Pith. sign in

REVIEW 3 major objections 5 minor 37 references

The paper argues that concentrating lookahead compute at confidence drops shifts the accuracy–token-cost frontier: it reports an average 87.3% token reduction with accuracy gains in 11 of 16 settings across four models and four competition-

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

2026-08-02 11:26 UTC pith:2THYUUVG

load-bearing objection A useful efficiency-focused decoding framework with a real token reduction, but the accuracy claims rest on small benchmarks and some reproducibility gaps. the 3 major comments →

arxiv 2607.22602 v1 pith:2THYUUVG submitted 2026-06-12 cs.AI

DeepLook: Deeper Thinking with Lookahead

Tingxin Yang , Zefeng Wang , Mengyue Wang , Xingcheng Zhou , Yunpu Ma This is my paper
classification cs.AI
keywords inference-time scalingtest-time computeconfidence triggeringlookahead rankingtoken efficiencymathematical reasoningself-consistencydecoding strategy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that the biggest efficiency gains in inference-time scaling come from concentrating compute where reasoning becomes uncertain, not from generating longer or more numerous full trajectories. It introduces DeepLook, a training-free decoding method that watches a model's own token-level confidence aggregated into segments, branches only when confidence drops relative to recent history, and evaluates candidate continuations with a fixed-horizon lookahead scored by Average Lookahead Confidence (ALC). The reported measurements show this selective mechanism matches or beats a confidence-filtered sampling baseline while using roughly one-eighth of the tokens, and sometimes beats a 512-sample self-consistency baseline. If true, this matters because it suggests reasoning models can get the benefits of test-time scaling at a fraction of the compute cost without any additional training.

Core claim

DeepLook claims that a monitor-and-intervene decoding strategy—segment-level confidence monitoring, triggered branching at local confidence drops, and ranking of fixed-horizon rollouts by Average Lookahead Confidence (ALC)—can shift the accuracy–token-cost Pareto frontier on competition mathematics. Across AIME24, AIME25, BRUMO25, and HMMT25 with DeepSeek-R1-8B, Qwen3-32B, GPT-OSS-20B, and GPT-OSS-120B, it reports higher accuracy than a confidence-filtered baseline (DeepConf-low) in 11 of 16 settings, ties once, and cuts dataset-level tokens by 87.3% on average, with per-setting gains like +3.1 on AIME25 with Qwen3-32B and +8.8 on BRUMO25 with GPT-OSS-20B. It also matches or exceeds a 512-sa

What carries the argument

Average Lookahead Confidence (ALC): after a confidence trigger, DeepLook samples K candidate branches, rolls each forward H segments, and scores each branch by the mean segment-level confidence over that rollout. Branches are ranked by ALC, pruned adaptively (keeping more branches when the confidence gap at the trigger is large), and the top branch resumes active monitoring while retained branches are completed greedily; final answers come from majority voting over completed branches. The ALC signal is what lets the method prefer continuations that regain confidence over ones that look plausible locally but stay uncertain, and the fixed-horizon lookahead is what distinguishes it from purely

Load-bearing premise

The load-bearing premise is that a branch's average self-confidence over a short lookahead rollout is a reliable indicator of whether that branch leads to a correct answer; if confidently wrong rollouts are common, ranking by ALC will systematically select the most confident errors and the mechanism fails, and the paper itself concedes that ALC ranks by distributional stability rather than correctness.

What would settle it

For a held-out set of trigger points from one benchmark, compute ALC for each of K sampled branches and record whether each branch reaches the correct answer. If the rank correlation between ALC and correctness is at or below zero, or if the ALC-selected branch is no more accurate than a randomly selected branch, the claimed mechanism is not selecting on quality.

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

If this is right

  • If the reported numbers hold, inference-time scaling does not require regenerating full trajectories: targeted lookahead at uncertainty bottlenecks captures most of the accuracy gains of brute-force sampling at roughly an order of magnitude lower token cost.
  • A single fixed trigger configuration (q=0.10, W=8, delta=0.02) works across four model families without per-model recalibration, which suggests the confidence-drop signal transfers across architectures.
  • Because the method is training-free and only needs white-box log-probability access, it can be dropped into existing open-weight reasoning models without finetuning.
  • The lookahead horizon H controls the trade-off: accuracy improves monotonically from H=1 to H=16, implying that short-horizon evaluation misses delayed reasoning errors and that longer lookahead is the main accuracy lever.
  • Component ablations indicate that neither confidence triggering alone nor voting alone reproduces the gains; both lookahead ranking and voting are needed, with lookahead ranking sometimes beating voting by itself.

