Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

A small compressor can write the reasoning chain as a few dense 'thought' vectors, and a 7B executor can answer with about half the tokens while staying within a few accuracy points of the uncompressed model.

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

2026-08-04 10:50 UTC pith:SE34D574

load-bearing objection Solid empirical contribution to CoT compression; the mechanism story is overstated and the headline numbers hide a real accuracy drop, but the method and ablations deserve review. the 4 major comments →

arxiv 2510.08647 v2 pith:SE34D574 submitted 2025-10-09 cs.CL cs.AI

Can Reasoning Path still be Effective as Input? Bridging Post-Reasoning to Chain-of-Thought Compression

Chengzhengxu Li , Xiaoming Liu , Zhaohan Zhang , Shengchao Liu , Guoxin Ma , Yu Lan , Cong Wang , Chao Shen This is my paper
classification cs.CL cs.AI
keywords Chain-of-Thought CompressionContinuous EmbeddingsSoft Token PromptingCompressor-Executor FrameworkReasoning EfficiencyLarge Language ModelsPost-Reasoning ParadigmUpfront Thoughts
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.

Chain-of-Thought (CoT) gives language models strong reasoning at the cost of long, token-hungry explanations. This paper asks whether the reasoning can be done up front: a small 'compressor' model converts the reasoning chain into a fixed sequence of dense continuous embeddings—called Upfront Thoughts—and the original large 'executor' model is trained to answer correctly from those embeddings plus a short continuation. The central claim is that this preserves most of the executor's reasoning ability while roughly halving the number of tokens it generates. On GSM8K, Qwen2.5-7B-Instruct drops from 298.6 to 140.4 average output tokens (compression ratio 0.5) with accuracy falling only from 92.17% to 86.55%—a result 3.08 points above the Tokenskip baseline. If correct, the paper shows that CoT compression is better framed as a two-model cooperation problem (compress-then-reason) than as a text-summarization or prompt-design problem.

Core claim

The paper's central discovery is that a chain-of-thought can be compressed into a small set of continuous soft-token embeddings before answering, and a large model can be trained to consume those embeddings and still reason almost as well as if it had written the full chain itself. Concretely, a compressor model is trained with prompt-tuning placeholders to emit Upfront Thought (UT) embeddings from which it can regenerate the original CoT text; a projector maps the UT into the executor's embedding space; and the executor is trained with loss Le = Lsem × R, where Lsem is the mean absolute error between the executor's hidden states for the compressed input (UT + truncated CoT) and for the orig

What carries the argument

Upfront Thought (UT): a fixed-length sequence of continuous embeddings, taken from the last hidden states of placeholder tokens in a small 'compressor' model, that condenses the reasoning chain. The load-bearing mechanism is the pair of training objectives in the Upfront Thought Utilization stage: the semantic loss Lsem forces the executor's last hidden states for the compressed input (UT plus a truncated short CoT) to match its hidden states for the original full CoT via mean absolute error, and the reward factor R multiplies that loss by the squared difference between the log-probability of the correct answer under the compressed input and under the original CoT. Together they train the ex

Load-bearing premise

The hidden-state alignment Lsem is assumed to capture exactly the information needed for a correct answer, but the paper's own decoded examples contain arithmetic mistakes and swapped question content, so it is not established that the Upfront Thought actually preserves the reasoning needed to solve the problem.

What would settle it

Take a trained UCoT system and replace the UT for a question with (a) the UT of a different question or (b) random vectors of the same length, then measure executor accuracy. If accuracy stays near the reported level, the specific semantic content of the UT is not what drives the result, contradicting the high-fidelity compression claim. A direct check: decode the UT to text and verify it matches the original CoT's arithmetic operations and quantities—the appendix examples already show mismatches such as '1/100 = 10'.

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

If this is right

  • On GSM8K with Qwen2.5-7B-Instruct at compression ratio 0.5, UCoT produces about 53% fewer tokens (140.4 vs 298.6) with accuracy 86.55%, only 5.62 points below the uncompressed model and 3.08 points above Tokenskip.
  • The same pattern holds on Llama-3.1-8B: at ratio 0.5, tokens drop from 212.1 to 101.8 with a 3.64-point accuracy drop, outperforming Tokenskip by 1.82 points.
  • UCoT extends to long-CoT reasoning models: on Deepseek-R1-Distill-Qwen-7B (HumanEval) it cuts tokens 50.67% with only a 5.20-point drop, and on Qwen3-8B (HumanEval) 48.05% with a 5.28-point drop, both beating Tokenskip.
  • Larger executors absorb compression better: Qwen2.5-14B loses only 1.79 points at compression ratio 0.5 on GSM8K, suggesting the method's value grows with model scale.
  • Ablations show both UT and reward factor are necessary (removing UT drops accuracy about 15 points on GSM8K; removing R drops it about 16 points), and accuracy rises monotonically as UT length grows from 16 to 64 tokens.

