in three stages: cold-start SFT on 500 examples, RLOO after the first reset on filtered difficult trajectories, and a second-reset stage on samples where all post-reset continuations fail. Experiments on AIME 2024/2025, ZebraLogic, AutoLogi, and GPQA-Diamond with Qwen3-8B/14B/32B report consistent improvements over fixed-ratio trigger + free writeback, length-penalized RLOO, TokenSkip†, and Halo, with 95% bootstrap CIs and paired-bootstrap significance claims.

Significance. If the empirical claims hold, the paper offers a useful reframing and a practical training objective for bounded-context reasoning. The strongest positive evidence is the comparison against fixed-ratio trigger + free writeback, which holds the trigger rule, prompt, stopping rule, and context window fixed and isolates the effect of RL on the written state. The cross-model consistency (8B/14B/32B) and the inclusion of confidence intervals and paired-bootstrap tests are commendable. The paper also explicitly identifies a plausible failure mode of outcome-reward long-chain RL near window exhaustion. However, the manuscript is currently under-supported for its central causal claim: it lacks a key control cell (reset + RL with an alternative reward), reports failed alternative rewards only qualitatively, selects a hyperparameter on the main benchmark, and provides incomplete training details without code or data. These issues are addressable and do not require abandoning the thesis, but they do require additional experiments and reporting before the central claim can be accepted at the standard of a journal publication.

major comments (5)

[§4.2, Table 1; §3.2, Eq. (2)] The causal attribution to the Rsucc8 objective is not established because the design lacks a reset+RL control cell with an alternative reward. Fixed-ratio trigger + free writeback isolates the presence of RL, and length-penalized RLOO isolates the presence of reset, but no condition holds reset+RL fixed while swapping Rsucc8 for a different reward. The failed attempts in Appendices H/I/J/K are described only qualitatively and are not evaluated under the same protocol, training budget, or confidence intervals. Please add a quantitative baseline such as reset+RL with a length penalty or a local-fidelity reward, and report it in Table 1 or an equivalent table.
[§4.4, Table 3] α=3/4 is selected from the AIME 2025 benchmark, and the same benchmark appears in the main results with this value. This creates a risk of selection bias. If the sensitivity analysis is used to choose α, it should be performed on a held-out validation split or the main tables should report results for a pre-specified α. Please clarify whether α is varied only at evaluation time or also during training, and report the selection procedure explicitly.
[§3.2, Eq. (2); §5] The training objective Rsucc8 requires an automatic verifier that can decide whether each of the 8 independent continuations reaches the correct final answer. All benchmarks used in the paper have checkable answers, but the Limitations section does not mention this structural assumption. Without a verifier, or a learned/approximate surrogate, the three-stage pipeline is undefined. The paper should state this scope condition and, ideally, provide evidence about sensitivity to verifier noise or partial verifiability.
[Appendix L] The 'Training and Evaluation Configuration Overview' omits many details that are essential for reproducing an empirical RL paper: learning rate, batch size, number of RL update steps, number of epochs for cold-start SFT, construction of the 500 SFT examples, GPU budget, and exact context-length accounting. The Reproducibility Statement only says code and materials 'will be released.' For a paper whose central claim is empirical, these missing details and the absence of released artifacts are a significant barrier to verification.
[Appendix N; Appendices E–K] The claim that 'only directly optimizing continuation success worked' rests on six failed experiments reported qualitatively in Appendices E–K and synthesized in Appendix N. These failures are not quantified with the same evaluation protocol, training budgets, or confidence intervals as the main tables, so the claim that the failure pattern is 'systematic' is not established at the same evidential standard. Please provide at least compact quantitative comparisons (e.g., a summary table with Avg@8 for each failed objective) or clearly label these as anecdotal.

minor comments (5)