Where Pith is reading between the lines

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

  • ALC is a self-confidence statistic, not a correctness verifier; a natural extension would be to replace or augment it with an externally trained verifier or a consistency check across branches, which might fix the cases where confident-but-wrong rollouts dominate.
  • The token-cost metric does not capture wall-clock latency or memory: running K lookahead branches in parallel raises KV-cache memory proportionally, so the practical efficiency gain may be smaller in latency-constrained or memory-constrained deployments.
  • The motivating observation—wrong traces show earlier and more frequent confidence drops—is measured on two model families; a testable extension is to check whether the same pattern holds on non-math reasoning tasks, and whether DeepLook-style triggering transfers to code or tool-use.
  • The method selects among branches generated from a shared prefix; an extension would be to use ALC not only for pruning but as a reward signal for training a lightweight router or policy that learns when to invoke lookahead.

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

3 major / 5 minor

Summary. DeepLook is a training-free monitor-and-intervene decoding framework. It converts token-level top-k log-probabilities into segment-level confidences (Eqs. 1–2), triggers exploration when a segment falls below a local percentile threshold (Eqs. 3–4), and then samples K branches at the trigger point. Each branch is rolled out for a fixed horizon H and scored by Average Lookahead Confidence (ALC, Eq. 5). Branches are pruned adaptively (Eq. 6), the highest-ALC branch resumes monitored decoding, and the final answer is a majority vote over completed retained traces (Eq. 7). On AIME24/25, BRUMO25, and HMMT25 across DeepSeek-R1-8B, Qwen3-32B, GPT-OSS-20B/120B, the paper reports that DeepLook@128 improves accuracy over DeepConf-low in 11 of 16 settings while reducing dataset-level token generation by 87.3% on average. The paper includes component ablations, horizon scaling, hyperparameter sensitivity, and a token-cost breakdown.

Significance. If the empirical claims are reproducible, the paper contributes a practical, training-free compute-allocation strategy with transparent token accounting. The pipeline is clearly specified, the headline token arithmetic is internally consistent (I verified the 87.3% average from Table 1), and the appendix provides useful sensitivity analyses. The main weakness is that the accuracy component of the claim is not yet statistically supported, and the causal mechanism attributed to ALC is explicitly conceded by the authors to rank distributional stability rather than correctness. The contribution would be strengthened substantially by repeated-seed results with confidence intervals and a direct per-branch validation of ALC as a correctness signal.

