pith. sign in

arxiv: 2606.23595 · v1 · pith:I66WEYEAnew · submitted 2026-06-22 · 💻 cs.AI

SPIRAL: Learning to Search and Aggregate

Jubayer Ibn Hamid , Ifdita Hasan Orney , Michael Y. Li , Omar Shaikh , Yoonho Lee , Dorsa Sadigh , Chelsea Finn , Noah Goodman This is my paper

Pith reviewed 2026-06-26 08:27 UTC · model grok-4.3

classification 💻 cs.AI
keywords SPIRALreinforcement learninglanguage model reasoninginference compute scalingchain-of-thoughttrace aggregationset reinforcement learningparallel sampling
0
0 comments X

The pith

SPIRAL trains language models end-to-end to generate parallel reasoning traces and then aggregate them into a final answer using combined set and standard reinforcement learning.

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

The paper presents SPIRAL as a way to move beyond training language models only for single sequential reasoning traces. Instead, the framework has the model produce a set of independent traces in parallel, each via chain-of-thought, and then generate an aggregation trace that conditions on the full set. Set reinforcement learning teaches the model to create traces that are collectively useful, while standard reinforcement learning optimizes the aggregation step, with everything trained jointly against the reward on the final output. Experiments on reasoning tasks demonstrate that this approach scales more efficiently with added inference compute than GRPO, delivering up to 11 times better scaling efficiency and 15 percent higher performance when the number of traces, their length, and the aggregation are all increased together.

Core claim

SPIRAL optimizes a language model across three inference primitives in one pipeline: it samples multiple independent sequential reasoning traces in parallel, then produces a final aggregation trace conditioned on that set, with the entire system trained end-to-end so that set reinforcement learning improves the collective utility of the traces and standard reinforcement learning improves the quality of the aggregation.

What carries the argument

The end-to-end combination of set reinforcement learning, which rewards traces for their joint usefulness to an aggregator, and standard reinforcement learning, which rewards the aggregator for producing better final responses.

If this is right

  • Scaling the number of parallel traces and the aggregation depth produces larger performance gains than scaling any single primitive alone.
  • The model learns to produce traces that are useful as a group rather than individually optimal.
  • End-to-end training removes the need for separate post-training stages for search versus aggregation.
  • Performance improvements hold when all three compute primitives are scaled simultaneously on reasoning benchmarks.

Where Pith is reading between the lines

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

  • The same joint training pattern could be applied to other multi-trace methods such as tree-of-thought or Monte Carlo tree search scaffolds.
  • Models trained this way might generalize the learned search-and-aggregate behavior to tasks outside the original reasoning benchmarks.
  • If trace quality collapses under larger set sizes, additional regularization on trace diversity would become necessary.

Load-bearing premise

That reinforcement learning signals alone will cause the model to generate a set of traces whose diversity and quality support effective aggregation without any separate mechanisms to enforce those properties.

What would settle it

A controlled comparison in which models trained with SPIRAL show no gain, or a loss, in final accuracy when the number of parallel traces is increased while holding total inference tokens fixed, relative to models trained only on aggregation.

Figures

Figures reproduced from arXiv: 2606.23595 by Chelsea Finn, Dorsa Sadigh, Ifdita Hasan Orney, Jubayer Ibn Hamid, Michael Y. Li, Noah Goodman, Omar Shaikh, Yoonho Lee.