Where Pith is reading between the lines

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

  • The paper's own decoded-UT examples suggest the UT may convey structural reasoning schemas rather than exact arithmetic content (it decodes to a different question's quantities, with errors like '1/100 = 10'). A testable reading: feed the executor UTs decoded from unrelated questions and see whether accuracy holds; if it does, the accuracy gain comes from a consistent reasoning template, not high-
  • Because the reward factor multiplies the loss by (r_UCoT − r_CoT)^2, training concentrates on examples where the compressed path initially disagrees with the original path's answer likelihood. The method's benefit may therefore be concentrated on a subset of problems; reporting accuracy conditioned on UT-decoding success or problem difficulty would reveal where it actually helps.
  • The UT itself costs M hidden-state computations on the compressor, and the reported speedups already include that cost. A fair deployment comparison should also account for the fixed UT prefix added to every query, especially at high compression ratios (0.9), where the prefix may negate part of the output-token savings.
  • The method is trained only on GSM8K and evaluated zero-shot on AIME, GPQA, ASDiv, and HumanEval. A natural stress test is training on a broader corpus or varying UT length per domain; the paper shows a single fixed-length UT generalizes surprisingly well, which is itself a strong claim worth probing.

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 / 5 minor

Summary. The paper proposes UCoT, a two-stage compressor-executor framework for chain-of-thought (CoT) compression. A small model (compressor) is trained to generate a fixed-length sequence of continuous embeddings (Upfront Thoughts, UT) from a question, and a larger model (executor) is trained to use these UT embeddings to produce a shorter CoT while maintaining accuracy. The first training stage (UTG) optimizes the compressor to reconstruct the original CoT from the UT via a language-modeling loss; the second stage (UTU) trains the executor with a semantic loss that aligns hidden states of the UT-conditioned generation with those of the original CoT, multiplied by a reward factor that penalizes incorrect answers. Experiments on GSM8K, MATH-500, and additional long-CoT tasks (GPQA, HumanEval, AIME, ASDiv) compare UCoT against Prompt, Truncation, CoD, and Tokenskip baselines at compression ratios 0.9, 0.7, and 0.5. The main empirical claim is that UCoT achieves higher accuracy than the SOTA method Tokenskip at comparable or lower token counts, e.g., on GSM8K with Qwen2.5-7B at ratio 0.5, accuracy is 86.55% versus 83.47% for Tokenskip while using fewer tokens (140.36 vs 157.58).

Significance. If the empirical findings hold, UCoT is a practical and relatively simple method for CoT compression that outperforms existing baselines on standard benchmarks while generalizing to long-reasoning models and out-of-domain tasks. The paper's strengths include a clear two-stage training pipeline, extensive comparisons across model families and datasets, ablations showing the importance of the UT and reward factor, and analysis of UT length and information volume. The method is reproducible in principle from the described details. However, the paper's central conceptual contribution—'High-fidelity CoT Information Compression'—is not supported by its own evidence: the decoded UTs in Appendix C.6 are described as providing 'structural information about reasoning rather than semantic information' and contain arithmetic errors. This undermines the claim that Eq. (6)'s semantic loss actually preserves reasoning content, and it calls into question the novelty of the compression mechanism as distinct from a soft structural prompt plus reward-based fine-tuning. The empirical accuracy improvements over Tokenskip are still plausible, but the paper's framing and internal consistency need substan