major comments (3)
  1. [§4.2, Table 1] The headline accuracy claim rests on single-run point estimates. The manuscript never states the number of problems per benchmark, and no seeds, standard errors, or significance tests are reported. On standard AIME/HMMT sets of 30 problems, a 3.3-point delta is one answer; several cells in Table 1 are sub-answer differences (e.g., AIME25 DeepSeek-R1-8B: +0.3; HMMT25 Qwen3-32B: −1.12; HMMT25 GPT-OSS-120B: +1.3). The aggregate '11 of 16 settings' is therefore within sampling noise. Please report dataset sizes, run multiple seeds (or explicitly state determinism), and provide paired bootstrap confidence intervals or McNemar tests for each accuracy delta.
  2. [§3.2–3.3, Eq. (5), Limitations] ALC is the load-bearing selection signal: Eq. (5) ranks branches, and §3.3 uses ALC both to pick the primary branch and to prune the retained set. The Limitations paragraph states 'ALC ranks by distributional stability rather than correctness.' Figure 7a shows the full method outperforms components on one benchmark, but it does not establish that high-ALC branches are correct more often than low-ALC branches; Appendix E.2 validates the trigger (n_unc), not ALC. Add a per-branch analysis at triggered segments comparing ALC distributions for eventually-correct vs eventually-wrong rollouts, or a controlled ablation replacing ALC with local confidence/random ranking. Without this, the claim that lookahead confidence is the cause of the gains is unsupported.
  3. [Abstract, §4.2, Fig. 5] The 'Pareto frontier' wording overstates what the data show. A Pareto frontier is a set of non-dominated operating points; Table 1 provides only one DeepLook point per setting, and in 4 of 16 settings DeepLook is less accurate than DeepConf-low (e.g., HMMT25 DeepSeek-R1-8B: 73.3% vs 77.6%). To support a frontier shift, the paper should present cost-accuracy curves (multiple budgets for DeepLook and for the baselines) and identify the non-dominated points per setting. Otherwise the claim should be softened to 'comparable or better accuracy at much lower cost in most settings.'
minor comments (5)
  1. [§3.1 and Algorithm 1] The symbol H is used both for the history window (Eq. 3, Algorithm line 7/29) and for the lookahead horizon (Eq. 5, Algorithm line 14). Use different symbols to avoid confusion.
  2. [Fig. 3] The confidence-to-color mapping is not defined in the caption. The 'Low (4.24) Mid (13.78) High (31.93)' scale is unclear in relation to the segment backgrounds.
  3. [§4.2] The sentence 'ties once' is ambiguous; it should specify the baseline (DeepConf-low), since Table 1 also shows ties against Cons@512.
  4. [Appendix E.3, Table 8] The text describes a one-at-a-time change from q=0.10 to q=0.05 as 'smooth degradation,' but a 3.3-point drop on a small benchmark may be within noise. Without error bars, 'smooth' is not supported.
  5. [§4.1] The paper does not specify how DeepConf-low's efficiency operating point is chosen or how its token cost is measured. This is needed to interpret the primary efficiency comparison.

Circularity Check

0 steps flagged

No circularity: the headline accuracy–token-cost result is measured against external ground-truth labels, and the paper explicitly disclaims ALC as a correctness verifier.

full rationale

The central claim is an empirical accuracy–token Pareto comparison (Table 1) evaluated on external benchmarks (AIME24, AIME25, BRUMO25, HMMT25) against ground-truth answers, with token costs counted from actual decoding, rollouts, and completions. No equation in the paper reduces the reported accuracy to the method's internal confidence inputs. ALC (Eq. 5) is defined as mean rollout confidence and is used as a ranking/trigger signal, but the paper explicitly does not claim that ALC predicts correctness: §3.2 says it is 'an internal stability heuristic' that is not 'an external correctness verifier,' and the Limitations paragraph concedes that 'ALC ranks by distributional stability rather than correctness.' That is a validity limitation on the causal interpretation, not a circular reduction: the paper's own experiments, not the definition of ALC, establish whether high-ALC selection improves accuracy. The comparison to DeepConf-low and Cons@512 is also external; DeepConf's confidence definition (Eq. 1) is adopted from an external reference (Fu et al. 2025), and the non-myopic motivation (Ma et al. 2024) is independently cited rather than self-imported. Hyperparameters (q, W, δ, H, K) were selected with knowledge of the benchmarks, which is a potential overfitting/selection concern but not a by-construction equivalence. No load-bearing self-citations, no imported uniqueness theorem, and no fitted parameter is renamed as a prediction. Therefore no specific circular step can be exhibited from the paper's own equations or citation chain, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 1 invented entities

The central claim rests on a small set of hand-set hyperparameters (with sensitivity analyzed on the same benchmarks that later appear as wins), on three empirical premises about confidence predicting errors and future confidence predicting correctness, and on one invented ranking quantity (ALC) whose validity is only tested in-house. No physical entities are postulated; the free-parameter count is dominated by trigger/lookahead choices rather than fitted physical quantities.

