Pith. sign in

REVIEW 3 major objections 5 minor 71 references

Chained Recursive Language Models for Multi-Iteration Reasoning

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Calling the same language model repeatedly in fresh roots, each handed only a plain-text summary, blackboard, and task artifacts, yields a 13.75-point average accuracy gain over single-call answering on four long-context benchmarks.

desk verdict A clearly specified inference-time chaining idea whose headline accuracy gain is not yet interpretable because the compute-matched comparison is missing. read the letter →

arxiv 2608.05124 v1 pith:255B6V63 submitted 2026-08-05 cs.CL cs.AIcs.ITcs.LGeess.SPmath.IT

Purbesh Mitra , Sennur Ulukus This is my paper
classification cs.CLcs.AIcs.ITcs.LGeess.SPmath.IT
keywords ChainedRLMrecursivelanguagemodelslong-contextreasoningartifact-mediatedhandoffmulti-iterationinference-timearchitecturecontextrottoolcalling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Chained Recursive Language Models (Chained RLM) is an inference-time architecture in which one LLM is called repeatedly as a sequence of fresh roots, each a new inference that receives the same problem and context but only a compact plain-text continuity state from earlier roots. The paper claims that passing a summary, a blackboard, and durable task-specific artifacts between roots lets later calls inspect, correct, and extend intermediate evidence instead of redoing the whole task inside one trajectory. On four long-context benchmarks, it reports an average absolute accuracy gain of 13.75 percentage points over a direct single-call baseline, with the largest gains on counting, ordering, and aggregation-heavy tasks. The cost is roughly 2.5 times the tokens and dollars of direct answering. A sympathetic reader would care because the method requires no training and turns intermediate reasoning into editable plain text, a concrete way to fight context rot.

What carries the argument

The central object is the artifact-mediated handoff: a root writes one or more plain-text artifacts, such as candidate ledgers, extraction tables, derivations, audit notes, or checklists, and ends with either FINAL or a plain-text HANDOFF with SUMMARY, BLACKBOARD, and NEXT. A later fresh root of the same model receives the original problem and context plus that chain state, and the Chain Rules instruct it to read the artifacts first, preserve their structure, and update or correct the information inside them. The blackboard is deliberately not parsed by the host, so the model is both author and reader of the continuity state; this makes the artifact the externalized working memory that is supposed to stop stale conclusions from propagating.

What would settle it

Log every root's read and write actions in the artifact workspace and check whether later roots read predecessor artifacts before doing new extraction and whether the artifact structure survives. If artifact-compliance does not correlate with final accuracy on the four benchmarks, or if a compute-matched ablation that withholds artifacts matches the chained accuracy, the reported gains would be explained by repeated attempts rather than by artifact-mediated continuation.

Watch

Extended reading notes

Core claim

The paper's central claim is that durable plain-text artifacts can carry reasoning state across multiple fresh calls of the same model, and that this artifact-mediated continuation is what makes chained inference more accurate than a single trajectory. In the proposed system, a root either submits FINAL or a three-section HANDOFF containing SUMMARY, BLACKBOARD, and NEXT; artifacts written by earlier roots are the main object of continuation, and later roots are instructed to preserve, correct, and audit them before finalizing. Against direct LLM answering on RULER, BABILong, LongBench v2, and OOLONG-real, Chained RLM reports accuracy gains of 5, 15, 11, and 24 percentage points respectively, and the paper attributes the uneven but consistently positive effect to tasks in which a model must keep partial evidence consistent across long contexts. Chained RLM averages 2.6 roots and 1.6 handoffs per task, roughly doubling input tokens and nearly quadrupling output tokens, and the paper presents this as the price of making intermediate state inspectable and auditable.

Load-bearing premise

The whole gain rests on the assumption that a later fresh instance of the same model will actually read the artifacts left by earlier roots and preserve or correct them rather than restarting extraction from scratch; the paper instructs this behavior but provides no evidence that models comply.