major comments (4)
  1. [§3.1 vs Appendix C.6] The paper claims in §3.1 that the compressor is trained to 'cover the information of an entire high-quality reasoning path' and lists 'High-fidelity CoT Information Compression' as Contribution 1. However, Appendix C.6 (Figures 4-5) explicitly states that the decoded UTs provide 'structural information about reasoning rather than semantic information' and shows decoded examples with arithmetic errors such as '1/100 = 10'. This is a direct contradiction. Since Eq. (6) (the L_sem loss) is the only mechanism aligning UT-conditioned hidden states with the original CoT, the paper needs to either provide evidence that hidden-state alignment tracks answer-relevant semantic content (e.g., probing, intervention, or controlled decoding experiments) or revise the contribution/framing to a structural-prompt-plus-reward method. As written, the mechanism for the accuracy improvement is unsupported.
  2. [Abstract and §4.1] The abstract and introduction state that UCoT's performance on GSM8K is '3.08% higher than that of the SOTA method.' This phrasing omits two facts that are essential for a fair reading: the comparison is to Tokenskip at compression ratio 0.5, and UCoT is 5.62 points below the uncompressed executor (92.17% to 86.55%). 'Maintains the powerful reasoning ability of executor' is an overstatement given a 5.62-point drop. The relative comparison should be reported alongside the absolute drop from the original model, and the wording should be adjusted to avoid implying that UCoT is 3.08% more accurate than the original.
  3. [Tables 1–3, Appendix A.3] All experimental results are said to be averaged over 5 random seeds, but no standard deviations, confidence intervals, or significance tests are reported. The headline margins are small in several cases (e.g., 0.17 points on HumanEval in Table 3, 0.45 points at ratio 0.9 in Table 1) and may be within seed variance. Please report per-seed variability and, where possible, paired significance tests for the main UCoT-vs-Tokenskip comparisons. This is needed to support the empirical claims quantitatively.
  4. [§4.4, Table 4] Token Gain and Information Gain are computed by decoding the UT with the compressor, but the downstream consumer is the executor, not the compressor. The positive correlation in Table 4 may reflect the compressor's ability to decode its own embeddings rather than the information actually available to the executor. To substantiate the claim that 'information volume in UT matters' for the executor, the authors should provide executor-side diagnostics, e.g., measuring the effect of UT information on executor hidden states or accuracy under controlled compressor checkpoints. The table also has a duplicated 'Latency' column header.
minor comments (5)
  1. [Throughout] Typos and formatting issues: 'intergated' (should be 'integrated'), 'satge' (stage), 'prmopt' (prompt), 'th number' (the number), 'Qwen2.5-7b-Instrust', 'middel' (middle), 'MA TH-500' in Figure 3, and 'Tokens'/'Lantency' in table captions.
  2. [Algorithm 2, line 10] In the pseudocode for Upfront Thought Utilization, line 10 says 'Compute loss L_c' but should be 'Compute loss L_e' (the executor loss defined in Eq. (8)).
  3. [Introduction, Related Work] The citation '(Kang et al.)' in the Introduction lacks a year; it should be '(Kang et al., 2024)' to match the reference list.
  4. [Table 1] The Prompt baseline at ratio 0.9 reports ActRatio 1.02-1.04, indicating it does not actually compress the CoT. The paper acknowledges this in §4.1, but it would be clearer to mark such rows explicitly as failing to meet the target compression ratio.
  5. [Figure 3] The figure legend includes a stray '1' and inconsistent labeling; the y-axis and legend should be cleaned up for readability.

Circularity Check

1 steps flagged

No significant circularity: UCoT's headline accuracy/token numbers are externally benchmarked and not forced by construction; the mild self-referentiality is confined to the §4.4 information-gain diagnostic and one non-load-bearing self-citation.

specific steps
  1. self definitional [§4.4 (Eqs. 10–11), compressor objective Eq. (4); Appendix C.6]
    "The positive correlation between the information volume in UT and the performance of UCoT suggests that our method compresses rich reasoning information into the embedding space of LLMs... We find that the UT provides structural information about reasoning rather than semantic information."

    The 'information volume' metrics in Eqs. (10)-(11) are defined as compressor Mc's token/cross-entropy gain from adding the UT, but Eq. (4) trains Mc to minimize -log P_Mc(Cn|Hn) — i.e., to maximize exactly that decodeability. Reporting rising Token Gain/Information Gain as evidence that UTs 'compress rich reasoning information' thus measures the training objective back as its own confirmation; the diagnostic is self-referential by construction. Appendix C.6 reinforces the gap: decoded UTs contain arithmetic errors ('1/100 = 10') and, per the authors, provide 'structural information about reasoning rather than semantic information,' contradicting the 'semantic'/'high-fidelity' mechanism claimed in §3.1-3.2 and Contribution 1. The headline accuracy numbers in Table 1 remain externally benchm