free parameters (8)
  • Lookahead horizon H = 16
    Default H=16 gives the best accuracy in Fig 7b (73.3→83.3 from H=1 to 16); not swept beyond 16. Controls how many future segments are used to rank branches.
  • Trigger quantile q = 0.10
    Chosen from one-at-a-time sensitivity on AIME25 DeepSeek-R1-8B (App E.3); 0.05 misses errors, 0.25 over-triggers, 0.10 is the ablation baseline.
  • History window W = 8
    App E.3: W=32 under-triggers, W=4 underperforms; 8 is the ablation baseline. Chosen on the same benchmark family reported as main results.
  • Hysteresis margin delta = 0.02
    App E.3: δ=0 fires 6.8 triggers/problem vs 5.5 at 0.02 with no accuracy gain; 0.02 chosen from the same ablation set.
  • Keep ratio bounds rho_min/rho_max = 0.10/0.25
    Table 9 shows fixed ρ=0.50 gives +3.3 over ρ=0.25 on Qwen3-32B/AIME25; the paper caps dynamic retention at 0.25 to bound completion cost, conceding an accuracy–cost trade.
  • Branch width K (budget B=K×R_max) = K=64 for DeepLook@128
    The headline results use the maximum tested budget (@128); Table 4 shows @32 and @64 give lower accuracy, so the reported Pareto point is the largest budget evaluated.
  • Pruning sensitivity s and base retention rho_base = unspecified
    Eq. (6) defines the keep ratio using s and ρ_base, but no default values are given in Table 2; a load-bearing parameter pair is left unset.
  • Segment lengths L_main and L_look = 512 / 32
    Chosen constants governing confidence aggregation granularity and the trigger firing rate per problem; no sensitivity analysis reported for either.
axioms (5)
  • domain assumption Token-level top-k negative log-likelihood (Eq. 1) tracks reasoning correctness better than chance.
    The entire trigger machinery reduces to this correlation. The paper cites Kadavath et al. and Kuhn et al. for calibration, but the specific segment-level aggregation is not validated against external correctness.
  • domain assumption Lookahead segment confidence (ALC, Eq. 5) is predictive of branch quality: high future confidence implies a better final answer.
    Primary-branch selection and pruning (§3.3, Eq. 6) both depend on ALC. The Limitations paragraph concedes it ranks “distributional stability rather than correctness,” making this an empirical premise the paper does not establish.
  • domain assumption Correct and incorrect traces differ early in uncertainty onset (Fig. 2, Fig. 10), and ≤2 interventions per trace capture the decisive errors.
    The adaptive computation is only useful if reasoning errors concentrate at a few segment positions. Fig. 2 is a descriptive density comparison on the evaluation traces, not a held-out predictive test.
  • domain assumption Majority voting over greedy completions of retained branches improves over single-path decoding (Eq. 7).
    The ablation shows voting helps on DeepSeek-R1-8B (63.96 vs 61.85 single-path) but hurts on Qwen3-32B (56.58 vs 60.00 single-path), so the axiom is model-dependent.
  • standard math Benchmark ground-truth answers from MathArena/Hugging Face datasets are correct as distributed.
    Standard evaluation assumption; no label audit of AIME24/25, BRUMO25, or HMMT25 is reported.
invented entities (1)
  • Average Lookahead Confidence (ALC) no independent evidence
    purpose: Ranks candidate continuations by mean segment confidence over a fixed-horizon rollout (Eq. 5); selects the primary branch and prunes the rest.
    A new scoring quantity defined entirely inside the paper. Its only empirical validation is the paper's own ablations on the same benchmarks used for the headline results; no external benchmark or released artifact supports it.

pith-pipeline@v1.3.0-alltime-deepseek · 19261 in / 26454 out tokens · 266862 ms · 2026-08-02T11:26:01.691458+00:00 · methodology

0 comments
read the original abstract

Inference-time scaling has emerged as a powerful paradigm for improving large language model reasoning, often delivering larger gains on difficult reasoning tasks than parameter scaling alone. However, existing approaches remain inefficient in how compute is allocated within a reasoning trace. Motivated by the observation that reasoning failures often exhibit an early onset of uncertainty before a wrong answer become explicit, we introduce DeepLook, a training-free monitor-and-intervene decoding framework that concentrates lookahead compute at uncertainty bottlenecks. DeepLook aggregates token-level confidence into segment-level signals, triggers when confidence drops relative to recent history, and explores candidate continuations with fixed-horizon lookahead. Branches are ranked by Average Lookahead Confidence (ALC), the average segment-level confidence over rollout continuations, then pruned and aggregated through voting. On four competition-style mathematics benchmarks across DeepSeek-R1-8B, Qwen3-32B, GPT-OSS-20B, and GPT-OSS-120B, DeepLook shifts the accuracy--token-cost Pareto frontier: it improves accuracy over DeepConf-low in 11 of 16 settings while reducing dataset-level token generation by 87.3% on average, including gains of +3.1 on AIME25 with Qwen3-32B and +8.8 on BRUMO25 with GPT-OSS-20B. These results show that selective, future-aware intervention yields substantially stronger accuracy--cost trade-offs than uniformly scaling complete reasoning trajectories. Code is available here.