Editorial extensions

If this is right

  • On four long-context benchmarks, Chained RLM improves pass@1 accuracy over a single direct LLM response by an average of 13.75 absolute percentage points.
  • The gain is largest where the task requires preserving and aggregating partial evidence: 24 points on OOLONG-real, 15 on BABILong, 11 on LongBench v2, and 5 on RULER.
  • The gain costs about 2.5 times the token and dollar budget, with an average of 2.6 roots and 1.6 handoffs per task.
  • Because the system is not trained and needs no specialized verifier, the same handoff mechanism can be applied to any tool-calling model on tasks with checkable intermediate artifacts.
  • If a later root ignores a good artifact and restarts from a worse state, chaining degenerates into repeated independent attempts, so reported accuracy depends on the model following the chain rules.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the reported comparison gives the chain multiple fresh attempts and the baseline only one, an ablation with the same number of fresh roots but artifacts withheld would separate the contribution of artifact continuity from the contribution of repeated sampling.
  • Editorial inference: the plain-text artifacts could themselves be used as compliance data, for example by measuring how often later roots edit the existing structure rather than creating new files, which would test the central mechanism directly.
  • Editorial inference: an adaptive stopping rule that reads the blackboard's confidence and open questions could tune the accuracy-versus-cost trade-off, since the paper's gains are uneven across task types.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Chained RLM, an inference-time architecture in which the same underlying LLM is invoked as a sequence of fresh roots, each receiving the original problem and context along with a compact plain-text continuity state composed of a summary, a blackboard, and durable artifacts. The system is described formally in Section 3, including the root execution, handoff format, and chain rules in Section 4. The authors evaluate Chained RLM on four long-context benchmarks (RULER, BABILong, LongBench v2, OOLONG-real) using GPT-5-mini, reporting an average absolute accuracy gain of 13.75 percentage points over a single-call regular LLM baseline at roughly 2.6 times the root calls and 2.5 times the token cost. The paper's central empirical claim is that artifact-mediated fresh-context continuation, rather than the additional compute, drives the observed improvement.

Significance. The architectural idea of using durable plain-text artifacts as externally inspectable checkpoints across fresh inference roots is clean, concrete, and potentially useful for long-context aggregation and multi-hop reasoning. The system model and chain rules are specific enough to reimplement, and the authors honestly acknowledge several limitations in Section 6. However, the empirical evidence as presented does not yet establish the central claim: the accuracy gains are not compared against a compute-matched control, and no trace-level evidence shows that later roots actually read and preserve artifacts. If the compute-matched experiment were added and showed comparable gains, the contribution would be a practical inference-time method for long-context tasks. As it stands, the paper is a system description with suggestive but uncontrolled results.

major comments (3)
  1. [§5.1, Table 2, §5.3] The central empirical claim—that Chained RLM improves accuracy by an average of 13.75 percentage points over a regular LLM baseline (Section 5.3)—is not supported by a controlled comparison. Table 2 shows that Chained RLM uses 1.8–3.4 root calls and roughly 1.8–3.4 times the input tokens of the single-call baseline, and Section 5.1 explicitly states that the compute-matched setting is deferred to 'the final experiment,' which is not reported. Without a baseline consuming comparable inference compute (e.g., multiple independent samples, self-consistency, or repeated roots without artifacts), the reported gain could be attributable to additional compute rather than to artifact-mediated continuation. This comparison is load-bearing for the paper's central claim and must be included before the accuracy gain can be interpreted.
  2. [Table 1, §5.2] Table 1 reports accuracy values without standard errors, confidence intervals, or the number of questions L used in Eq. (4) for each benchmark, and the 'toolcall' column is empty. The conclusion's statement that Chained RLM improves over 'vanilla tool-calling RLMs' is not supported by any reported tool-calling baseline in the table. The evaluation should report per-benchmark L, variance estimates or raw counts, and either fill the toolcall column with an actual baseline or remove the claim about surpassing vanilla tool-calling RLMs.
  3. [§4.1, Chain Rules, §6] The mechanism's causal role rests on the assumption that later roots will read, preserve, and update artifacts rather than restart extraction from scratch (Chain Rules 3–6). Section 6 acknowledges that the host does not enforce artifact reading, that artifact quality is model-dependent, and that 'the whole chain can drift.' The paper provides no quantitative evidence about compliance—for example, how often roots called list_artifacts(), whether artifacts were modified, or how often a root ignored the handoff. Without such measurements, the observed accuracy gains could be explained by repeated independent fresh attempts rather than artifact continuation. At minimum, the paper should report a trace-level analysis of artifact usage across a sample of tasks.
minor comments (5)
  1. [§4.2] The phrase 'artifact-mediated handoff of teh resoning chain' contains two typos; it should read 'of the reasoning chain.'
  2. [§5.2] The word 'resoyrce' should be 'resource' in the sentence 'These secondary metrics are crucial for determining the accuracy and resoyrce trade-off.'
  3. [§3.4] The text 'Y AML' should be 'YAML' in the sentence 'We do not use JSON, XML, Y AML, or a host-parsed schema for the handoff.'
  4. [§5.3] The citation 'RULER Hsieh et al., 2024' is missing its closing parenthesis, and the sentence 'the cost increase is due to the the price of making intermediate reasoning more robust' contains a duplicated 'the.'
  5. [§1] The opening sentence 'Recently, long context reasoning in large language models (LLMs) have become' has a subject-verb agreement error; 'reasoning' is singular, so it should be 'has become.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported accuracy gains are external benchmark measurements, the architecture is defined independently of the results, and the cited related work is background rather than load-bearing evidence.

