Pith. sign in

REVIEW 3 major objections 5 minor 25 references

A one-step distillation of an agent's own trial-and-error history into model weights keeps 64.8–93.4% of the in-context learning gain, with no extra environment calls.

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

2026-08-01 08:35 UTC pith:EEKW44A4

load-bearing objection Experience Distillation is a worthwhile applied method with an honest empirical core, but the TaleSuite sample-efficiency claim in Fig. 1 is inflated by a unit mix and the paper currently ships no code or data. the 3 major comments →

arxiv 2607.21051 v1 pith:EEKW44A4 submitted 2026-07-23 cs.CL

Sample-Efficient Learning from Agent Experience

Chenhui Gou , Haoqin Tu , Yunhao Fang , Jianfei Cai , Hamid Rezatofighi This is my paper
classification cs.CL
keywords Experience Distillationcontext distillationin-context learningagent experiencesample efficiencysoftware engineering agentstext-adventure gamesteacher-sampled distillation
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 proposes Experience Distillation, a way to transfer the benefit an agent gets from seeing its own past trial-and-error history into the model's weights, without spending any additional environment interactions. It does this by replaying recorded histories and asking an experience-conditioned teacher to generate only the next decision at each point; those teacher decisions then serve as supervised targets. On 749 software-engineering tasks and six text-adventure games, the trained student keeps 64.8% and 93.4% of the in-context learning gain, while direct supervised fine-tuning on the same experience retains only 3.8% and actually loses ground on the text games. The paper argues this turns in-context learning from a temporary inference-time effect into a persistent capability, and shows the pipeline matches classical RL baselines with at least 9.6× fewer environment samples.

Core claim

The central claim is that trajectory-level context distillation can be made sample-efficient by reducing each branch to a single teacher decision sampled at a recorded history, eliminating both the world model and further environment rollouts. The one-step branch rollout yields a teacher-sampled forward-KL objective: minimize the negative log-likelihood of the teacher's next action conditioned only on the history prefix. With branch packing and an enhanced-reasoning prompt, this objective lets a student reproduce most of the behavior that ICL elicits, even though the experience context is absent at evaluation. The paper's headline results are 64.8% ICL-gain retention (51.4% pass@1 vs 5.3% ze

What carries the argument

The central object is the one-step branched rollout: at each recorded history prefix, the experience-conditioned teacher samples a single next action and that action becomes the student's training label at that prefix. This turns context distillation into a teacher-sampled next-token prediction objective and avoids both additional environment interaction and world-model rollout error. Branch packing concatenates successive teacher decisions with recorded observations into one sequence, and an enhanced-reasoning prompt makes the teacher's decisions more informative. The key contrast is teacher-sampled forward KL (used by default) versus student-sampled reverse KL, which the paper finds stays

Load-bearing premise

Everything rests on the teacher's next action at a recorded history being a reliable training signal even though the environment never verifies it; the paper's own case study shows the teacher sometimes repeats hypotheses that the accumulated trials already rejected.

What would settle it

Take a fixed experience corpus, replace the teacher-sampled next actions with actions sampled from the zeroshot base model at the same histories, and distill. If ICL-gain retention stays near 64.8%, teacher decision quality is not what carries the method; if it falls to SFT levels, the teacher-sampled signal is load-bearing.

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

If this is right

  • If the results hold, an agent that learns by trial and error can consolidate that learning into weights and later act without carrying the history in its context window.
  • The environment-sample budget for this form of learning is set by the ICL experience-collection stage; distillation itself adds zero samples, so the 9.6× reduction over RL is a direct property of the method.
  • Direct SFT on raw collected trajectories is not a workable substitute: it recovers only 3.8% of the ICL gain on SWE and none on TaleSuite.
  • Distilled multi-task experience transfers to out-of-distribution tasks and accumulates over repeated collect-and-distill cycles, so capabilities built this way can grow without retaining old contexts.
  • Teacher-sampled decisions are the effective supervision; reversing the KL to student-sampled trajectories leaves performance near zeroshot.