Figure 1
Figure 1. Figure 1: Overview of Spiral. Top: Spiral trains a language model (LM) to use sequential, parallel, and aggregative inference compute end-to-end. Given an input x, we independently sample n parallel reasoning traces. The LM then synthesizes these into a final aggregation trace y∗. Using only the final reward r(x, y∗), Spiral optimizes the parallel generations via set RL and the aggregation trace via standard RL. Con… view at source ↗
Figure 2
Figure 2. Figure 2: Spiral consists of two levels of generation. First, the model samples N1 search traces, y1, . . . , yN1 ∼ πθ(· | x), from the problem x. It then uniformly samples K sets, each consisting of n search traces. For each set, it samples N2 aggregation traces conditioned on both the problem and the set Gi , y Gi 1 , . . . , y Gi N2 ∼ πθ(· | x, Gi). Rewards are evaluated only at the end of this full inference com… view at source ↗
Figure 3
Figure 3. Figure 3: Pass@k evaluation on test sets. The x-axis is the number of independent attempts, k, used in the evaluation and the y-axis is the coverage of the test set [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Pass@1 evaluation under recursive self-aggregation [VJM+25]. The x-axis is the number of recursive self-aggregation steps used and the y-axis is the pass@1 rate. set. As such, the model is not encouraged to immediately collapse its entropy on a generation; search traces that are incorrect but still enable the model to craft high quality aggregation traces are explicitly encouraged by Spiral. In fact, as fi… view at source ↗
Figure 5
Figure 5. Figure 5: Token-level entropy over training. For Spiral, we plot the entropy over the search traces only to make the comparison to GRPO fair [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of models under scaling sequential compute. The x-axis is the maximum number of tokens the model is allowed to sample within a chain-of-thought and the y-axis is the pass@1 rate. model and GRPO, achieving up to 13.5% higher performance. This suggests that Spiral learns search and refinement behaviors that are better suited to scaling parallel and aggregative inference compute at test time. Next,… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of inference methods scaling parallel traces. The x-axis is the number of parallel traces sampled. GRPO is not trained to aggregate, whereas Spiral trains a model to search and aggregate. Majority voting is a rule-based aggregation procedure whereas self-aggregation is model-based. We plot pass@1 value for reference [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of inference compute scaling and model pairs against token usage. The x-axis is the maximum token budget provided to each model and inference compute method pair. Sequential compute scaling is up to 32k tokens due to the context length of our base model. Finally, we make a comparison of the various inference compute methods in terms of the efficiency in their token usage. In this comparison, we … view at source ↗
read the original abstract

Language model reasoning can be substantially improved at test time via scaffolds that scale inference compute across different primitives -- sequential reasoning within a trace, independently sampled parallel traces, and aggregation of multiple reasoning traces into a final response. During post-training, however, language models are optimized only for sequential reasoning within a single trace. We introduce Sequential-Parallel-Aggregative Reinforcement Learning (SPIRAL), a framework in which a language model is trained to use all three primitives, as part of a unified inference compute pipeline. Concretely, the language model first samples a set of independent traces in parallel, each produced through sequential chain-of-thought reasoning, and then generates a final aggregation trace conditioned on those traces; all components are optimized end-to-end against the reward of the final aggregated response. To train this system, SPIRAL uses set reinforcement learning to teach models to produce a set of traces that are collectively useful for an aggregator and standard reinforcement learning to teach models to aggregate the set into improved final responses. Our experiments on reasoning tasks show that SPIRAL effectively scales with inference compute, outperforming GRPO by up to 11$\times$ scaling efficiency and 15% higher performance when all three compute primitives are scaled.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces SPIRAL, a framework for end-to-end training of language models to perform sequential reasoning, parallel trace sampling, and aggregation of multiple traces using a combination of set reinforcement learning and standard reinforcement learning. The model is optimized against the reward of the final aggregated response. The paper claims that this approach allows effective scaling with inference compute, outperforming GRPO by up to 11× in scaling efficiency and 15% in performance on reasoning tasks when scaling all three primitives.