full rationale

The paper does not derive its accuracy results from its own definitions or fitted parameters. The central claim in Section 5.3 ("Chained RLM improves accuracy on all four benchmarks, with an average absolute gain of 13.75 percentage points over the regular LLM baseline") is an empirical comparison against externally defined benchmarks (RULER, BABILong, LongBench v2, OOLONG-real), and the pass@1 metric in Eq. (4) is a standard statistical estimator, not a quantity defined in terms of the chaining mechanism. No equation in the paper is fitted to the accuracy values, and no fitted parameter is renamed as a prediction. The handoff and artifact rules in Section 4.1 and the system prompt are architectural instructions, not mathematical constraints that force a particular accuracy outcome; they describe a procedure whose effect is measured, not assumed. The self-citations to MOTIF (Mitra and Ulukus, 2025) are used only as related work and are not load-bearing for the paper's empirical claim. The paper also does not import a uniqueness theorem or ansatz from prior work by the same authors; the fresh-root continuation design is introduced in this paper and evaluated directly. The limitations the paper itself states in Section 6 (artifact quality is model-dependent, the chain can drift, and compute is higher) are honest scope notes rather than evidence of circularity. The one substantive weakness is the deferred compute-matched comparison noted in Section 5.1: "For a fair comparison, the final experiment also includes a compute-matched setting, but in the first version we use the direct LLM baseline as a clean lower-complexity reference point." This means the reported 13.75-point gain may partly reflect additional inference compute rather than artifact-mediated continuation specifically. That is a correctness or interpretability risk, not circularity: the comparison still tests an independent behavioral hypothesis using external benchmarks, and no quantity in the claim is defined in terms of the benchmark outcomes themselves. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No parameters were fitted to data; the reported numbers are benchmark measurements. The main hand-chosen settings are the maximum chain length and per-root budgets, which are not reported. The architecture depends on behavioral assumptions about instruction-following and on a plain-text handoff format that is not stress-tested. No new entities such as particles or forces are introduced.

free parameters (2)
  • Maximum chain length R = not reported
    Defined in Section 3 as the chain budget. Section 5.1 says it is recorded, but only average root calls (1.8 to 3.4) are shown, so the actual per-benchmark setting is an unstated choice that directly affects accuracy and cost.
  • Per-root iteration and sub-LLM call limits = not reported
    Section 5.1 states these are recorded for a fair comparison, but the values are never given. They control how much work each root can do and therefore affect the reported results.
assumptions (3)
  • domain assumption A later fresh root of the same model will follow the instruction to read, preserve, and update existing artifacts rather than restart extraction from scratch.
    Chain Rules 1 through 6 and Section 4.1 depend on this behavior. The paper itself acknowledges in Section 6 that artifact quality is model-dependent and the chain can drift.
  • domain assumption The plain-text handoff format with exactly three sections can be reliably produced by the model and parsed by the host.
    Section 3.4 specifies the format and forbids structured formats, but no robustness data are provided, and a malformed handoff would break the chain.
  • domain assumption The model's context window can hold the original problem, original context, and chain state at every root.
    Table 2 reports up to 132k input tokens per Chained RLM task, so the architecture assumes the underlying model can accommodate the original context plus summary, blackboard, and artifacts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Chained Recursive Language Models for Multi-Iteration Reasoning." pith.science (2026). https://pith.science/paper/255B6V63

@misc{pith2026260805124,
  author       = {Pith},
  title        = {Pith review of: Chained Recursive Language Models for Multi-Iteration Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/255B6V63}},
  note         = {Machine review of arXiv:2608.05124}
}
read the original abstract