Where Pith is reading between the lines

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

  • A practical testable extension: filter or reweight teacher decisions by later trial outcomes (e.g., whether the action appears in the accepted trajectory) could remove the stale rejected hypotheses the paper documents and push retention above 64.8%.
  • The one-step branch recipe is not tied to ICL; it could distill decisions from any context-conditioned teacher, such as a verifier or human feedback at a specific turn, as long as the teacher can be queried at recorded states.
  • If this scales, the scarce resource in agent learning shifts from environment samples to the diversity and informativeness of the collected experience, since distillation is cheap once a rich multi-trial record exists.
  • The reported sample-efficiency gap may understate the method's advantage if RL baselines are charged with only their best-checkpoint budget; a fairer comparison would match compute and report the full training curve.

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. The paper introduces Experience Distillation, a context-distillation method for agent policies that converts experience-conditioned in-context learning (ICL) gains into model weights without additional environment interaction. The implementation samples one-step teacher decisions at recorded history points, packs multiple branches into a single sequence, and trains the student with next-token prediction on the teacher-sampled decisions. Experiments on 749 curated software-engineering tasks and six TaleSuite text-adventure tasks report that Experience Distillation retains 64.8% and 93.4% of the ICL gain, respectively, while direct SFT retains only 3.8% and -2.6%. The paper further claims that ICL followed by Experience Distillation matches or exceeds PPO/GRPO baselines with at least 9.6x fewer environment samples, and includes extensive ablations, controlled diagnostics, and qualitative case studies.

Significance. If the reported results hold, the core idea is valuable: it offers a practical way to consolidate sample-efficient ICL improvements into weights while avoiding both additional environment rollouts and learned-world-model error compounding. The paper is unusually thorough in its ablations: it compares the one-step model-free objective with model-based rollouts, analyzes branch packing, studies teacher-sampled versus student-sampled objectives, provides task-level GRPO curves, includes a controlled ColorButton diagnostic, and documents qualitative evidence chains for eight SWE tasks. The explicit discussion of limitations in Appendix C.11 is a strength. However, the headline sample-efficiency claim contains an arithmetic inconsistency in the TaleSuite ratio, and the SWE retention numbers are conditional on a success-filtered experience set. These issues do not destroy the contribution, but they require correction and qualification before the central claims can be accepted as stated.