Figures

Figures reproduced from arXiv: 2607.22602 by Mengyue Wang, Tingxin Yang, Xingcheng Zhou, Yunpu Ma, Zefeng Wang.

Figure 2
Figure 2. Figure 2: Uncertainty diagnostics correlate with errors on DeepSeek-R1-8B. We analyze the density of (a) the number of uncertain segments and (b) the position of the first uncertain segment. Wrong traces (orange) tend to exhibit more frequent uncertainty and, crucially, an earlier uncertainty onset compared to correct traces (blue). This separation motivates our strategy to use early confidence drops as triggers for… view at source ↗
Figure 3
Figure 3. Figure 3: Token- vs. segment-level confidence. Each segment spans two lines for readability (L=32). The segment background encodes segment-level confidence C¯, where red backgrounds indicate uncertain segments. Individual token colors (Ct) reveal local confidence fluctuations, motivating segment-level aggregation for robust detection of reasoning failures. quality [Kadavath et al., 2022, Kuhn et al., 2023]. Yet conf… view at source ↗
Figure 4
Figure 4. Figure 4: DeepLook pipeline. (1) The reasoning trace is gathered into segments, and token-level top￾k log-probability statistics are aggregated into a segment confidence score C¯(Si). (2) A local history window Hi defines a percentile threshold τi ; exploration is triggered only when the current segment falls below this local baseline. (3) From the trigger segment, DeepLook samples K lookahead branches and rolls eac… view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy vs. token cost trade-offs across benchmarks. Each panel plots dataset-level token cost against accuracy for one benchmark. Colors identify the base model, with DeepLook@128 shown as prominent stars and the corresponding Path@1, DeepConf-low, and Cons@512 references shown with lighter markers. Faint colored segments connect Path@1 to DeepConf-low within the same model, visualizing the baseline accu… view at source ↗
Figure 6
Figure 6. Figure 6: Overcoming reasoning hallucination. DeepLook utilizes lookahead horizon H to filter deceptive paths plausible locally but leading to collapse. Adaptive pruning scales coverage with uncertainty ∆i . branches according to their average confidence over lookahead rollout segments, DeepLook can reject trajectories that look locally plausible but remain uncertain later. In this sense, the benefit comes not from … view at source ↗
Figure 7
Figure 7. Figure 7: Ablation and scaling analysis. (a) Performance improves incrementally as components are integrated; ∆ denotes gain over the Greedy baseline. (b) Accuracy scales robustly with the lookahead horizon H with high compute efficiency. roles: lookahead refines the candidate set by pruning low-ALC branches, while voting resolves residual ambiguities among the remaining high-quality trajectories. Impact of confiden… view at source ↗
Figure 8
Figure 8. Figure 8: Prompt logic for model_type=deepseek/qwen. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Prompt logic for GPT models with reasoning_effort. A.4 Existing asset licenses [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Motivating analysis on Qwen3-32B. Incorrect traces show both more frequent uncertainty and earlier uncertainty onset than correct traces [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Avg. nunc for correct vs. incorrect traces. DeepSeek-R1-8B on HMMT25 [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: DeepLook exposes global constraints that self-consistency misses. In this HMMT combinatorics example, Cons@512 repeatedly selects the locally plausible odd-vertex heuristic and converges to 202. By intervening at the unstable pairing step, DeepLook uses lookahead to test whether the proposed edge removals remain globally consistent, detects the boundary-parity obstruction, and redirects the solution to th… view at source ↗
Figure 13
Figure 13. Figure 13: DeepLook concentrates compute on fragile enumeration steps. In this AIME 2024 counting example, Cons@512 preserves a mechanically plausible inclusion-exclusion error and over￾counts the rotational cases, giving 373. DeepLook identifies the low-confidence overlap calculation as the decision point, uses lookahead rollouts to reveal duplicated symmetry classes, and prunes the overcounted branch to recover th… view at source ↗
Figure 14
Figure 14. Figure 14: DeepLook rejects deceptive algebraic branches through lookahead verification. In this continuous-math example, Cons@512 treats the result of repeated squaring as reliable and outputs an extraneous root. DeepLook instead evaluates the downstream stability of that branch, checks the candidate against the original equation and domain behavior, and switches to the globally valid conclusion that no real soluti… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

