Pith. sign in

REVIEW 3 major objections 6 minor 38 references

In-context search turns exponentially rare correct solutions into high-probability success after only polynomially many sequential attempts, if reflections catch early mistakes.

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 · grok-4.5

2026-07-10 22:44 UTC pith:JEQVT3VI

load-bearing objection Clean poly-vs-exp sampling-complexity theory for in-context search that isolates early error localization as the make-or-break condition; proofs are careful, empirics only qualitative. the 3 major comments →

arxiv 2607.06720 v1 pith:JEQVT3VI submitted 2026-07-07 cs.AI cs.CL

When Does In-Context Search Help? A Sampling-Complexity Theory of Reflection-Driven Reasoning

Yotam Wolf , Noam Wies , Amnon Shashua This is my paper
classification cs.AI cs.CL
keywords in-context searchsampling complexityself-reflectionreasoning tracesposterior reweightinglarge reasoning modelsRLVRearly error localization
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.

Modern reasoning models keep a full history of failed attempts and reflections in context, then try again. This paper asks when that sequential process is actually better than just sampling many independent answers from the base model. The answer is governed by reflection quality: if the model can reliably name the earliest wrong step after seeing the same failure a few times, each update prunes a whole subtree of continuations, and the number of needed attempts scales only polynomially with reasoning depth and branching. When reflections only flag late mistakes, or when positive feedback is used instead, the first-step distribution never improves and sequential conditioning gives no asymptotic gain. The same posterior reweighting rule is shown to be learnable from search rollouts with polynomial sample complexity and to arise as the optimal stage-wise policy extension under a noisy-reflection model of reinforcement learning with verifiable rewards. Qualitative checks on synthetic error-injection traces and on real large reasoning models on AIME 2025 are consistent with progressive redistribution of probability mass rather than monotone accumulation of progress.

Core claim

When reflections localize the earliest incorrect prefix with non-negligible probability after a few exposures, in-context search under the paper's logit-reweighting rule solves problems whose base-model pass rate decays exponentially with depth using only a polynomial number of sequential attempts; when that early-localization property fails, conditioning on past attempts offers no asymptotic advantage over parallel sampling.

What carries the argument

Adaptive early reflection plus in-context logit reweighting: after an incorrect trace, reflection returns a transition; the model subtracts a fixed rate times the number of times that transition has been flagged from its base logit, thereby exponentially suppressing the whole subtree of continuations that pass through it.

Load-bearing premise

The model must, after seeing the same earliest wrong prefix a few times, correctly name that earliest step with non-negligible probability; if it only ever flags late errors, the exponential-to-polynomial gain disappears.

What would settle it

On hard problems of increasing depth, measure whether a real large reasoning model returns the true earliest incorrect step with probability bounded away from zero after a fixed number of repeated exposures of that prefix; if that success probability stays near zero or only late steps are ever flagged, the polynomial-round guarantee does not apply.

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

If this is right

  • Problems whose zero-shot pass rate is exponentially small in depth become solvable with high probability after only polynomially many sequential attempts under early reflection.
  • Approximate step-wise imitation of efficient search transcripts is enough; standard cross-entropy training recovers the same pass-rate guarantee with polynomial sample complexity.
  • Under a stage-wise model of RL with verifiable rewards, the optimal policy extension implements exactly the same exponential reweighting rule, giving a mechanism for search-like reasoning to emerge from lengthening CoTs.
  • Late or positive-only feedback can leave the first-step distribution unchanged or even lower overall success probability, so reflection design should prioritize early negative localization over reinforcement of believed-correct steps.

Where Pith is reading between the lines

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

  • If reflection quality itself improves with model scale or with deliberate self-awareness training, the theory predicts a sharp transition from exponential to polynomial inference-time cost rather than a gradual gain.
  • Reasoning loops that restart from already-invalidated prefixes are predicted to be a dominant practical failure mode; measuring loop rate may be a better diagnostic of search efficiency than raw pass rate alone.
  • The same early-localization criterion could be used as an auxiliary training objective or as a filter when selecting which reflections to keep in context, potentially amplifying the polynomial regime.