major comments (3)
  1. [Section 4.3 / Figure 1 / Appendix B.1] The TaleSuite label '57.2x fewer samples' is inconsistent with the figure's own axis definition. Appendix B.1 reports the GRPO run as 448 total environment samples for six tasks (74.7 samples/task), while Table 10 gives 47 total experience trials (7.8 trials/task) for the EPD pipeline. The printed ratio 57.2 is computed as 448 / (47/6), i.e., total GRPO samples divided by per-task EPD samples. With consistent per-task units, the ratio is (448/6)/(47/6) = 9.5. Thus the 57.2x claim is unsupported by the paper's own numbers. This is load-bearing because the abstract promises 'at least 9.6x fewer environment samples'; the corrected TaleSuite ratio is just below 9.6. Please revise the figure label, caption, and abstract, and report exact per-task budgets for both pipelines.
  2. [Section 3.2, Eq. (2) / Appendix C.3.3] The distillation objective maximizes the likelihood of teacher-sampled decisions a'_t ~ pi_{theta0}(.|h_t^exp, tau_exp) without any environment verification. Appendix C.3.3 explicitly shows that the teacher can regenerate decisions that the accumulated trials already rejected (MapStore2 Targets 2 and 3). Consequently, the training signal is not guaranteed to consist of 'improved decisions' from ICL, despite the method's rationale. This does not by itself invalidate the empirical retention in Table 1, but it weakens the conceptual claim. The paper should quantify, at least in the case studies, how many teacher targets correspond to accepted repair components versus rejected hypotheses, or add an ablation that filters or down-weights rejected-hypothesis targets.
  3. [Appendix A.7 / Table 1] The curated SWE experience retains only trajectories that reach an accepted commit after more than one trial. The reported ICL reference (76.4% pass@1) and the EPD retention (64.8% G_ICL) are therefore conditional on a success-filtered task set. The paper does not state how many of the originally considered tasks were excluded by this criterion. This selection is favorable to the method and limits the external validity of the headline retention number. Please report the number of excluded tasks and, if possible, include results on the unfiltered task set or discuss the expected effect of including tasks without successful experience.
minor comments (5)
  1. [Section 4.3] The EPD sample budgets are not stated directly in the main text. For SWE, the implied budget is 504.9/9.6 ≈ 52.6 environment samples per task; for TaleSuite it is 47/6 ≈ 7.8 trials per task. These exact numbers should be reported alongside Figure 1 so the reader can reproduce the ratios.
  2. [Figure 1] The x-axis uses a log(1+x) transform, but the visible axis starts at 10. Clarify how the zeroshot and ICL+SFT points, which appear to have zero environment samples, are plotted.
  3. [Table 1] The SFT G_ICL value on TaleSuite is negative (-2.6%). The definition of G_ICL in Section 4.1 should state whether negative values are allowed or clipped at zero.
  4. [Appendix A.7] Report the distribution of n in {8,...,12} for the curated SWE experience collection, and how many tasks were excluded because no retained trajectory reached an accepted commit after more than one trial.
  5. [Section 3.3, Eq. (4)] The relationship between Eq. (2) and the packed objective Eq. (4) is acknowledged as an approximation. Table 3 shows comparable performance, but a short quantitative statement of the divergence (e.g., teacher-sequence likelihood or per-position KL) would make the approximation more transparent.

Circularity Check

0 steps flagged

No significant circularity: the distillation objective is an empirical supervised target, and the headline numbers are measured against external ICL, SFT, and RL baselines.

full rationale

Experience Distillation's core equations (Eq. 2 and Eq. 4) define a teacher-sampled supervised objective: sample a'_t from the experience-conditioned base model at recorded histories and train the student to maximize log pi_theta(a'_t | h_t^exp). This is a distillation target, not a fitted parameter renamed as a prediction; whether the student retains ICL gains is measured against the external ICL reference and zeroshot, and could in principle be low, as SFT's 3.8% result shows. The G_ICL normalization is a reporting convention, not a circular definition. Comparisons with PPO and GRPO use independent RL baselines, and no fitted constant enters Eqs. 2-4. No load-bearing self-citation or imported uniqueness theorem appears. The TaleSuite '57.2x' label in Fig. 1 is an arithmetic unit inconsistency (total GRPO samples divided by per-task EPD samples rather than both per-task), which is a correctness concern, not circularity. The acknowledged unverified teacher labels in Appendix C.3.3 are a data-quality limitation, not a circular step. Overall, the derivation is self-contained and the central claims rest on empirical comparisons rather than on definitional equivalences.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central method introduces no unobserved entity; its commitments are the reliability of teacher decisions and the approximation in branch packing.