37 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    2025 , eprint =

    Deep Think with Confidence , author =. 2025 , eprint =. doi:10.48550/arXiv.2508.15260 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.15260 2025

  2. [2]

    2026 , eprint =

    DeepPrune: Parallel Scaling without Inter-trace Redundancy , author =. 2026 , eprint =. doi:10.48550/arXiv.2510.08483 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2510.08483 2026

  3. [3]

    2024 , eprint =

    Non-myopic Generation of Language Models for Reasoning and Planning , author =. 2024 , eprint =. doi:10.48550/arXiv.2410.17195 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.17195 2024

  4. [4]

    $\phi$-Decoding: Adaptive Foresight Sampling for Balanced Inference-Time Exploration and Exploitation

    Xu, Fangzhi and Yan, Hang and Ma, Chang and Zhao, Haiteng and Liu, Jun and Lin, Qika and Wu, Zhiyong , year =. doi:10.48550/arXiv.2503.13288 , url =. 2503.13288 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.13288

  5. [5]

    2025 , eprint =

    Scalable Best-of-N Selection for Large Language Models via Self-Certainty , author =. 2025 , eprint =. doi:10.48550/arXiv.2502.18581 , url =

    doi:10.48550/arxiv.2502.18581 2025

  6. [6]

    2026 , eprint =

    Lookahead Tree-Based Rollouts for Enhanced Trajectory-Level Exploration in Reinforcement Learning with Verifiable Rewards , author =. 2026 , eprint =. doi:10.48550/arXiv.2510.24302 , note =

    doi:10.48550/arxiv.2510.24302 2026

  7. [7]

    2026 , eprint =

    MITS: Enhanced Tree Search Reasoning for LLMs via Pointwise Mutual Information , author =. 2026 , eprint =. doi:10.48550/arXiv.2510.03632 , url =

    doi:10.48550/arxiv.2510.03632 2026

  8. [8]

    2025 , eprint =

    Qwen3 Technical Report , author =. 2025 , eprint =. doi:10.48550/arXiv.2505.09388 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.09388 2025

  9. [9]

    doi:10.1038/s41586-025-09422-z , url =

    2025 , journal =. doi:10.1038/s41586-025-09422-z , url =. 2501.12948 , archivePrefix =

    Pith/arXiv arXiv 2025

  10. [10]

    2025 , eprint =

    gpt-oss-120b & gpt-oss-20b Model Card , author =. 2025 , eprint =. doi:10.48550/arXiv.2508.10925 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.10925 2025

  11. [11]

    Are More

    Chen, Lingjiao and Davis, Jared Quincy and Hanin, Boris and Bailis, Peter and Stoica, Ion and Zaharia, Matei and Zou, James , year =. Are More. 2403.02419 , archivePrefix =

    Pith/arXiv arXiv

  12. [12]

    2024 , eprint =

    Large Language Monkeys: Scaling Inference Compute with Repeated Sampling , author =. 2024 , eprint =

    2024

  13. [13]

    2025 , eprint =

    Efficient Reasoning Models: A Survey , author =. 2025 , eprint =

    2025

  14. [14]

    2025 , eprint =

    Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models , author =. 2025 , eprint =

    2025

  15. [15]

    2025 , eprint =

    Efficient Inference for Large Reasoning Models: A Survey , author =. 2025 , eprint =

    2025

  16. [16]

    2025 , eprint =

    A Survey of Efficient Reasoning for Large Reasoning Models: Language, Multimodality, and Beyond , author =. 2025 , eprint =

    2025

  17. [17]

    2025 , eprint =

    L1: Controlling How Long a Reasoning Model Thinks with Reinforcement Learning , author =. 2025 , eprint =