[§3.3, Stage 3] Please specify how the second-reset reward is aggregated when multiple failed first-reset continuations are present, and whether the same G′=16 sampling protocol and leave-one-out baseline are reused. Currently the text says the 'same reward definition is reused' but does not give the grouped-sampling details.
[Appendix L] Decoding parameters are listed as Temperature=0.6, TopP=0.95, TopK=20, MinP=0. Please define MinP and report the tokenization scheme used for the 32k context window (e.g., tokens vs. characters).
[Table 3] The column headers '1/2', '2/3', etc. denote α values, but this is not stated in the caption. Please make the meaning of each column explicit and clarify whether the training configuration is truly identical across all settings.
[Figure 1] The figure caption and the annotated labels (e.g., 'Wrong hypothesis', '[RESET]') are too terse. Please expand the caption to explain the flow and the meaning of the reset token in the interface-construction panel.
[Appendix C, Table 4] The 14B and 32B tables do not include a column naming the methods; they rely on the row order from Table 1. Please add method names to each row for readability.

Circularity Check

0 steps flagged

No significant circularity; the main empirical claim is directly optimized by Rsucc8, but this is reward-metric alignment rather than a derivation-level circular step.

full rationale

THINKRESET's reasoning chain is empirical rather than derivational: the paper proposes a reformulation, instantiates it with a three-stage RL pipeline, and validates it on benchmarks. I found no load-bearing self-citation, no imported uniqueness theorem, and no ansatz smuggled in via citation. The one formal near-identity is that the training reward Rsucc8 (Eq. 2) is the same quantity as the Avg@8 evaluation metric; Appendix L explicitly states that "training rewards also use the average over 8 continuations, keeping the training objective aligned with evaluation." That is an objective-alignment choice, not a fit-then-predict step: the paper does not derive a prediction from a fitted parameter, and the comparison against baselines is a genuine (if imperfect) empirical test. The defensible criticism is attribution: because there is no reset+RL baseline with a different reward, gains cannot be uniquely assigned to the interface-construction objective rather than to the reset+RL combination; this is an experimental-design gap, not circularity. Appendix K's discussion of a frozen external judge explicitly addresses the self-evaluation concern and explains the trade-off, and the failed-experiment appendix is self-reported evidence, which affects evidential weight but not circularity. Score 2 reflects these minor evidential weaknesses while concluding that no claim reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central contribution is an empirical training method rather than a derivation. The reported improvements rest on choices that are not derived from first principles: the trigger ratio alpha, the 8-continuation reward definition, the 500 cold-start examples, the 16-sample RLOO batches, and the 32k context window. These are selected by the authors; the paper provides sensitivity data only for alpha, and only on one benchmark.

free parameters (6)

alpha (trigger ratio) = 3/4
Fixed-ratio trigger threshold; selected from Table 3 sensitivity on AIME 2025, which is also a main reported benchmark.
N=8 continuations for Rsucc8 and Avg@8 = 8
The reward and evaluation metric are defined by 8 independent continuations; no sensitivity analysis is reported for this choice.
Cold-start SFT example count = 500
Manually designed examples used to stabilize initial writeback behavior; composition of these examples is not specified.
G'=16 sampled trajectories per problem = 16
Stage 2 sampling budget; if fewer than 2 retained trajectories, the problem is skipped for the RLOO update.
Context window C = 32k
All methods use the same 32k window; the method's value depends on this bounded-context setting.
Decoding parameters = Temp=0.6, TopP=0.95, TopK=20, MinP=0
Used for both training and evaluation; not swept.

axioms (5)

standard math The RLOO policy-gradient estimator (Eq. 3) is a valid estimator of the expected reward gradient.
Uses REINFORCE with a leave-one-out baseline following Ahmadian et al.; standard and not the paper's contribution.
domain assumption A sample of 8 continuations estimates continuation success well enough to serve as a training reward.
No variance analysis is given; both training and evaluation use 8 samples, so the metric is stable only up to sampling noise.
domain assumption Final answers on all benchmarks can be automatically checked.
Needed to compute 1[pi_theta(x xor s => a_star)] in Eq. (2); not true for many open-ended or unverifiable tasks.
ad hoc to paper Filtering out trajectories that solve correctly before the threshold prevents degeneration toward trajectory retention.
Behavioral assumption in Section 3.3; no analysis shows that including such samples would degrade interface quality.
domain assumption Outcome-reward long-chain RL near window exhaustion systematically encourages premature guessing.
Stated in Sections 1 and 3.3 as an identified failure mode, but no empirical or formal demonstration is provided; the narrative motivates the method.