free parameters (3)
  • branch-packed sequences per experience trial = 16 (main experiments)
    Figure 3 shows G_ICL rises from 31.8% at 1 sequence to 64.8% at 16; the headline result uses the top of this scaling curve, so the claimed retention depends on a chosen compute budget not fixed by the method.
  • SWE experience-collection budget and retention filter = n=8–12 processes × up to k=10 trials; retain trajectory only if it reaches an accepted commit after more than one trial
    Appendix A.7: this hand-chosen collection rule defines the experience corpus and excludes tasks where ICL never succeeds; it directly affects ICL reference and all downstream numbers.
  • enhanced teacher reasoning prompt I = fixed prompt, not quoted in full
    Section 3.3 and Appendix A.1: added after empirical observation that longer reasoning improves post-distillation performance; its exact wording is an unstated implementation detail.
axioms (4)
  • domain assumption The teacher's one-step decisions at recorded histories are good enough to serve as training labels without environment verification.
    Eq. 2/4 trains on a'_t sampled from π_{θ0}; no environment check confirms correctness. Qualitative appendix examples show some teacher targets revisit rejected hypotheses (C.3.3).
  • domain assumption The base model exhibits strong in-context learning from the collected experience, so the ICL reference is a meaningful ceiling.
    Table 1 reports ICL 76.4% pass@1 and 45.6 normalized score; this is an empirical property of the in-house base models, not a guaranteed property of LLMs.
  • domain assumption Branch packing (Eq. 4) approximates the exact one-step objective (Eq. 2) well enough that conclusions transfer.
    Section 3.3 admits Eq. 4 is 'not an exact implementation' because earlier teacher decisions are in context; Table 3 shows comparable scores but only on six TaleSuite tasks.
  • standard math The KL chain-rule factorization of trajectory distributions is valid under the POMDP formulation.
    Section 2 defines trajectory distribution; Eq. 1 uses chain rule for KL; standard under the stated factorization.

pith-pipeline@v1.3.0-alltime-deepseek · 37461 in / 14228 out tokens · 158726 ms · 2026-08-01T08:35:45.635652+00:00 · methodology

0 comments
read the original abstract

Real-world agent learning is often constrained by costly environment interactions, such as running time-consuming experiments or obtaining human feedback. In-context learning offers a highly sample-efficient way for agents to learn from their own interaction histories, but its gains disappear once that experience is removed from the context. Separately, context distillation provides a mechanism for internalizing contextual information into model weights. However, applying it to agents' interaction histories without sacrificing environment sample efficiency remains underexplored. We term this problem Experience Distillation and develop an implementation that requires no further environment interaction beyond the collected experience. Experiments on 749 curated software-engineering tasks and six text-adventure games show that it retains at least 64.8\% of the gains from in-context learning across both domains, whereas direct supervised fine-tuning on the collected experience recovers only 3.8\%. Compared with classical reinforcement-learning baselines, in-context learning from trial-and-error experience followed by Experience Distillation matches their performance with at least \(9.6\times\) fewer environment samples.

Figures

Figures reproduced from arXiv: 2607.21051 by Chenhui Gou, Hamid Rezatofighi, Haoqin Tu, Jianfei Cai, Yunhao Fang.

Figure 1
Figure 1. Figure 1: Environment-sample efficiency of ICL followed by Experience Distillation (ICL + EPD) [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Continual Experience Distillation on six TaleSuite tasks. Cycle 0 is the initial checkpoint; [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Teacher-generated data scaling for Experience Distillation on 749 curated SWE tasks. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of learning rate and training duration on Experience Distillation (EPD). Crosses [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: GRPO learning curve on six TaleSuite tasks. Left: mean normalized score by GRPO [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Task-level GRPO learning curves for six TaleSuite tasks. Every marker is an observed nor [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: ColorButton action space and two observed GRPO runs. Left: three decisions define 18 [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: PPO learning curve on 749 curated SWE tasks. Each point reports pass@1 averaged over [PITH_FULL_IMAGE:figures/full_fig_p029_8.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

25 extracted references · 1 canonical work pages

  1. [1]

    Every available checkpoint is shown for each task; missing checkpoints remain unobserved rather than being imputed

    To determine whether the aggregate improvement is shared across tasks, Figure 6 disaggregates the same GRPO run into task-level learning curves. Every available checkpoint is shown for each task; missing checkpoints remain unobserved rather than being imputed. 23 1 10 20 30 0 25 50 75 100 (a) Balances 1 10 20 30 0 25 50 75 100 (b) Detective 1 10 20 30 0 2...