Significance. If the results are robust, this would represent a meaningful advance in training models for multi-primitive inference-time compute scaling. The end-to-end optimization is a strength, as is the focus on collective utility of trace sets.

major comments (1)
  1. [Abstract] Abstract: The set reinforcement learning component is presented as teaching the model to produce traces that are collectively useful for the aggregator, but the description provides no auxiliary objective, regularization term, or sampling constraint to encourage diversity or complementarity among the parallel traces. This is load-bearing for the central scaling claims, as the reported 11× scaling efficiency when increasing parallel samples would be undermined if gradients favor redundant high-reward patterns rather than complementary ones.
minor comments (1)
  1. [Abstract] Abstract: The distinction between 'set reinforcement learning' and 'standard reinforcement learning' is invoked without a precise definition of the objectives or how the set-level reward is computed and backpropagated.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback. We address the major comment on the set reinforcement learning component below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The set reinforcement learning component is presented as teaching the model to produce traces that are collectively useful for the aggregator, but the description provides no auxiliary objective, regularization term, or sampling constraint to encourage diversity or complementarity among the parallel traces. This is load-bearing for the central scaling claims, as the reported 11× scaling efficiency when increasing parallel samples would be undermined if gradients favor redundant high-reward patterns rather than complementary ones.

    Authors: We thank the referee for highlighting this important point. In SPIRAL the set reinforcement learning objective optimizes trace generation end-to-end against the reward of the aggregator's final output. Because the aggregator is conditioned on the full set and receives reward only on the aggregated answer, traces that are redundant contribute no additional value and therefore receive lower effective credit under the set-level signal. This mechanism encourages complementarity without requiring an auxiliary diversity term. The abstract condenses this as 'collectively useful,' but we agree the presentation would benefit from greater explicitness. We will revise the abstract and method section to clarify the end-to-end objective and will add an empirical analysis of trace diversity to the appendix. revision: yes

Circularity Check

0 steps flagged

No circularity in SPIRAL derivation or claims

full rationale

The paper presents SPIRAL as an empirical training framework that optimizes a language model end-to-end against the final aggregated reward using set RL for trace generation and standard RL for aggregation. No equations, self-definitions, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. Performance claims (scaling efficiency vs GRPO) rest on experimental comparisons rather than any closed mathematical loop or ansatz smuggled via prior work. This is a standard RL-for-LLM paper whose central method does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no details on specific free parameters, axioms, or invented entities; assessment is limited to the high-level description given.

pith-pipeline@v0.9.1-grok · 5766 in / 1009 out tokens · 23365 ms · 2026-06-26T08:27:20.140108+00:00 · methodology

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

84 extracted references · 3 canonical work pages

  1. [1]

    STaR: Bootstrapping Reasoning With Reasoning , url =

    Zelikman, Eric and Wu, Yuhuai and Mu, Jesse and Goodman, Noah , booktitle =. STaR: Bootstrapping Reasoning With Reasoning , url =

  2. [2]

    Proceedings of the Nineteenth International Conference on Machine Learning , pages =

    Kakade, Sham and Langford, John , title =. Proceedings of the Nineteenth International Conference on Machine Learning , pages =. 2002 , isbn =

    2002

  3. [3]

    2026 , eprint=

    Polychromic Objectives for Reinforcement Learning , author=. 2026 , eprint=

    2026

  4. [4]

    2025 , eprint=

    Pass@K Policy Optimization: Solving Harder Reinforcement Learning Problems , author=. 2025 , eprint=

    2025

  5. [5]

    2025 , eprint=

    Jointly Reinforcing Diversity and Quality in Language Model Generations , author=. 2025 , eprint=

    2025

  6. [6]

    2017 , eprint=

    Trust Region Policy Optimization , author=. 2017 , eprint=

    2017

  7. [7]

    2025 , eprint=

    VinePPO: Refining Credit Assignment in RL Training of LLMs , author=. 2025 , eprint=

    2025

  8. [8]

    International Conference on Machine Learning , year=

    Approximately Optimal Approximate Reinforcement Learning , author=. International Conference on Machine Learning , year=

  9. [9]

    2017 , eprint=

    Proximal Policy Optimization Algorithms , author=. 2017 , eprint=

    2017

  10. [10]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    2024

  11. [11]

    2026 , eprint=

    Poly-EPO: Training Exploratory Reasoning Models , author=. 2026 , eprint=

    2026

  12. [12]