pith-pipeline@v1.3.0-alltime-deepseek · 17030 in / 11439 out tokens · 104536 ms · 2026-08-03T00:49:08.393493+00:00 · methodology

0 comments

read the original abstract

Long chain-of-thought reasoning improves performance on complex problems, but it also introduces redundancy accumulation, context overflow, and error anchoring. We argue that under bounded context windows, the core bottleneck is not trajectory compression or test-time control, but the absence of a reusable intermediate interface that can replace discarded history and support continued solving. We further identify a key failure mode of outcome-reward-driven long-chain reinforcement learning: when the model has not solved the task before the window is nearly exhausted, the final-answer reward encourages premature guessing rather than continued careful reasoning. We propose ThinkReset, a text-space instantiation of this view. ThinkReset explicitly constructs reusable intermediate interfaces through interface writeback and reset, and directly optimizes post-reset continuation success. Across multiple long-horizon reasoning benchmarks, this perspective consistently improves success rates under fixed context windows.

Figures

Figures reproduced from arXiv: 2607.28642 by Fei Ding, Runhao Liu, Yongkang Zhang, Yuhao Liao, Zijian Zeng.

discussion (0)

Reference graph

Works this paper leans on

26 extracted references · 8 linked inside Pith

[1]

Back to basics: Revisiting REINFORCE -style optimization for learning from human feedback in LLM s

Arash Ahmadian, Chris Cremer, Matthias Gall \'e , Marzieh Fadaee, Julia Kreutzer, Olivier Pietquin, Ahmet \"U st \"u n, and Sara Hooker. Back to basics: Revisiting REINFORCE -style optimization for learning from human feedback in LLM s. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Proceedings of the 62nd Annual Meeting of the Association for C...

2024
[2]

Think deep, not just long: Measuring llm reasoning effort via deep-thinking tokens, 2026

Wei-Lin Chen, Liqian Peng, Tian Tan, Chao Zhao, Blake JianHang Chen, Ziqian Lin, Alec Go, and Yu Meng. Think deep, not just long: Measuring llm reasoning effort via deep-thinking tokens, 2026. URL https://arxiv.org/abs/2602.13517

Pith/arXiv arXiv 2026
[3]

Metacognitive reuse: Turning recurring llm reasoning into concise behaviors, 2025

Aniket Didolkar, Nicolas Ballas, Sanjeev Arora, and Anirudh Goyal. Metacognitive reuse: Turning recurring llm reasoning into concise behaviors, 2025. URL https://arxiv.org/abs/2509.13237

arXiv 2025
[4]

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, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, ...

2025
[5]

Training large language models to reason in a continuous latent space

Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason E Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space. In Workshop on Reasoning and Planning for Large Language Models, 2025. URL https://openreview.net/forum?id=KrWSrrYGpT

2025
[6]

Deepmath-103k: A large-scale, challenging, decontaminated, and verifiable mathematical dataset for advancing reasoning, 2025

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. Deepmath-103k: A large-scale, challenging, decontaminated, and verifiable mathematical dataset for advancing reasoning, 2025. URL https://arxiv.org/abs/2504.11456

Pith/arXiv arXiv 2025
[7]

Thinkprune: Pruning long chain-of-thought of LLM s via reinforcement learning

Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang. Thinkprune: Pruning long chain-of-thought of LLM s via reinforcement learning. Transactions on Machine Learning Research, 2026. ISSN 2835-8856. URL https://openreview.net/forum?id=V51gPu1uQD

2026
[8]