    2024

  2. [7]

    Priyanshu Gupta, Shashank Kirtania, Ananya Singha, Sumit Gulwani, Arjun Radhakrishna, Gustavo Soares, and Sherry Shi

    doi: 10.1038/s41586-025-09422-z. Priyanshu Gupta, Shashank Kirtania, Ananya Singha, Sumit Gulwani, Arjun Radhakrishna, Gustavo Soares, and Sherry Shi. MetaReflection: Learning instructions for language agents using past reflections. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp. 8369–8385,

    doi:10.1038/s41586-025-09422-z 2024

  3. [8]

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

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531,

    Pith/arXiv arXiv

  4. [9]

    Distilling step-by-step! outperform- ing larger language models with less training data and smaller model sizes.arXiv preprint arXiv:2305.02301,

    Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperform- ing larger language models with less training data and smaller model sizes.arXiv preprint arXiv:2305.02301,

    Pith/arXiv arXiv

  5. [12]

    Offline reinforcement learning: Tuto- rial, review, and perspectives on open problems.arXiv preprint arXiv:2005.01643,

    Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline reinforcement learning: Tuto- rial, review, and perspectives on open problems.arXiv preprint arXiv:2005.01643,

    Pith/arXiv arXiv 2005

  6. [13]

    Bodhisattwa Prasad Majumder, Bhavana Dalvi Mishra, Peter Jansen, Oyvind Tafjord, Niket Tandon, Li Zhang, Chris Callison-Burch, and Peter Clark

    doi: 10.64434/tml.20251026.https://thinkingmachines.ai/blog/ on-policy-distillation. Bodhisattwa Prasad Majumder, Bhavana Dalvi Mishra, Peter Jansen, Oyvind Tafjord, Niket Tandon, Li Zhang, Chris Callison-Burch, and Peter Clark. CLIN: A continually learning language agent for rapid task adaptation and generalization. InConference on Language Modeling,

    work page doi:10.64434/tml.20251026.https://thinkingmachines.ai/blog/

  7. [15]

    In-context learning and induction heads.arXiv preprint arXiv:2209.11895,

    Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. In-context learning and induction heads.arXiv preprint arXiv:2209.11895,

    Pith/arXiv arXiv

  8. [16]

    Privileged information distillation for language models.arXiv preprint arXiv:2602.04942,

    Emiliano Penaloza, Dheeraj Vattikonda, Nicolas Gontier, Alexandre Lacoste, Laurent Charlin, and Massimo Caccia. Privileged information distillation for language models.arXiv preprint arXiv:2602.04942,

    Pith/arXiv arXiv

  9. [17]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    Pith/arXiv arXiv

  10. [18]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. DeepSeekMath: Pushing the limits of mathe- matical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Pith/arXiv arXiv

  11. [19]

    Charlie Snell, Dan Klein, and Ruiqi Zhong

    doi: 10.1038/nature16961. Charlie Snell, Dan Klein, and Ruiqi Zhong. Learning by distilling context.arXiv preprint arXiv:2209.15189,

    Pith/arXiv arXiv

  12. [20]

    Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents.arXiv preprint arXiv:2604.10674,

    Hao Wang, Guozhi Wang, Han Xiao, Yufeng Zhou, Yue Pan, Jichao Wang, Ke Xu, Yafei Wen, Xiaohu Ruan, Xiaoxin Chen, and Honggang Qi. Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents.arXiv preprint arXiv:2604.10674,

    Pith/arXiv arXiv

  13. [21]

    Wang, Zeb Kurth-Nelson, Dhruva Tirumala, Hubert Soyer, Joel Z

    Jane X. Wang, Zeb Kurth-Nelson, Dhruva Tirumala, Hubert Soyer, Joel Z. Leibo, Remi Munos, Charles Blundell, Dharshan Kumaran, and Matt Botvinick. Learning to reinforcement learn.arXiv preprint arXiv:1611.05763,

    Pith/arXiv arXiv

  14. [22]

    Online experiential learning for language models.arXiv preprint arXiv:2603.16856, 2026a

    Tianzhu Ye, Li Dong, Qingxiu Dong, Xun Wu, Shaohan Huang, and Furu Wei. Online experiential learning for language models.arXiv preprint arXiv:2603.16856, 2026a. Tianzhu Ye, Li Dong, Xun Wu, Shaohan Huang, and Furu Wei. On-policy context distillation for language models.arXiv preprint arXiv:2602.12275, 2026b. Tianhe Yu, Garrett Thomas, Lantao Yu, Stefano E...