    2025 , eprint=

    The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models , author=. 2025 , eprint=

    2025

  13. [13]

    2025 , eprint=

    LLMs Can Generate a Better Answer by Aggregating Their Own Responses , author=. 2025 , eprint=

    2025

  14. [14]

    arXiv preprint arXiv:2509.26626 , year=

    Recursive self-aggregation unlocks deep thinking in large language models , author=. arXiv preprint arXiv:2509.26626 , year=

    arXiv

  15. [15]

    International Conference on Learning Representations , volume=

    Mixture-of-agents enhances large language model capabilities , author=. International Conference on Learning Representations , volume=

  16. [16]

    Advances in Neural Information Processing Systems , volume=

    Atom of thoughts for markov llm test-time scaling , author=. Advances in Neural Information Processing Systems , volume=

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    Chain of agents: Large language models collaborating on long-context tasks , author=. Advances in Neural Information Processing Systems , volume=

  18. [18]

    International Conference on Learning Representations , volume=

    Skeleton-of-thought: Prompting llms for efficient parallel generation , author=. International Conference on Learning Representations , volume=

  19. [19]

    2025 , eprint=

    THREAD: Thinking Deeper with Recursive Spawning , author=. 2025 , eprint=

    2025

  20. [20]

    arXiv preprint arXiv:2504.07081 , year=

    Self-steering language models , author=. arXiv preprint arXiv:2504.07081 , year=

    arXiv

  21. [21]

    Forty-first international conference on machine learning , year=

    Improving factuality and reasoning in language models through multiagent debate , author=. Forty-first international conference on machine learning , year=

  22. [22]

    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=

  23. [23]

    2021 , eprint=

    Training Verifiers to Solve Math Word Problems , author=. 2021 , eprint=

    2021

  24. [24]

    arXiv preprint arXiv:1712.01815 , year=

    Mastering chess and shogi by self-play with a general reinforcement learning algorithm , author=. arXiv preprint arXiv:1712.01815 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2203.11171 , year=

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

    Pith/arXiv arXiv

  26. [26]

    2026 , eprint=

    OpenAI o1 System Card , author=. 2026 , eprint=

    2026

  27. [27]

    arXiv preprint arXiv:2507.06261 , year=

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2602.10177 , year=

    Towards autonomous mathematics research , author=. arXiv preprint arXiv:2602.10177 , year=

    arXiv

  29. [29]

    arXiv preprint arXiv:2507.15855 , year=

    Winning gold at imo 2025 with a model-agnostic verification-and-refinement pipeline , author=. arXiv preprint arXiv:2507.15855 , year=

    arXiv 2025

  30. [30]

    arXiv preprint arXiv:2501.12948 , year=

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

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2506.13131 , year=

    Alphaevolve: A coding agent for scientific and algorithmic discovery , author=. arXiv preprint arXiv:2506.13131 , year=

    Pith/arXiv arXiv

  32. [32]

    2025 , url =

    Welcome to the Era of Experience , author =. 2025 , url =

    2025

  33. [33]

    2025 , eprint=

    Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs , author=. 2025 , eprint=

    2025

  34. [34]

    arXiv preprint arXiv:2407.21787 , year=

    Large language monkeys: Scaling inference compute with repeated sampling , author=. arXiv preprint arXiv:2407.21787 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2408.03314 , year=

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

    Pith/arXiv arXiv

  36. [36]

    arXiv preprint arXiv:2601.16175 , year=

    Learning to discover at test time , author=. arXiv preprint arXiv:2601.16175 , year=

    Pith/arXiv arXiv

  37. [37]

    2025 , eprint=

    Learning Adaptive Parallel Reasoning with Language Models , author=. 2025 , eprint=

    2025

  38. [38]