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 / 6 minor

Summary. The paper models in-context search in large reasoning models as approximate sequential inference over reasoning traces: a base model supplies a prior, self-reflection supplies noisy localization of errors, and conditioning implements posterior reweighting of transitions (Definition 3.4). Under γ-local support, adaptive early reflection (Definition 3.2), and contextual reachability (Assumption 3.5), Theorem 4.3 / C.1 shows that problems whose zero-shot pass rate can be exponentially small in depth n are solved with high probability after only a polynomial number of sequential attempts, Õ(n W m_r + (n W/(p_r η)) log(1/γ) + p_r^{-2} log(1/δ)). Propositions 4.1–4.2 identify complementary failure modes (late localization; positive feedback). The authors further show robustness of approximate step-wise updates (Proposition 5.1), polynomial sample complexity of cross-entropy imitation of efficient rollouts, and that the same reweighting arises as the optimal stagewise policy extension under a noisy-reflection model of RLVR (Theorem 5.2). Appendices extend the main bound to multiple solutions, bounded false positives, and bounded delay. Qualitative experiments on synthetic traces and AIME 2025 trajectories with DeepSeek-R1 distillations are used to check reflection delay and prefix-conditioned pass-rate dynamics.

Significance. If the stated conditions hold, the work gives a clean sampling-complexity separation between parallel sampling from a base model and sequential reflection-driven search, and supplies a mechanistic account of when and why in-context search can be exponentially more efficient. The proofs are detailed (suppression lemmas, martingale concentration, self-consistent horizon arguments), and the extensions to false positives, delay, learnability, and stagewise RLVR optimality are genuine technical contributions rather than corollaries. The paper is also careful about negative results (late reflection, positive feedback) and includes an honest limitations appendix. These strengths make the manuscript a useful theoretical reference for the growing literature on test-time search and RLVR-trained reasoning models, even though the empirical section is only qualitative.

major comments (3)
  1. [Definition 3.2, Theorem 4.3, §6.1 / Figure 2] Definition 3.2 (adaptive early reflection) is load-bearing for the poly-vs-exp separation of Theorem 4.3 / C.1: Proposition 4.1 shows that if localization is confined to depths > n/2, the first-step distribution never improves and the exponential gap disappears. The only direct empirical check is Figure 2 on synthetic arithmetic/equation traces with a single injected error, which shows non-trivial mass at zero delay but does not estimate (p_r, m_r) on multi-error, multi-path AIME-style problems under the same model that generates the CoTs. The manuscript should either (i) add a quantitative measurement of earliest-error hit rate after repeated exposures on hard traces, or (ii) more sharply separate the conditional theoretical claim from any implication that current LRMs operate in the efficient regime.
  2. [Assumption 3.5, Appendix J, Theorem 4.3] Assumption 3.5 (contextual reachability) rules out restarting from already-invalidated prefixes. Appendix J reports that a large fraction of failed AIME trajectories exhibit reasoning loops (e.g., 0.83/0.48 for the 1.5B model under greedy/nucleus), so the assumption is frequently violated in practice. The poly bound of Theorem 4.3 is stated only under this assumption; the paper should quantify or bound how partial violations degrade T (e.g., via a loop-budget parameter), or explicitly mark the main theorem as characterizing an idealized non-looping regime and discuss efficiency under observed loop rates.
  3. [Definition 3.4, Proposition 5.1, Theorem 5.2] Definition 3.4 implements posterior updates as explicit logit subtraction ℓ ← ℓ_0 − η n(h,a;C). Real LRM search is free-form conditioning on full history, not a known multiplicative downweight of a named transition. Proposition 5.1 shows that approximate next-step imitation of an efficient transcript distribution preserves pass rate, and Theorem 5.2 derives the same form as optimal under a stylized noisy-reflection likelihood, but neither result formally justifies that transformer conditioning on natural-language reflections approximates the specific local reweighting used in the sampling-complexity analysis. A short discussion of this modeling gap—and of what would falsify the approximation—would strengthen the bridge from theory to practice.
minor comments (6)
  1. [Figure 1] Figure 1 is helpful but the caption and panel (b) are sparse; labeling the reflected node and the suppressed subtree would make the heuristic clearer for readers who skip the formal definitions.
  2. [§3.1] In §3.1 the pass-rate bound is written p^n = O(exp(-n)); the equality is only asymptotic order, not identity. Prefer ≤ p^n = O(exp(-n)).
  3. [Theorem 4.3 / Appendix C] Theorem 4.3 and the multi-solution version C.1 both hide log factors in Õ; a short table of leading dependencies on n, W, k, p_r, m_r, η, γ, δ would help practitioners.
  4. [§6, Appendix I.1] Appendix I.1 (synthetic trees) is a useful controlled check of the theory but is only briefly referenced from the main text; a one-sentence pointer in §6 would improve discoverability.
  5. [Throughout] Typos / style: “Despitethesesuccesses” (p.1), “F ailure Modes” / “Rew ards” (section headers with stray spaces), and occasional missing spaces after periods in the related-work block.
  6. [Figure 3(c), Appendix M] Figure 3(c) uses a truncated log-ratio with +0.01 floor (Appendix M); state this in the main-text caption so readers do not over-interpret the left tail.