    2025

  18. [18]

    ThinkPrune: Pruning Long Chain-of-Thought of

    Hou, Bairu and Zhang, Yang and Ji, Jiabao and Liu, Yujian and Qian, Kaizhi and Andreas, Jacob and Chang, Shiyu , year =. ThinkPrune: Pruning Long Chain-of-Thought of. 2504.01296 , archivePrefix =

    Pith/arXiv arXiv

  19. [19]

    TokenSkip: Controllable Chain-of-Thought Compression in

    Xia, Heming and Leong, Chak Tou and Wang, Wenjie and Li, Yongqi and Li, Wenjie , year =. TokenSkip: Controllable Chain-of-Thought Compression in. 2502.12067 , archivePrefix =

    arXiv

  20. [20]

    2025 , eprint =

    Chain of Draft: Thinking Faster by Writing Less , author =. 2025 , eprint =

    2025

  21. [21]

    2025 , eprint =

    Dynamic Early Exit in Reasoning Models , author =. 2025 , eprint =

    2025

  22. [22]

    2025 , eprint =

    AdaptThink: Reasoning Models Can Learn When to Think , author =. 2025 , eprint =

    2025

  23. [23]

    2022 , eprint =

    Large Language Models are Zero-Shot Reasoners , author =. 2022 , eprint =. doi:10.48550/arXiv.2205.11916 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2205.11916 2022

  24. [24]

    2023 , journal =

    ReAct: Synergizing Reasoning and Acting in Language Models , author =. 2023 , journal =. doi:10.48550/arXiv.2210.03629 , url =. 2210.03629 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2210.03629 2023

  25. [25]

    2023 , journal =

    Self-Refine: Iterative Refinement with Self-Feedback , author =. 2023 , journal =. doi:10.48550/arXiv.2303.17651 , url =. 2303.17651 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.17651 2023

  26. [26]

    2022 , journal =

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. 2022 , journal =. 2201.11903 , archivePrefix =

    Pith/arXiv arXiv 2022

  27. [27]

    2023 , journal =

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author =. 2023 , journal =. 2203.11171 , archivePrefix =

    Pith/arXiv arXiv 2023

  28. [28]

    2023 , journal =

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author =. 2023 , journal =. 2305.10601 , archivePrefix =

    Pith/arXiv arXiv 2023

  29. [29]

    Snell, Charlie Victor and Lee, Jaehoon and Xu, Kelvin and Kumar, Aviral , year =. Scaling. The Thirteenth International Conference on Learning Representations , eprint =. doi:10.48550/arXiv.2408.03314 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2408.03314

  30. [30]

    2024 , journal =

    Let's Verify Step by Step , author =. 2024 , journal =. 2305.20050 , archivePrefix =

    Pith/arXiv arXiv 2024

  31. [31]

    2022 , journal =

    Language Models (Mostly) Know What They Know , author =. 2022 , journal =. 2207.05221 , archivePrefix =

    Pith/arXiv arXiv 2022

  32. [32]

    2023 , journal =

    Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation , author =. 2023 , journal =. 2302.09664 , archivePrefix =

    Pith/arXiv arXiv 2023

  33. [33]

    2023 , journal =

    Reasoning with Language Model is Planning with World Model , author =. 2023 , journal =. 2305.14992 , archivePrefix =

    Pith/arXiv arXiv 2023

  34. [34]

    2024 , booktitle =

    Graph of Thoughts: Solving Elaborate Problems with Large Language Models , author =. 2024 , booktitle =. 2308.09687 , archivePrefix =

    Pith/arXiv arXiv 2024

  35. [35]

    2024 , booktitle =

    Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models , author =. 2024 , booktitle =. 2310.04406 , archivePrefix =

    Pith/arXiv arXiv 2024

  36. [36]

    2024 , journal =

    Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking , author =. 2024 , journal =. 2403.09629 , archivePrefix =

    Pith/arXiv arXiv 2024

  37. [37]

    2024 , booktitle =

    The Pitfalls of Next-Token Prediction , author =. 2024 , booktitle =. 2403.06963 , archivePrefix =

    Pith/arXiv arXiv 2024