full rationale

UCoT's central claims — GSM8K/MATH accuracy at set compression ratios, token counts, latency, and gains over Tokenskip/CoD on Qwen2.5-7B and Llama-3.1-8B (Table 1), plus extra-long-CoT results (Table 3) — are measured on standard held-out benchmarks against external baselines, so they do not reduce to fitted parameters or to the paper's own definitions. The 50% token reduction is the method's design target (Cutoff in Eq. 5 shapes the executor's outputs), and the reported ActRatio confirms the model follows the target while retaining accuracy; this is an empirical capability result, not a prediction that equals its input. The compressor-executor loop is a self-distillation loop (executor generates CoTs in Eq. 1; compressor compresses them; executor is then trained to use the UTs), but test accuracy is judged against ground truth, so the loop does not force the outcome. The closest thing to circularity is §4.4: Token Gain and Information Gain are computed with the very compressor that Eq. (4) trained to maximize exactly that decodeability, so the 'information volume' diagnostic partly measures the training objective itself; and Appendix C.6 concedes decoded UTs carry structural rather than semantic information ('1/100 = 10'), which undercuts the 'high-fidelity'/'semantic' mechanism claimed in §3.1-3.2 and Contribution 1. That is a genuine internal-evidence gap (a correctness risk), but it does not make the headline results circular. One minor self-citation exists — Wang et al. (2024) Stumbling Blocks (co-authored by two of the present authors), cited only for perturbation-attack definitions in the Appendix C.3 robustness study — and it is not load-bearing. No uniqueness theorem, ansatz-via-citation, or renamed known result is present.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 1 invented entities

The central claim rests on the compressor being able to pack useful reasoning into a fixed-length continuous representation and on the executor being trained to decode it. The method's main free parameters are the UT length and the compression ratio; the semantic-loss proxy and the transferability of self-generated CoTs are unproven domain assumptions. No genuinely new physical or scientific entity is introduced.

free parameters (4)
  • UT length M = 64 in main experiments; 16/32 analyzed in §4.5
    Number of soft-token embeddings emitted by the compressor. Accuracy improves monotonically with M (Fig. 3), so M is a hand-chosen capacity parameter that directly controls the accuracy-efficiency trade-off.
  • Compression ratio α = 0.9, 0.7, 0.5
    Target output-length budget used in the Cutoff function (Eq. 5). This is an experimental condition rather than a fitted value, but the headline claims are tied to the α=0.5 column.
  • Projector middle-layer width |H_middle| = 2048
    Architectural hyperparameter for the two-layer projector; no ablation or sensitivity analysis is reported.
  • LoRA ranks and scaling = compressor r=8, α=16; executor r=16, α=32
    Standard LoRA hyperparameters used in both training stages; no sensitivity analysis is provided.
axioms (3)
  • domain assumption Last-hidden-state MAE (L_sem) is a valid proxy for semantic equivalence of reasoning paths
    Eq. (6) aligns HUCoT and HCoT in the executor's last hidden layer; no theoretical or independent empirical validation is provided that this distance tracks reasoning correctness.
  • domain assumption A 1.5B model can encode enough of the executor's reasoning into M fixed continuous embeddings to serve as a useful prior
    Core to the compressor-executor design; Appendix C.6 shows decoded UTs contain arithmetic errors, so this assumption holds only partially.
  • domain assumption Training on CoTs generated by the executor itself transfers to test questions and to other tasks without domain-specific fine-tuning
    Training data D is built from Cn = Me(Ze⊕Qn) in §3.1 and zero-shot evaluation is used in §A.5; no analysis of distribution shift is provided.
invented entities (1)
  • Upfront Thought (UT) no independent evidence
    purpose: Continuous soft-token embeddings that carry compressed reasoning information from the compressor to the executor
    Introduced by the method; no external falsifiable prediction. Its informativeness is measured only through internally defined Token Gain/Information Gain metrics and downstream accuracy.

pith-pipeline@v1.3.0-alltime-deepseek · 26911 in / 13322 out tokens · 103832 ms · 2026-08-04T10:50:49.019205+00:00 · methodology