Making slow thinking faster: Compressing LLM chain-of-thought via step entropy

Zeju Li, Jianyuan Zhong, Ziyang Zheng, Xiangyu Wen, Zhijian Xu, Yingying Cheng, Fan Zhang, and Qiang Xu. Making slow thinking faster: Compressing LLM chain-of-thought via step entropy. In The Fourteenth International Conference on Learning Representations, 2026 a . URL https://openreview.net/forum?id=cGLqQfS5wH

2026
[9]

Limited reasoning space: The cage of long-horizon reasoning in llms, 2026 b

Zhenyu Li, Guanlin Wu, Cheems Wang, and Yongqiang Zhao. Limited reasoning space: The cage of long-horizon reasoning in llms, 2026 b . URL https://arxiv.org/abs/2602.19281

arXiv 2026
[10]

The pensieve paradigm: Stateful language models mastering their own context

Xiaoyuan Liu, Tian Liang, Dongyang Ma, Deyu Zhou, Haitao Mi, Pinjia He, and Yan Wang. The pensieve paradigm: Stateful language models mastering their own context. In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=GymjF88oGQ

2026
[11]

Beyond context limits: Subconscious threads for long-horizon reasoning, 2025

Hongyin Luo, Nathaniel Morgan, Tina Li, Derek Zhao, Ai Vy Ngo, Philip Schroeder, Lijie Yang, Assaf Ben-Kish, Jack O'Brien, and James Glass. Beyond context limits: Subconscious threads for long-horizon reasoning, 2025. URL https://arxiv.org/abs/2507.16784

Pith/arXiv arXiv 2025
[12]

Reasoning models can be effective without thinking, 2025

Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. Reasoning models can be effective without thinking, 2025. URL https://arxiv.org/abs/2504.09858

Pith/arXiv arXiv 2025
[13]

Vicky Zhao, Lili Qiu, and Dongmei Zhang

Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor R \"u hle, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, and Dongmei Zhang. LLML ingua-2: Data distillation for efficient and faithful task-agnostic prompt compression. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Findings of the Association ...

2024
[14]

A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond, 2025

Xiaoye Qu, Yafu Li, Zhao-Chen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, Peng Li, Wei Wei, Jing Shao, Chaochao Lu, Yue Zhang, Xian-Sheng Hua, Bowen Zhou, and Yu Cheng. A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond, 2025. URL https://arxiv.org/abs/2503.21614

arXiv 2025
[15]

Stop overthinking: A survey on efficient reasoning for large language models

Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Na Zou, Hanjie Chen, and Xia Hu. Stop overthinking: A survey on efficient reasoning for large language models. Transactions on Machine Learning Research, 2025. ISSN 2835-8856. URL https://openreview.net/forum?id=HvoG8SxggZ

2025
[16]

Think silently, think fast: Dynamic latent compression of llm reasoning chains

Wenhui Tan, Jiaze Li, Jianzhong Ju, Zhenbo Luo, Ruihua Song, and Jian Luan. Think silently, think fast: Dynamic latent compression of llm reasoning chains. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen (eds.), Advances in Neural Information Processing Systems, volume 38, pp.\ 4646--4668. Curran Associates, Inc., 2025. ...

arXiv 2025
[17]

Towards efficient large language reasoning models via extreme-ratio chain-of-thought compression, 2026

Yuntian Tang, Bohan Jia, Wenxuan Huang, Lianyue Zhang, Jiao Xie, Wenxi Li, Wei Li, Jie Hu, Xinghao Chen Rongrong Ji, and Shaohui Lin. Towards efficient large language reasoning models via extreme-ratio chain-of-thought compression, 2026. URL https://arxiv.org/abs/2602.08324

Pith/arXiv arXiv 2026
[18]

Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Congcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, Haiqing Guo, Han Zhu, Hao Ding, Hao Hu, Hao Yang, Hao Zhang, Haotian Yao, Haotian Zhao, Haoyu Lu, Haoze Li, Haoz...