Circularity Check

1 steps flagged

No significant circularity: sampling-complexity theorems and RLVR optimality are derived from explicit generative/reflection models and Bayes, not forced by fits or self-citation chains.

specific steps
  1. self citation load bearing [Sec. 1 / Related Work / Sec. 3 intro]
    "Building on the search-based perspective of Shalev-Shwartz and Shashua (2025), we model reasoning as inference over reasoning traces... In contrast, modern LRMs typically retain the full history of attempts in-context, raising the question of when this unstructured form can match the same guarantees."

    Overlapping author (Shashua) prior work is cited as the source of the search-theoretic framing and tree representation. This is background motivation only; the paper’s own theorems (4.3, 5.2, etc.) are proved from freshly stated assumptions and update rules without importing uniqueness or numerical claims from that paper. Not load-bearing for the poly-vs-exp result.

full rationale

The central claims (Theorem 4.3/C.1 poly-vs-exp separation under early reflection; Props 4.1–4.2 failure modes; Prop 5.1 transfer via KL/TV; Theorem 5.2 optimal stage extension) are mathematical consequences of the paper’s own definitions (γ-local support, adaptive reflection Def. 3.2, logit reweighting Def. 3.4, noisy-reflection likelihood) and standard concentration/Bayes arguments, fully proved in the appendices. They do not reduce by construction to fitted constants or to a restatement of the target. The sole self-citation of note is to Shalev-Shwartz & Shashua (2025) for the explicit tree-search motivation and prefix-tree notation; it supplies background, not a uniqueness theorem or load-bearing premise that forces the in-context results. Empirical sections are qualitative checks of assumptions, not predictions obtained by fitting the same quantities. Score 1 only for the minor background self-citation; the derivation chain itself is self-contained and non-circular.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The central poly-vs-exp claim rests on a small set of modeling axioms (local support, adaptive early reflection, contextual reachability, multiplicative logit reweighting) plus free parameters that quantify reflection quality and update strength. No data-fitted constants enter the theorems; the parameters are left symbolic. The invented entities are the idealized reflection oracle and the in-context reweighting rule themselves.

free parameters (6)
  • p_r (reflection success probability)
    Lower-bounds the probability that reflection returns the earliest incorrect step once the failure mode has been seen m_r times; appears in the leading 1/p_r and 1/p_r² terms of T.
  • m_r (exposure threshold)
    Number of prior failed attempts with the same earliest incorrect prefix required before the p_r guarantee applies; multiplies the knW term.
  • η (logit update rate)
    Multiplicative down-weight strength per reflection; appears as 1/η in the suppression count L.
  • γ (local support)
    Minimum probability of a correct next step from any correct prefix under the base model; enters only logarithmically.
  • Δ_r (reflection delay)
    In the delayed-reflection extension, the depth window of useful reflections; produces the exponential factor B_Δr = Σ W^j.
  • d (false-positive budget)
    Maximum number of times any correct transition may be incorrectly flagged; adds an additive d inside the log term.
axioms (5)
  • domain assumption γ-local support (Assumption 3.1): every correct prefix has next-step probability ≥ γ of staying on a correct path.
    Ensures correct solutions remain reachable; without it inference cannot recover them. Invoked throughout the suppression lemmas.
  • ad hoc to paper Adaptive reflection (Definition 3.2 / Assumption H.1 / I.1): after m_r exposures, earliest incorrect step is returned with probability ≥ p_r (or within delay Δ_r).
    The paper’s central modeling choice that turns exponential search into polynomial; not derived from first principles of LLMs.
  • ad hoc to paper Contextual reachability (Assumption 3.5): the model never restarts from a prefix already marked invalid.
    Rules out reasoning loops; Appendix J shows real models frequently violate it, so the efficient regime is idealized.
  • ad hoc to paper In-context reweighting (Definition 3.4): each reflection subtracts η from the logit of the identified transition.
    Idealized multiplicative update; Proposition 5.1 shows approximate step-wise versions suffice, but the exact form is postulated.
  • domain assumption Noisy reflection model for RLVR (Appendix D): reflections mark the first incorrect step with probability p_r > 1/2, conditionally independent given y*.
    Used only for the optimality derivation (Theorem 5.2); abstracts the feedback available under verifiable rewards.
invented entities (2)
  • Adaptive reflection signal Z_t no independent evidence
    purpose: Supplies the information that allows posterior mass to be removed from whole subtrees rather than single leaves.
    Defined in Definition 3.2; independent evidence is only the qualitative delay histograms on synthetic arithmetic traces, not a measured (p_r, m_r) on hard problems.
  • In-context logit reweighting rule ℓ ← ℓ_0 - η n(h,a;C) no independent evidence
    purpose: Implements the posterior update that concentrates probability on viable prefixes.
    Definition 3.4; shown to be optimal under the noisy-reflection likelihood, but the exact multiplicative form is an idealization of what conditioning on history actually does.