0 comments
read the original abstract

Recent developments have enabled advanced reasoning in Large Language Models (LLMs) via long Chain-of-Thought (CoT), trading efficiency during inference for performance. Existing works focus on compressing generated CoT in reasoning, which impairs the necessary information for deriving the correct answer. In this work, we propose post-reasoning, a reasoning paradigm that takes CoT as a part of context to simplify the reasoning task for LLMs. We find that post-reasoning significantly reduces the generation length of LLMs, but its effectiveness hinges on the efficiency and the reliability of the contextual CoT generation. Therefore, we propose Upfront CoT (UCoT), an efficient post-reasoning framework for CoT compression. UCoT trains a lightweight model (compressor) to provide contextual CoT in form of soft tokens and trains the LLM (executor) to leverage this contextual CoT for producing the final answer. Extensive experiments show that UCoT maintains the powerful reasoning ability of executor while significantly reducing the length of CoT. It is worth mentioning that when applying UCoT to the Qwen2.5-7B-Instruct model, the usage of tokens on GSM8K dataset is reduced by 50%, while the performance is 3.08% higher than that of the state-of-the-art (SOTA) method.

Figures

Figures reproduced from arXiv: 2510.08647 by Chao Shen, Chengzhengxu Li, Cong Wang, Guoxin Ma, Shengchao Liu, Xiaoming Liu, Yu Lan, Zhaohan Zhang.

Figure 1
Figure 1. Figure 1: A comparison of reasoning accuracy and the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Upfront Chain-of-Thought. In Upfront Thought Generation stage (§ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance comparison between original CoT and UTs of different lengths. inferencing the GSM8K dataset, the model’s ac￾curacy improved by 4.75% when the UT length increased from 16 to 64, gradually approaching that of the original CoT reasoning. This result fully demonstrates that as the length of the UT increases, the compressor can provide more complete and accurate reasoning information for the executo… view at source ↗
Figure 4
Figure 4. Figure 4: A case study on GSM8K dataset. Output using Qwen2.5-7B-Instruct as the executor and Qwen2.5-1.5B [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: A case study on Math dataset. Output using Qwen2.5-7B-Instruct as the executor and Qwen2.5-1.5B as [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A case study on GSM8K dataset. Output using Qwen2.5-7B-Instruct as the executor with the CoT [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: A case study on GSM8K dataset. Output using Llama-3.1-8B-Instruct as the executor with the CoT [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: A case study on Math dataset. Output using Qwen2.5-7B-Instruct as the executor with the CoT compression [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: A case study on Math dataset. Output using Llama-3.1-8B-Instruct as the executor with the CoT [PITH_FULL_IMAGE:figures/full_fig_p020_9.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 3 Pith papers

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

  1. ReSum: Synergizing LLM Reasoning and Summarization with Reinforcement Learning

    cs.AI 2026-06 unverdicted novelty 7.0

    ReSum trains LLMs via RLVR to self-summarize reasoning trajectories, yielding 4% average performance gains and 18.6% shorter rollouts through contrastive rollout branches.

  2. ReSum: Synergizing LLM Reasoning and Summarization with Reinforcement Learning

    cs.AI 2026-06 conditional novelty 6.0

    ReSum's contrastive RL branching on self-summarization points improves LLM math reasoning accuracy by about 4% and shortens rollouts by about 18.6% across tested backbones.

  3. When Less is More: The LLM Scaling Paradox in Context Compression

    cs.LG 2026-02 unverdicted novelty 6.0

    Larger LLM compressors in lossy setups often yield less faithful context reconstructions due to knowledge overwriting and semantic drift, with mid-sized models outperforming larger ones across 27 tested configurations.

Reference graph

Works this paper leans on

13 extracted references · 8 linked inside Pith · cited by 2 Pith papers

  1. [2]

    Therefore, her mom's rate is:\[\frac{4}{2} = 2 \text{ pieces per minute}\] 3

    **Calculate her mom's rate of placing pieces:** Her mom can place half as many pieces per minute as Kalinda. Therefore, her mom's rate is:\[\frac{4}{2} = 2 \text{ pieces per minute}\] 3. **Determine their combined rate:** To find their combined rate, we add Kalinda's rate and her mom's rate: \[4 + 2 = 6 \text{ pieces per minute}\] 4. **Calculate the total...