Pith/arXiv arXiv 2025
[19]

Compass: Enhancing agent long-horizon reasoning with evolving context, 2025

Guangya Wan, Mingyang Ling, Xiaoqi Ren, Rujun Han, Sheng Li, and Zizhao Zhang. Compass: Enhancing agent long-horizon reasoning with evolving context, 2025. URL https://arxiv.org/abs/2510.08790

arXiv 2025
[20]

R1-compress: Long chain-of-thought compression via chunk compression and search

Yibo Wang, Haotian Luo, Huanjin Yao, Tiansheng Huang, Haiying He, Rui Liu, Naiqiang Tan, Jiaxing Huang, Xiaochun Cao, Dacheng Tao, and Li Shen. R1-compress: Long chain-of-thought compression via chunk compression and search. In First Workshop on Foundations of Reasoning in Language Models, 2025. URL https://openreview.net/forum?id=H4UTzUhspV

2025
[21]

Chain-of-thought prompting elicits reasoning in large language models

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (eds.), Advances in Neural Information Processing Systems, volume 35, pp.\ 24824--24837. Curran Associa...

2022
[22]

Resum: Unlocking long-horizon search intelligence via context summarization, 2026

Xixi Wu, Kuan Li, Yida Zhao, Liwen Zhang, Litu Ou, Huifeng Yin, Zhongwang Zhang, Xinmiao Yu, Dingchu Zhang, Yong Jiang, Pengjun Xie, Fei Huang, Minhao Cheng, Shuai Wang, Hong Cheng, and Jingren Zhou. Resum: Unlocking long-horizon search intelligence via context summarization, 2026. URL https://arxiv.org/abs/2509.13313

arXiv 2026
[23]

T oken S kip: Controllable chain-of-thought compression in LLM s

Heming Xia, Chak Tou Leong, Wenjie Wang, Yongqi Li, and Wenjie Li. T oken S kip: Controllable chain-of-thought compression in LLM s. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (eds.), Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp.\ 3351--3363, Suzhou, China, November 2025....

doi:10.18653/v1/2025.emnlp-main.165 2025
[24]

Qwen3 technical report, 2025

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, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

Pith/arXiv arXiv 2025
[25]

Recursive models for long-horizon reasoning, 2026

Chenxiao Yang, Nathan Srebro, and Zhiyuan Li. Recursive models for long-horizon reasoning, 2026. URL https://arxiv.org/abs/2603.02112

Pith/arXiv arXiv 2026
[26]

L ight T hinker: Thinking step-by-step compression

Jintian Zhang, Yuqi Zhu, Mengshu Sun, Yujie Luo, Shuofei Qiao, Lun Du, Da Zheng, Huajun Chen, and Ningyu Zhang. L ight T hinker: Thinking step-by-step compression. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (eds.), Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp.\ 13307--133...

doi:10.18653/v1/2025.emnlp-main.673 2025

[1] [1]

Back to basics: Revisiting REINFORCE -style optimization for learning from human feedback in LLM s

Arash Ahmadian, Chris Cremer, Matthias Gall \'e , Marzieh Fadaee, Julia Kreutzer, Olivier Pietquin, Ahmet \"U st \"u n, and Sara Hooker. Back to basics: Revisiting REINFORCE -style optimization for learning from human feedback in LLM s. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Proceedings of the 62nd Annual Meeting of the Association for C...

2024

[2] [2]

Think deep, not just long: Measuring llm reasoning effort via deep-thinking tokens, 2026

Wei-Lin Chen, Liqian Peng, Tian Tan, Chao Zhao, Blake JianHang Chen, Ziqian Lin, Alec Go, and Yu Meng. Think deep, not just long: Measuring llm reasoning effort via deep-thinking tokens, 2026. URL https://arxiv.org/abs/2602.13517

Pith/arXiv arXiv 2026

[3] [3]

Metacognitive reuse: Turning recurring llm reasoning into concise behaviors, 2025

Aniket Didolkar, Nicolas Ballas, Sanjeev Arora, and Anirudh Goyal. Metacognitive reuse: Turning recurring llm reasoning into concise behaviors, 2025. URL https://arxiv.org/abs/2509.13237

arXiv 2025

[4] [4]

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, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, ...