Long context reasoning in large language models (LLMs) is usually constrained by the fact that a single inference trajectory has to simultaneously explore the context, store intermediate state, verify evidence, and produce the final answer. This becomes particularly difficult in tasks that require extraction, counting, ordering, or multi-hop reasoning, where an early mistake can propagate until the final response. In this work, we propose Chained Recursive Language Models (Chained RLM), an inference-time architecture, in which the same underlying model is called repeatedly as a sequence of fresh reasoning roots. Each root receives the original problem and context, but does not inherit the full conversational history. Instead, it receives a compact plain-text summary, a plain-text blackboard, and some durable task-specific artifacts written by predecessor roots. The motivation is to manage the context by chopping into partial tasks rather than one large inference response; in each staged computation, intermediate artifacts can be inspected, corrected, and extended by a later fresh inference by the same model. We describe the system model, handoff mechanism, artifact workspace, and evaluation protocol for this system. We study when fresh-context artifact continuation gives a measurable gain in accuracy over direct LLM answering even with recursive tool-calling.

Figures

Figures reproduced from arXiv: 2608.05124 by the authors.

Figure 1
Figure 1. Chained RLM architecture with 3 RLM calls. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 8 canonical work pages

  1. [1]

    arXiv preprint arXiv:2501.12948 , year=

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

    arXiv

  2. [2]

    arXiv preprint arXiv:2409.12618 , year=

    Iteration of thought: Leveraging inner dialogue for autonomous large language model reasoning , author=. arXiv preprint arXiv:2409.12618 , year=

    arXiv

  3. [3]

    arXiv preprint arXiv:2503.06692 , year=

    Inftythink: Breaking the length limits of long-context reasoning in large language models , author=. arXiv preprint arXiv:2503.06692 , year=

    arXiv

  4. [4]

    arXiv preprint arXiv:2503.19855 , year=

    Think twice: Enhancing llm reasoning by scaling multi-round test-time thinking , author=. arXiv preprint arXiv:2503.19855 , year=

    arXiv

  5. [5]

    arXiv preprint arXiv:2502.10858 , year=

    Is depth all you need? an exploration of iterative reasoning in llms , author=. arXiv preprint arXiv:2502.10858 , year=

    arXiv

  6. [6]

    arXiv preprint arXiv:2504.04736 , year=

    Synthetic data generation & multi-step rl for reasoning & tool use , author=. arXiv preprint arXiv:2504.04736 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2308.10379 , year=

    Algorithm of thoughts: Enhancing exploration of ideas in large language models , author=. arXiv preprint arXiv:2308.10379 , year=

    arXiv

  8. [8]

    arXiv preprint arXiv:2503.09567 , year=

    Towards reasoning era: A survey of long chain-of-thought for reasoning large language models , author=. arXiv preprint arXiv:2503.09567 , year=

    arXiv

Show all 71 references
  1. [9]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    arXiv

  2. [10]

    arXiv preprint arXiv:2501.19393 , year=

    s1: Simple test-time scaling , author=. arXiv preprint arXiv:2501.19393 , year=

    arXiv

  3. [11]

    arXiv preprint arXiv:2307.03172 , year=

    Lost in the middle: How language models use long contexts , author=. arXiv preprint arXiv:2307.03172 , year=

    arXiv

  4. [12]

    arXiv preprint arXiv:2404.07143 , volume=

    Leave no context behind: Efficient infinite context transformers with infini-attention , author=. arXiv preprint arXiv:2404.07143 , volume=

    arXiv

  5. [13]

    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=

  6. [14]

    arXiv preprint arXiv:2502.12962 , year=

    Infinite Retrieval: Attention Enhanced LLMs in Long-Context Processing , author=. arXiv preprint arXiv:2502.12962 , year=

    arXiv

  7. [15]

    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=

  8. [16]

    arXiv preprint arXiv:2505.11827 , year=

    Not All Thoughts are Generated Equal: Efficient LLM Reasoning via Multi-Turn Reinforcement Learning , author=. arXiv preprint arXiv:2505.11827 , year=

    arXiv

  9. [17]

    arXiv preprint arXiv:2406.04692 , year=

    Mixture-of-agents enhances large language model capabilities , author=. arXiv preprint arXiv:2406.04692 , year=

    arXiv

  10. [18]

    arXiv preprint arXiv:2412.21200 , year=

    Distributed Mixture-of-Agents for Edge Inference with Large Language Models , author=. arXiv preprint arXiv:2412.21200 , year=

    arXiv

  11. [19]