  2. [3]

    Yingqian Cui, Pengfei He, Jingying Zeng, Hui Liu, Xianfeng Tang, Zhenwei Dai, Yan Han, Chen Luo, Jing Huang, Zhen Li, and 1 others

    Training Verifiers to Solve Math Word Prob- lems.arXiv e-prints, arXiv:2110.14168. Yingqian Cui, Pengfei He, Jingying Zeng, Hui Liu, Xianfeng Tang, Zhenwei Dai, Yan Han, Chen Luo, Jing Huang, Zhen Li, and 1 others. 2025. Stepwise perplexity-guided refinement for efficient chain-of- thought reasoning in large language models.arXiv preprint arXiv:2502.13260...

    Pith/arXiv arXiv 2025

  3. [4]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Letman, and 1 others

    Scaling up test-time compute with latent rea- soning: A recurrent depth approach.arXiv preprint arXiv:2502.05171. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Letman, and 1 others. 2024. The Llama 3 Herd of Models.arXiv e-prints, arXiv:2407.21783. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, R...

    Pith/arXiv arXiv 2024

  4. [6]

    single-unit division → single-time summation → single-period multiplication → multi-period multi- plication

    In addition, for the few-shot part of the CoD prompt, we used the original design and do not change it. For Truncation (Xia et al., 2025), we forcibly truncate the full CoT generated by the LLM to the corresponding compression ratio, and then tested the performance of the LLM. For To- kenskip (Xia et al., 2025), we employed its pro- vided short CoT datase...

    2025

  5. [8]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan

    Qwen3 technical report.arXiv preprint arXiv:2505.09388. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan

    Pith/arXiv arXiv

  6. [9]

    Please reduce your thought process by 50 per- cent

    Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822. Junchi Yu, Ran He, and Rex Ying. 2023. Thought propagation: An analogical approach to complex rea- soning with large language models.arXiv preprint arXiv:2310.03965. Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Pe- te...

    Pith/arXiv arXiv 2023

  7. [12]

    **Calculate total time required complete puzzle find time divide total number pieces by combined rate.{Total time (in minutes)} = {360 pieces}}{6 pieces per minute}} = 60 minutes} 5. **Convert total time from minutes to hours 60 minutes in hour.{Total time (in hours)} = {60 minutes}}{60 minutes per hour}} = 1 hour} time for Kalinda and mom to complete puz...

  8. [13]

    - This means he walked \( \frac{10,000}{2} = 5,000 \) steps to and from school

    **Steps Completed Before Jogging**: - He completed half of his steps on his walks to and from school. - This means he walked \( \frac{10,000}{2} = 5,000 \) steps to and from school. - Additionally, he took another 1,000 steps going for a short walk with his friend. - Therefore, the total steps taken before jogging is \( 5,000 + 1,000 = 6,000 \) steps. 3. ...

    2000

  9. [2020]

    InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 975–984

    A diverse corpus for evaluating and developing english math word problem solvers. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 975–984. Sania Nayab, Giulio Rossolini, Marco Simoni, Andrea Saracino, Giorgio Buttazzo, Nicolamaria Manes, and Fabrizio Giacomelli. 2024. Concise thoughts: Impact of output leng...

    Pith/arXiv arXiv 2024

  10. [2021]

    Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wanxiang Che

    Evaluating large language models trained on code. Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wanxiang Che. 2025. Towards reasoning era: A survey of long chain-of-thought for reasoning large language models.arXiv preprint arXiv:2503.09567. Sijia Chen, Baochun Li, and Di Niu. 2024. Boosti...

    Pith/arXiv arXiv 2025

  11. [2023]

    InProceedings of the Eighteenth European Conference on Computer Systems, pages 233–248

    Tabi: An efficient multi-level inference sys- tem for large language models. InProceedings of the Eighteenth European Conference on Computer Systems, pages 233–248. 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 elic- its reasoning in large language models.Adva...

    Pith/arXiv arXiv 2022

  12. [2024]

    Sanjana Ramprasad, Elisa Ferracane, and Zachary C Lipton

    Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression.arXiv preprint arXiv:2403.12968. Sanjana Ramprasad, Elisa Ferracane, and Zachary C Lipton. 2024. Analyzing llm behavior in dialogue summarization: Unveiling circumstantial hallucina- tion trends.arXiv preprint arXiv:2406.03487. David Rein, Betty Li Hou, Asa Cooper S...

    Pith/arXiv arXiv 2024

  13. [2025]

    Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching.arXiv preprint arXiv:2503.05179. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastr...

    arXiv