2025

[5] [5]

Training large language models to reason in a continuous latent space

Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason E Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space. In Workshop on Reasoning and Planning for Large Language Models, 2025. URL https://openreview.net/forum?id=KrWSrrYGpT

2025

[6] [6]

Deepmath-103k: A large-scale, challenging, decontaminated, and verifiable mathematical dataset for advancing reasoning, 2025

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. Deepmath-103k: A large-scale, challenging, decontaminated, and verifiable mathematical dataset for advancing reasoning, 2025. URL https://arxiv.org/abs/2504.11456

Pith/arXiv arXiv 2025

[7] [7]

Thinkprune: Pruning long chain-of-thought of LLM s via reinforcement learning

Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang. Thinkprune: Pruning long chain-of-thought of LLM s via reinforcement learning. Transactions on Machine Learning Research, 2026. ISSN 2835-8856. URL https://openreview.net/forum?id=V51gPu1uQD

2026

[8] [8]

Making slow thinking faster: Compressing LLM chain-of-thought via step entropy

Zeju Li, Jianyuan Zhong, Ziyang Zheng, Xiangyu Wen, Zhijian Xu, Yingying Cheng, Fan Zhang, and Qiang Xu. Making slow thinking faster: Compressing LLM chain-of-thought via step entropy. In The Fourteenth International Conference on Learning Representations, 2026 a . URL https://openreview.net/forum?id=cGLqQfS5wH

2026

[9] [9]

Limited reasoning space: The cage of long-horizon reasoning in llms, 2026 b

Zhenyu Li, Guanlin Wu, Cheems Wang, and Yongqiang Zhao. Limited reasoning space: The cage of long-horizon reasoning in llms, 2026 b . URL https://arxiv.org/abs/2602.19281

arXiv 2026

[10] [10]

The pensieve paradigm: Stateful language models mastering their own context

Xiaoyuan Liu, Tian Liang, Dongyang Ma, Deyu Zhou, Haitao Mi, Pinjia He, and Yan Wang. The pensieve paradigm: Stateful language models mastering their own context. In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=GymjF88oGQ

2026

[11] [11]

Beyond context limits: Subconscious threads for long-horizon reasoning, 2025

Hongyin Luo, Nathaniel Morgan, Tina Li, Derek Zhao, Ai Vy Ngo, Philip Schroeder, Lijie Yang, Assaf Ben-Kish, Jack O'Brien, and James Glass. Beyond context limits: Subconscious threads for long-horizon reasoning, 2025. URL https://arxiv.org/abs/2507.16784

Pith/arXiv arXiv 2025

[12] [12]

Reasoning models can be effective without thinking, 2025

Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. Reasoning models can be effective without thinking, 2025. URL https://arxiv.org/abs/2504.09858

Pith/arXiv arXiv 2025

[13] [13]

Vicky Zhao, Lili Qiu, and Dongmei Zhang

Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor R \"u hle, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, and Dongmei Zhang. LLML ingua-2: Data distillation for efficient and faithful task-agnostic prompt compression. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Findings of the Association ...

2024

[14] [14]

A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond, 2025

Xiaoye Qu, Yafu Li, Zhao-Chen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, Peng Li, Wei Wei, Jing Shao, Chaochao Lu, Yue Zhang, Xian-Sheng Hua, Bowen Zhou, and Yu Cheng. A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond, 2025. URL https://arxiv.org/abs/2503.21614

arXiv 2025

[15] [15]

Stop overthinking: A survey on efficient reasoning for large language models

Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Na Zou, Hanjie Chen, and Xia Hu. Stop overthinking: A survey on efficient reasoning for large language models. Transactions on Machine Learning Research, 2025. ISSN 2835-8856. URL https://openreview.net/forum?id=HvoG8SxggZ

2025

[16] [16]