    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=

    arXiv

  12. [20]

    arXiv preprint arXiv:2412.16720 , year=

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

    arXiv

  13. [21]

    arXiv preprint arXiv:2312.11805 , year=

    Gemini: a family of highly capable multimodal models , author=. arXiv preprint arXiv:2312.11805 , year=

    arXiv

  14. [22]

    arXiv preprint arXiv:2103.03874 , year=

    Measuring mathematical problem solving with the math dataset , author=. arXiv preprint arXiv:2103.03874 , year=

    arXiv

  15. [23]

    The Twelfth International Conference on Learning Representations , year=

    Let's verify step by step , author=. The Twelfth International Conference on Learning Representations , year=

  16. [24]

    arXiv preprint arXiv:2110.14168 , year=

    Training verifiers to solve math word problems , author=. arXiv preprint arXiv:2110.14168 , year=

    arXiv

  17. [25]

    Granular Format Rewards for Eliciting Mathematical Reasoning Capabilities in Small Language Models , author=

  18. [26]

    Daniel Han, Michael Han and Unsloth team , title =

  19. [27]

    Hugging Face repository , volume=

    Numinamath: The largest public dataset in ai4maths with 860k pairs of competition math problems and solutions , author=. Hugging Face repository , volume=

  20. [28]

    arXiv preprint arXiv:2107.03374 , year=

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

    arXiv

  21. [29]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. ICLR , volume=

  22. [30]

    arXiv preprint arXiv:2411.03284 , year=

    Smoa: Improving multi-agent large language models with sparse mixture-of-agents , author=. arXiv preprint arXiv:2411.03284 , year=

    arXiv

  23. [31]

    arXiv preprint arXiv:2502.00674 , year=

    Rethinking Mixture-of-Agents: Is Mixing Different Large Language Models Beneficial? , author=. arXiv preprint arXiv:2502.00674 , year=

    arXiv

  24. [32]

    arXiv preprint arXiv:2507.19457 , year=

    Gepa: Reflective prompt evolution can outperform reinforcement learning , author=. arXiv preprint arXiv:2507.19457 , year=

    arXiv

  25. [33]

    arXiv preprint arXiv:2505.13975 , year=

    DRP: Distilled Reasoning Pruning with Skill-aware Step Decomposition for Efficient Large Reasoning Models , author=. arXiv preprint arXiv:2505.13975 , year=

    arXiv

  26. [34]

    arXiv preprint arXiv:2510.14901 , year=

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

    arXiv

  27. [35]

    arXiv preprint arXiv:1503.02531 , year=

    Distilling the knowledge in a neural network , author=. arXiv preprint arXiv:1503.02531 , year=

    arXiv

  28. [36]

    The twelfth international conference on learning representations , year=

    On-policy distillation of language models: Learning from self-generated mistakes , author=. The twelfth international conference on learning representations , year=

  29. [37]

    Thinking Machines Lab: Connectionism , year =

    Kevin Lu and Thinking Machines Lab , title =. Thinking Machines Lab: Connectionism , year =

  30. [38]

    Proceedings of the 2024 conference on empirical methods in natural language processing , pages=

    A survey on in-context learning , author=. Proceedings of the 2024 conference on empirical methods in natural language processing , pages=

    2024

  31. [39]

    Incomplete Ideas (blog) , volume=

    The bitter lesson , author=. Incomplete Ideas (blog) , volume=

  32. [40]

    arXiv preprint arXiv:2304.12210 , year=

    A cookbook of self-supervised learning , author=. arXiv preprint arXiv:2304.12210 , year=

    arXiv

  33. [41]

    arXiv preprint arXiv:2509.14745 , year=

    On the use of agentic coding: An empirical study of pull requests on github , author=. arXiv preprint arXiv:2509.14745 , year=

  34. [42]

    arXiv preprint arXiv:2510.12399 , year=

    A Survey of Vibe Coding with Large Language Models , author=. arXiv preprint arXiv:2510.12399 , year=

  35. [43]

    arXiv preprint arXiv:2507.15855 , volume=

    Gemini 2.5 pro capable of winning gold at imo 2025 , author=. arXiv preprint arXiv:2507.15855 , volume=

    2025

  36. [44]

    arXiv preprint arXiv:2412.04604 , year=

    Arc prize 2024: Technical report , author=. arXiv preprint arXiv:2412.04604 , year=

    2024 arXiv

  37. [45]

    arXiv preprint arXiv:2511.16072 , year=

    Early science acceleration experiments with GPT-5 , author=. arXiv preprint arXiv:2511.16072 , year=