    Pith/arXiv arXiv

  15. [24]

    EdgeBench: Unveiling scaling laws of learning from real-world environments

    Deyao Zhu et al. EdgeBench: Unveiling scaling laws of learning from real-world environments. arXiv preprint arXiv:2607.05155,

    Pith/arXiv arXiv

  16. [2016]

    In-context learning distillation for efficient few-shot fine-tuning.arXiv preprint arXiv:2412.13243,

    Yifei Duan, Liu Li, Zirui Zhai, and Jinxia Yao. In-context learning distillation for efficient few-shot fine-tuning.arXiv preprint arXiv:2412.13243,

    Pith/arXiv arXiv

  17. [2017]

    A general language assistant as a labora- tory for alignment.arXiv preprint arXiv:2112.00861,

    Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones, Nicholas Joseph, Ben Mann, Nova DasSarma, Nelson Elhage, Zac Hatfield-Dodds, Danny Her- nandez, Jackson Kernion, Kamal Ndousse, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, and Jared Kaplan. A general language assistant as a labo...

    Pith/arXiv arXiv

  18. [2018]

    TALES: Text adventure learning environment suite.arXiv preprint arXiv:2504.14128,

    Christopher Zhang Cui, Xingdi Yuan, Ziang Xiao, Prithviraj Ammanabrolu, and Marc-Alexandre Côté. TALES: Text adventure learning environment suite.arXiv preprint arXiv:2504.14128,

    Pith/arXiv arXiv

  19. [2020]

    Doc-to-LoRA: Learning to instantly internalize contexts.arXiv preprint arXiv:2602.15902,

    Rujikorn Charakorn, Edoardo Cetin, Shinnosuke Uesaka, and Robert Tjarko Lange. Doc-to-LoRA: Learning to instantly internalize contexts.arXiv preprint arXiv:2602.15902,

    arXiv

  20. [2021]

    Brown et al

    Tom B. Brown et al. Language models are few-shot learners. InAdvances in Neural Information Processing Systems, volume 33, pp. 1877–1901,

    1901

  21. [2022]

    Reinforcement learning via self-distillation.arXiv preprint arXiv:2601.20802,

    Jonas Hübotter et al. Reinforcement learning via self-distillation.arXiv preprint arXiv:2601.20802,

    Pith/arXiv arXiv

  22. [2023]

    In-context learning distilla- tion: Transferring few-shot learning ability of pre-trained language models.arXiv preprint arXiv:2212.10670,

    Yukun Huang, Yanda Chen, Zhou Yu, and Kathleen McKeown. In-context learning distilla- tion: Transferring few-shot learning ability of pre-trained language models.arXiv preprint arXiv:2212.10670,

    Pith/arXiv arXiv

  23. [2024]

    Context distillation as latent memory management.arXiv preprint arXiv:2605.28889,

    Ziyang Zheng, Zeju Li, Xiangyu Wen, Jianyuan Zhong, Junhua Huang, Lei Chen, Mingxuan Yuan, and Qiang Xu. Context distillation as latent memory management.arXiv preprint arXiv:2605.28889,

    Pith/arXiv arXiv

  24. [2025]

    Bartlett, Ilya Sutskever, and Pieter Abbeel

    Yan Duan, John Schulman, Xi Chen, Peter L. Bartlett, Ilya Sutskever, and Pieter Abbeel. RL2: Fast reinforcement learning via slow reinforcement learning.arXiv preprint arXiv:1611.02779,

    Pith/arXiv arXiv

  25. [2026]

    Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pp

    Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pp. 11048–11064,

    2022