Think silently, think fast: Dynamic latent compression of llm reasoning chains

Wenhui Tan, Jiaze Li, Jianzhong Ju, Zhenbo Luo, Ruihua Song, and Jian Luan. Think silently, think fast: Dynamic latent compression of llm reasoning chains. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen (eds.), Advances in Neural Information Processing Systems, volume 38, pp.\ 4646--4668. Curran Associates, Inc., 2025. ...

arXiv 2025

[17] [17]

Towards efficient large language reasoning models via extreme-ratio chain-of-thought compression, 2026

Yuntian Tang, Bohan Jia, Wenxuan Huang, Lianyue Zhang, Jiao Xie, Wenxi Li, Wei Li, Jie Hu, Xinghao Chen Rongrong Ji, and Shaohui Lin. Towards efficient large language reasoning models via extreme-ratio chain-of-thought compression, 2026. URL https://arxiv.org/abs/2602.08324

Pith/arXiv arXiv 2026

[18] [18]

Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Congcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, Haiqing Guo, Han Zhu, Hao Ding, Hao Hu, Hao Yang, Hao Zhang, Haotian Yao, Haotian Zhao, Haoyu Lu, Haoze Li, Haoz...

Pith/arXiv arXiv 2025

[19] [19]

Compass: Enhancing agent long-horizon reasoning with evolving context, 2025

Guangya Wan, Mingyang Ling, Xiaoqi Ren, Rujun Han, Sheng Li, and Zizhao Zhang. Compass: Enhancing agent long-horizon reasoning with evolving context, 2025. URL https://arxiv.org/abs/2510.08790

arXiv 2025

[20] [20]

R1-compress: Long chain-of-thought compression via chunk compression and search

Yibo Wang, Haotian Luo, Huanjin Yao, Tiansheng Huang, Haiying He, Rui Liu, Naiqiang Tan, Jiaxing Huang, Xiaochun Cao, Dacheng Tao, and Li Shen. R1-compress: Long chain-of-thought compression via chunk compression and search. In First Workshop on Foundations of Reasoning in Language Models, 2025. URL https://openreview.net/forum?id=H4UTzUhspV

2025

[21] [21]

Chain-of-thought prompting elicits reasoning in large language models

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (eds.), Advances in Neural Information Processing Systems, volume 35, pp.\ 24824--24837. Curran Associa...

2022

[22] [22]

Resum: Unlocking long-horizon search intelligence via context summarization, 2026

Xixi Wu, Kuan Li, Yida Zhao, Liwen Zhang, Litu Ou, Huifeng Yin, Zhongwang Zhang, Xinmiao Yu, Dingchu Zhang, Yong Jiang, Pengjun Xie, Fei Huang, Minhao Cheng, Shuai Wang, Hong Cheng, and Jingren Zhou. Resum: Unlocking long-horizon search intelligence via context summarization, 2026. URL https://arxiv.org/abs/2509.13313

arXiv 2026

[23] [23]

T oken S kip: Controllable chain-of-thought compression in LLM s

Heming Xia, Chak Tou Leong, Wenjie Wang, Yongqi Li, and Wenjie Li. T oken S kip: Controllable chain-of-thought compression in LLM s. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (eds.), Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp.\ 3351--3363, Suzhou, China, November 2025....

doi:10.18653/v1/2025.emnlp-main.165 2025

[24] [24]

Qwen3 technical report, 2025

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, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

Pith/arXiv arXiv 2025

[25] [25]

Recursive models for long-horizon reasoning, 2026

Chenxiao Yang, Nathan Srebro, and Zhiyuan Li. Recursive models for long-horizon reasoning, 2026. URL https://arxiv.org/abs/2603.02112

Pith/arXiv arXiv 2026

[26] [26]

L ight T hinker: Thinking step-by-step compression

Jintian Zhang, Yuqi Zhu, Mengshu Sun, Yujie Luo, Shuofei Qiao, Lun Du, Da Zheng, Huajun Chen, and Ningyu Zhang. L ight T hinker: Thinking step-by-step compression. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (eds.), Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp.\ 13307--133...

doi:10.18653/v1/2025.emnlp-main.673 2025