  38. [46]

    BioRxiv , pages=

    Scaling large language models for next-generation single-cell analysis , author=. BioRxiv , pages=. 2025 , publisher=

    2025

  39. [47]

    arXiv preprint arXiv:2504.13837 , year=

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

    arXiv

  40. [48]

    2nd AI for Math Workshop@ ICML 2025 , year=

    The invisible leash: Why rlvr may not escape its origin , author=. 2nd AI for Math Workshop@ ICML 2025 , year=

    2025

  41. [49]

    Ieee Access , volume=

    Contrastive representation learning: A framework and review , author=. Ieee Access , volume=. 2020 , publisher=

    2020

  42. [50]

    Nature , volume=

    AI models collapse when trained on recursively generated data , author=. Nature , volume=. 2024 , publisher=

    2024

  43. [51]

    Advances in Neural Information Processing Systems , volume=

    Star: Bootstrapping reasoning with reasoning , author=. Advances in Neural Information Processing Systems , volume=

  44. [52]

    arXiv preprint arXiv:2504.18116 , year=

    Think, prune, train, improve: Scaling reasoning without scaling models , author=. arXiv preprint arXiv:2504.18116 , year=

    arXiv

  45. [53]

    arXiv preprint arXiv:2502.03860 , year=

    Bolt: Bootstrap long chain-of-thought in language models without distillation , author=. arXiv preprint arXiv:2502.03860 , year=

    arXiv

  46. [54]

    arXiv preprint arXiv:2503.20783 , year=

    Understanding r1-zero-like training: A critical perspective , author=. arXiv preprint arXiv:2503.20783 , year=

    arXiv

  47. [55]

    arXiv preprint arXiv:2503.14476 , year=

    Dapo: An open-source llm reinforcement learning system at scale , author=. arXiv preprint arXiv:2503.14476 , year=

    arXiv

  48. [56]

    arXiv preprint arXiv:2507.18071 , year=

    Group sequence policy optimization , author=. arXiv preprint arXiv:2507.18071 , year=

    arXiv

  49. [57]

    arXiv preprint arXiv:2511.07919 , year=

    Feedback Descent: Open-Ended Text Optimization via Pairwise Comparison , author=. arXiv preprint arXiv:2511.07919 , year=

  50. [58]

    arXiv preprint arXiv:2210.03629 , year=

    React: Synergizing reasoning and acting in language models , author=. arXiv preprint arXiv:2210.03629 , year=

    arXiv

  51. [59]

    , author=

    MemGPT: towards LLMs as operating systems. , author=. 2023 , publisher=

    2023

  52. [60]

    Advances in neural information processing systems , volume=

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

  53. [61]

    arXiv preprint arXiv:2310.03714 , year=

    Dspy: Compiling declarative language model calls into self-improving pipelines , author=. arXiv preprint arXiv:2310.03714 , year=

    arXiv

  54. [62]

    arXiv preprint arXiv:2512.24601 , year=

    Recursive language models , author=. arXiv preprint arXiv:2512.24601 , year=

    arXiv

  55. [63]

    arXiv preprint arXiv:2404.06654 , year=

    RULER: What's the real context size of your long-context language models? , author=. arXiv preprint arXiv:2404.06654 , year=

    arXiv

  56. [64]

    Advances in Neural Information Processing Systems , volume=

    Babilong: Testing the limits of llms with long context reasoning-in-a-haystack , author=. Advances in Neural Information Processing Systems , volume=

  57. [65]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Longbench v2: Towards deeper understanding and reasoning on realistic long-context multitasks , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  58. [66]

    arXiv preprint arXiv:2511.02817 , year=

    Oolong: Evaluating long context reasoning and aggregation capabilities , author=. arXiv preprint arXiv:2511.02817 , year=

  59. [67]

    arXiv preprint arXiv:2203.11171 , year=

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

    arXiv

  60. [68]

    Journal of Artificial Intelligence Research , volume=

    Agentic large language models, a survey , author=. Journal of Artificial Intelligence Research , volume=

  61. [69]

    URL https://research

    Context rot: How increasing input tokens impacts llm performance , author=. URL https://research. trychroma. com/context-rot, retrieved October , volume=

  62. [70]

    Advances in neural information processing systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in neural information processing systems , volume=

  63. [71]

    Mitra, Purbesh and Ulukus, Sennur , journal=

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.