pith-pipeline@v1.1.0-grok45 · 31396 in / 3682 out tokens · 51443 ms · 2026-07-10T22:44:49.297198+00:00 · methodology

0 comments
read the original abstract

Training large language models (LLMs) with extended reasoning has enabled in-context search, in which models iteratively generate, critique, and revise solution attempts. We provide a theoretical analysis of in-context search by modeling it as approximate inference over reasoning traces, where the base model defines a prior and self-reflection provides feedback for posterior updates, and study the resulting inference-time sampling complexity - the number of sequential attempts needed to achieve high success probability. We show that when reflections reliably localize early mistakes, in-context search can yield exponential improvements over the base model, solving problems with exponentially small zero-shot pass rates using only a polynomial number of sequential attempts, whereas when this property fails, conditioning on past attempts offers no asymptotic benefit over parallel sampling. We further show that these gains are robust and learnable: approximate posterior updates suffice, and cross-entropy training on search rollouts recovers the required behavior with polynomial sample complexity. Finally, we show that under a stagewise abstraction of reinforcement learning with verifiable rewards, the optimal policy extension implements the same posterior reweighting rule. We validate key qualitative predictions of the theory on real large reasoning models.

Figures

Figures reproduced from arXiv: 2607.06720 by Amnon Shashua, Noam Wies, Yotam Wolf.

Figure 1
Figure 1. Figure 1: (a) Visualization of search heuristic. On the first attempt the model samples a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Reflection delay distributions for arithmetic and equation-solving tasks. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Evolution of pass rate conditioned on CoT prefixes along reasoning trajectories [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Comparison of sample efficiency to find the correct path in trees of increasing [PITH_FULL_IMAGE:figures/full_fig_p034_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Pass-rate evolution under explicit sequential self-reflection rounds on AIME [PITH_FULL_IMAGE:figures/full_fig_p035_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Evolution of prefix-conditioned pass rate along reasoning trajectories on AIME [PITH_FULL_IMAGE:figures/full_fig_p036_6.png] 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

38 extracted references · 38 canonical work pages · 23 internal anchors

  1. [1]

    Scheherazade: Evaluating Chain-of-Thought Math Reasoning in LLMs with Chain-of-Problems

    Scheherazade: Evaluating chain-of-thought math reasoning in llms with chain-of-problems , author=. arXiv preprint arXiv:2410.00151 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?

    Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? , author=. arXiv preprint arXiv:2504.13837 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Reasoning with Sampling: Your Base Model is Smarter Than You Think

    Reasoning with sampling: Your base model is smarter than you think , author=. arXiv preprint arXiv:2510.14901 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2510.14967 , year=

    Information Gain-based Policy Optimization: A Simple and Effective Approach for Multi-Turn LLM Agents , author=. arXiv preprint arXiv:2510.14967 , year=

    work page arXiv

  5. [5]

    arXiv preprint arXiv:2510.11701 , year=

    Demystifying Reinforcement Learning in Agentic Reasoning , author=. arXiv preprint arXiv:2510.11701 , year=

    work page arXiv

  6. [6]

    Verifying Meta-Awareness via Predictive Rewards in Reasoning Models

    Meta-Awareness Enhances Reasoning Models: Self-Alignment Reinforcement Learning , author=. arXiv preprint arXiv:2510.03259 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2509.26628 , year=

    Attention as a Compass: Efficient Exploration for Process-Supervised RL in Reasoning Models , author=. arXiv preprint arXiv:2509.26628 , year=

    work page arXiv

  8. [8]

    LLM-guided Hierarchical Search for End-to-end Reasoning Intensive Retrieval

    LLM-guided Hierarchical Retrieval , author=. arXiv preprint arXiv:2510.13217 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    On Provable Length and Compositional Generalization

    On provable length and compositional generalization , author=. arXiv preprint arXiv:2402.04875 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Sub-Task Decomposition Enables Learning in Sequence to Sequence Tasks