    Miguel Calvo-Fullana, Santiago Paternain, Luiz F

    Noam Brown and Tuomas Sandholm , title =. Science , volume =. 2019 , doi =. https://www.science.org/doi/pdf/10.1126/science.aay2400 , abstract =

    work page doi:10.1126/science.aay2400 2019

  39. [39]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Math-shepherd: Verify and reinforce llms step-by-step without human annotations , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  40. [40]

    2023 , eprint=

    Self-Refine: Iterative Refinement with Self-Feedback , author=. 2023 , eprint=

    2023

  41. [41]

    Advances in neural information processing systems , volume=

    Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume=

  42. [42]

    2026 , eprint=

    OpenDeepThink: Parallel Reasoning via Bradley--Terry Aggregation , author=. 2026 , eprint=

    2026

  43. [43]

    2025 , eprint=

    Population-Evolve: a Parallel Sampling and Evolutionary Method for LLM Math Reasoning , author=. 2025 , eprint=

    2025

  44. [44]

    2025 , eprint=

    ParaThinker: Native Parallel Thinking as a New Paradigm to Scale LLM Test-time Compute , author=. 2025 , eprint=

    2025

  45. [45]

    2026 , eprint=

    PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning , author=. 2026 , eprint=

    2026

  46. [46]

    2025 , eprint=

    Multiverse: Your Language Models Secretly Decide How to Parallelize and Merge Generation , author=. 2025 , eprint=

    2025

  47. [47]

    2025 , eprint=

    Learning to Reason Across Parallel Samples for LLM Reasoning , author=. 2025 , eprint=

    2025

  48. [48]

    arXiv preprint arXiv:2503.19595 , year=

    Optimizing language models for inference time objectives using reinforcement learning , author=. arXiv preprint arXiv:2503.19595 , year=

    arXiv

  49. [49]

    and McAllester, David and Singh, Satinder and Mansour, Yishay , title =

    Sutton, Richard S. and McAllester, David and Singh, Satinder and Mansour, Yishay , title =. Proceedings of the 13th International Conference on Neural Information Processing Systems , pages =. 1999 , publisher =

    1999

  50. [50]

    and Barto, Andrew G

    Sutton, Richard S. and Barto, Andrew G. , title =. 2018 , isbn =

    2018

  51. [51]

    A Natural Policy Gradient , url =

    Kakade, Sham M , booktitle =. A Natural Policy Gradient , url =

  52. [52]

    Sutton and Mohammad Ghavamzadeh and Mark Lee , keywords =

    Shalabh Bhatnagar and Richard S. Sutton and Mohammad Ghavamzadeh and Mark Lee , keywords =. Natural actor–critic algorithms , journal =. 2009 , issn =. doi:https://doi.org/10.1016/j.automatica.2009.07.008 , url =

    work page doi:10.1016/j.automatica.2009.07.008 2009

  53. [53]

    2013 , eprint=

    Off-Policy Actor-Critic , author=. 2013 , eprint=

    2013

  54. [54]

    Lillicrap and Jonathan J

    Timothy P. Lillicrap and Jonathan J. Hunt and Alexander Pritzel and Nicolas Heess and Tom Erez and Yuval Tassa and David Silver and Daan Wierstra , editor =. Continuous control with deep reinforcement learning , booktitle =. 2016 , url =

    2016

  55. [55]

    2017 , eprint=

    Sample Efficient Actor-Critic with Experience Replay , author=. 2017 , eprint=

    2017

  56. [56]

    2025 , eprint=

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale , author=. 2025 , eprint=

    2025

  57. [57]

    2025 , eprint=

    Group Sequence Policy Optimization , author=. 2025 , eprint=

    2025

  58. [58]

    2025 , eprint=

    Understanding R1-Zero-Like Training: A Critical Perspective , author=. 2025 , eprint=

    2025

  59. [59]

    arXiv preprint arXiv:2506.13585 , year=

    Minimax-m1: Scaling test-time compute efficiently with lightning attention , author=. arXiv preprint arXiv:2506.13585 , year=

    Pith/arXiv arXiv

  60. [60]

    POLARIS: A Post-Training Recipe for Scaling Reinforcement Learning on Advanced Reasoning Models , url =

    An, Chenxin and Xie, Zhihui and Li, Xiaonan and Li, Lei and Zhang, Jun and Gong, Shansan and Zhong, Ming and Xu, Jingjing and Qiu, Xipeng and Wang, Mingxuan and Kong, Lingpeng , year =. POLARIS: A Post-Training Recipe for Scaling Reinforcement Learning on Advanced Reasoning Models , url =

  61. [61]

    ICML 2025 Workshop on Long-Context Foundation Models , year=

    Simple, Scalable Reasoning via Iterated Summarization , author=. ICML 2025 Workshop on Long-Context Foundation Models , year=

    2025

  62. [62]

    2026 , eprint=

    Meta-Harness: End-to-End Optimization of Model Harnesses , author=. 2026 , eprint=

    2026

  63. [63]

    2025 , eprint=

    Rethinking Thinking Tokens: LLMs as Improvement Operators , author=. 2025 , eprint=

    2025

  64. [64]

    2026 , eprint=

    Scaling Test-Time Compute for Agentic Coding , author=. 2026 , eprint=

    2026

  65. [65]

    Large Language Models Sensitivity to The Order of Options in Multiple-Choice Questions

    Pezeshkpour, Pouya and Hruschka, Estevam. Large Language Models Sensitivity to The Order of Options in Multiple-Choice Questions. Findings of the Association for Computational Linguistics: NAACL 2024. 2024. doi:10.18653/v1/2024.findings-naacl.130

    work page doi:10.18653/v1/2024.findings-naacl.130 2024

  66. [66]

    Advances in Neural Information Processing Systems , volume=

    Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? , author=. Advances in Neural Information Processing Systems , volume=

  67. [67]

    arXiv preprint arXiv:2210.03493 , year=

    Automatic chain of thought prompting in large language models , author=. arXiv preprint arXiv:2210.03493 , year=

    Pith/arXiv arXiv

  68. [68]

    2026 , eprint=

    Composer 2 Technical Report , author=. 2026 , eprint=

    2026

  69. [69]

    2026 , url =

    Thinking Machines Lab , title =. 2026 , url =

    2026

  70. [70]

    2026 , eprint=

    Neural Garbage Collection: Learning to Forget while Learning to Reason , author=. 2026 , eprint=

    2026

  71. [71]

    2026 , eprint=

    Forget, Then Recall: Learnable Compression and Selective Unfolding via Gist Sparse Attention , author=. 2026 , eprint=

    2026

  72. [72]

    2025 , eprint=

    Creating General User Models from Computer Use , author=. 2025 , eprint=

    2025

  73. [73]

    2026 , eprint=

    Learning Next Action Predictors from Human-Computer Interaction , author=. 2026 , eprint=

    2026

  74. [74]

    2026 , eprint=

    Using Reward Uncertainty to Induce Diverse Behaviour in Reinforcement Learning , author=. 2026 , eprint=

    2026

  75. [75]

    2025 , eprint=

    S*: Test Time Scaling for Code Generation , author=. 2025 , eprint=

    2025

  76. [76]

    2026 , eprint=

    Remarks on the disproof of the unit distance conjecture , author=. 2026 , eprint=

    2026

  77. [77]

    2026 , eprint=

    First Proof , author=. 2026 , eprint=

    2026

  78. [78]

    2026 , eprint=

    First Proof Second Batch , author=. 2026 , eprint=

    2026

  79. [79]

    2026 , month = feb, day =

    Harness Engineering: Leveraging Codex in an Agent-First World , author =. 2026 , month = feb, day =

    2026

  80. [80]

    2026 , eprint=

    V_1 : Unifying Generation and Self-Verification for Parallel Reasoners , author=. 2026 , eprint=

    2026

Showing first 80 references.