    Sub-task decomposition enables learning in sequence to sequence tasks , author=. arXiv preprint arXiv:2204.02892 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    The Expressive Power of Transformers with Chain of Thought

    The expresssive power of transformers with chain of thought , author=. arXiv preprint arXiv:2310.07923 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Auto-Regressive Next-Token Predictors are Universal Learners

    Auto-regressive next-token predictors are universal learners , author=. arXiv preprint arXiv:2309.06979 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    On Limitations of the Transformer Architecture

    On Limitations of the Transformer Architecture , author=. arXiv preprint arXiv:2402.08164 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models , year=

    Do large language models have compositional ability? an investigation into limitations and scalability , author=. ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models , year=

    work page 2024

  15. [15]

    arXiv preprint arXiv:2503.17523 , year=

    Bayesian teaching enables probabilistic reasoning in large language models , author=. arXiv preprint arXiv:2503.17523 , year=

    work page arXiv

  16. [16]

    arXiv preprint arXiv:2509.10739 , year=

    Reasoning Under Uncertainty: Exploring Probabilistic Reasoning Capabilities of LLMs , author=. arXiv preprint arXiv:2509.10739 , year=

    work page arXiv

  17. [17]

    arXiv preprint arXiv:2505.20561 , year=

    Beyond Markovian: Reflective Exploration via Bayes-Adaptive RL for LLM Reasoning , author=. arXiv preprint arXiv:2505.20561 , year=

    work page arXiv

  18. [18]

    arXiv preprint arXiv:2509.25810 , year=

    Learning to Reason as Action Abstractions with Scalable Mid-Training RL , author=. arXiv preprint arXiv:2509.25810 , year=

    work page arXiv

  19. [19]

    From Reasoning to Super-Intelligence: A Search-Theoretic Perspective

    From Reasoning to Super-Intelligence: A Search-Theoretic Perspective , author=. arXiv preprint arXiv:2507.15865 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    Evaluating Large Language Models Trained on Code

    Evaluating large language models trained on code , author=. arXiv preprint arXiv:2107.03374 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    Science , volume=

    Competition-level code generation with alphacode , author=. Science , volume=. 2022 , publisher=

    work page 2022

  22. [22]

    Google DeepMind AlphaCode Team, AlphaCode 2 Technical Report ,year=

    work page

  23. [23]

    Code Generation with AlphaCodium: From Prompt Engineering to Flow Engineering

    Code generation with alphacodium: From prompt engineering to flow engineering , author=. arXiv preprint arXiv:2401.08500 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    2024 , publisher=

    Openai o3 breakthrough high score on arc-agi-pub , author=. 2024 , publisher=

    work page 2024

  26. [26]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

    work page

  27. [27]

    OpenAI o1 System Card

    Openai o1 system card , author=. arXiv preprint arXiv:2412.16720 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

    Scaling llm test-time compute optimally can be more effective than scaling model parameters , author=. arXiv preprint arXiv:2408.03314 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    Demystifying Long Chain-of-Thought Reasoning in LLMs

    Demystifying long chain-of-thought reasoning in llms , author=. arXiv preprint arXiv:2502.03373 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    Reflexion: Language Agents with Verbal Reinforcement Learning

    Reflexion: Language agents with verbal reinforcement learning, 2023 , author=. URL https://arxiv. org/abs/2303.11366 , volume=

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  31. [31]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models

    Self-consistency improves chain of thought reasoning in language models , author=. arXiv preprint arXiv:2203.11171 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    Advances in neural information processing systems , volume=

    Tree of thoughts: Deliberate problem solving with large language models , author=. Advances in neural information processing systems , volume=

    work page

  33. [33]

    Kimi k1.5: Scaling Reinforcement Learning with LLMs

    Kimi k1. 5: Scaling reinforcement learning with llms , author=. arXiv preprint arXiv:2501.12599 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  34. [34]

    Rethinking the Unsolvable: When In-Context Search Meets Test-Time Scaling

    Rethinking the Unsolvable: When In-Context Search Meets Test-Time Scaling , author=. arXiv preprint arXiv:2505.22290 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [35]

    Qwen3 Technical Report

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  36. [36]

    Improving Test-Time Search for LLMs with Backtracking Against In-Context Value Verifiers , author=

    work page

  37. [37]

    Towards System 2 Reasoning in LLMs: Learning How to Think With Meta Chain-of-Thought

    Towards system 2 reasoning in llms: Learning how to think with meta chain-of-thought , author=. arXiv preprint arXiv:2501.04682 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  38. [38]

    International Conference on Machine Learning , pages=

    Transformers learn in-context by gradient descent , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    work page 2023