Pith. sign in

REVIEW 3 major objections 6 minor 17 references

MERA: Model Evolution and Routing with Skill Adaptation for Agentic Systems at Scale

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Retraining small coding agents beats routing: 28.7% to 49.7% pass

desk verdict A genuinely novel invoke-level adaptation loop with honest, mostly defensible code-generation evidence, but the headline 28.7→49.7 claim likely confounds weight evolution with the SkillBook prompt, so the central capability claim is not yet proven. read the letter →

arxiv 2608.10333 v1 pith:DE7DTXW5 submitted 2026-08-11 cs.LG

Yuhang Yao , Zeyu Wang , Wanyi Chen , Tongyun Yang , Yuhang Han , Jie Xiao , Chengke Bao , Tianyi Zhao
show 3 more authors
Lynn Ai Eric Yang Tianyu Shi
This is my paper · ORCID
classification cs.LG
keywords MERAmodelevolutioninvocation-leveladaptationverifier-backedfallbackSkillBookmulti-cycledistillationagentroutingGRPO
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

MERA is an attempt to prove that the small model inside an agent system can be made more capable from its own execution traces, rather than merely being routed around. The paper proposes a closed loop in which failed invocations are replayed to produce execution-verified teacher demonstrations, recurring procedures are condensed into a SkillBook, and a small-model adapter is fine-tuned with SFT and optional GRPO; routing and verifier-backed fallback then decide which invocations can be served cheaply. The claim is that four such cycles raise a 1.5B coding model from 28.7% to 49.7% pass on held-out tasks, and that the deployed system preserves near-large-model quality at reduced cost. A sympathetic reader would care because the result would mean cost savings in agent deployment do not have to stop at the boundary of what the small model already knows.

What carries the argument

The load-bearing mechanism is the verifier-backed joint replay gate operating on canonical step slices. Each invocation is canonicalized into a slice holding the prompt, local context, tool schema, output, verifier result, retry and fallback metadata, and skill assignment. Three tracks consume those slices: SkillBook aggregates stable prompt signatures into procedural prompt memory, the LLM adapter trains on hard examples with SFT or GRPO, and an input-only router learns cheap-versus-strong labels from executable outcomes. Updates run in Skill→LLM→Router order so the adapter sees the current SkillBook procedure and the router sees current-cycle student outcomes, and nothing is admitted unless replay shows the combined policy preserves verifier-passing quality. Routing is deliberately input-only and weak by design; safety comes from the verifier fallback path.

What would settle it

Run the admitted final-cycle state on a fresh task set whose ground truth includes hidden properties not covered by the provided executable tests. If verifier-passing student outputs fail those hidden properties at a materially higher rate after adaptation than before, the verifier has missed a semantic failure mode and the replay admission gate has overestimated the safety of the promoted state.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that treating a single model invocation, rather than a whole task, as the unit of adaptation turns ordinary runtime traces into reusable supervision. Failed invocations are replayed against a verifier to obtain teacher outputs; recurring local structure is promoted into SkillBook entries; and the student LoRA adapter is trained on hard slices, with the router fit last so its labels reflect the current-cycle student. Joint replay then admits a candidate runtime state only if end-to-end quality is preserved. The quantitative claim, averaged over three seeds on 582 held-out HumanEval+MBPP tasks, is that the base Qwen2.5-Coder-1.5B improves from 28.7% to 44.2% under SFT and to 49.7% with matched SFT+GRPO, and that with verifier-backed fallback the deployed policy keeps 88.3% pass at 60.8% of always-Luna cost. The paper is explicit that the learned pre-router is weak; most deployed quality preservation comes from verification and fallback, not from routing accuracy.

Load-bearing premise

The loop hinges on the verifier catching every failure mode that matters and on replay of logged invocation slices predicting online behavior; if either fails, down-routing or skill promotion can be admitted while being unsafe in deployment.

Editorial extensions

If this is right

  • Cost savings from routing are no longer capped by the student's current ability, because the student itself improves each cycle from its own failed invocations.
  • Cheap execution can expand over time: the same verifier-gated replay that admits a skill or adapter also identifies new invocation slices that can be safely down-routed.
  • End-task quality can be held near the large-model level even when the learned router is inaccurate, as long as verification and fallback cover the failures.
  • On tool-use benchmarks such as TAU-2, an adapted 2B agent can reach the level of an unadapted 4B endpoint, suggesting adaptation can substitute for some model size.
  • Matched GRPO exceeds matched SFT by 5.5–6.7 points on the code-generation split, attributing part of the gain to reinforcement learning on top of distillation rather than to SFT alone.

Reading between the lines

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

  • The admission gate's safety is bounded by verifier coverage, so the method would become stronger if verifiers extended beyond executable assertions to semantic or hidden-test checks; nothing in the loop prevents such an extension.
  • Because replay cannot capture the distribution shift caused by changing the runtime policy, the paper's own staging rule implies production deployments still need online canary validation; the replay gate is a necessary but not sufficient condition for safe promotion.
  • The Skill→LLM→Router schedule suggests cycle cadence and staleness are tunable: if traces are replayed with stale skills, the next cycle can absorb residual error, so one could test how much overlap across cycles is needed.
  • The same invocation-granularity loop might transfer to agent domains with cheap checkable verifiers, such as structured data extraction or API-call validation, though the paper only demonstrates code generation and one tool-use benchmark.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces MERA, a multi-cycle adaptation protocol for deployed LLM agents. The method treats a single model invocation as the unit of adaptation: online traces are replayed to construct execution-verified teacher demonstrations, a SkillBook of reusable prompt templates is updated, a small-model LoRA adapter is fine-tuned with SFT or SFT+GRPO, and an input-only router is trained on executable labels; a joint replay gate admits only runtime states that preserve end-task quality with verifier-backed fallback. The main empirical claims are that four cycles raise Qwen2.5-Coder-1.5B direct pass from 28.7% to 44.2% (SFT) and 49.7% (SFT+GRPO) on 582 held-out HumanEval+MBPP tasks, that with verifier fallback the deployed policy retains 88.3% pass at 60.8% of always-Luna cost, and that a fine-tuned Qwen3.5-2B improves from 14/35 to 18/35 on TAU-2. The paper also gives a finance break-even planning calculation and an explicit limitations section.

Significance. If the small-model capability claim is established, the paper makes a useful systems contribution: it shifts the routing literature from fixed-student cost savings to verifier-backed multi-cycle model improvement, and it is careful to separate direct model quality from cascade quality. Strengths of the present version include the use of executable, benchmark-provided tests rather than self-generated tests; a held-out split with disjoint task identifiers; three training seeds with cached and matched teacher outputs; explicit reporting of the underpowered TAU-2 comparison; and public code and data. However, the central evidence for capability improvement is currently undermined by the SkillBook confound in the headline table, so the significance claim is conditional on the matched-control fix.

major comments (3)
  1. [§3.3, §3.4, Table 1] The headline direct-pass gain (Table 1: 28.7% → 44.2% / 49.7%) is not a matched comparison of the base model and the fine-tuned model. §3.3 states that 'The LLM adapter is trained and queried with the SkillBook procedure prepended to its prompt,' and §3.4 describes SkillBook entries that prepend static task-format instructions, bounded successful exemplars, and pitfalls to a new task. Table 1's note says only that Direct SLM pass disables routing and fallback; it does not say that SkillBook is disabled. Consequently the trained rows appear to measure the adapter plus a few-shot skill prompt, while the Base SLM row measures the bare model, so the 21-point claimed lift conflates weight-level model evolution with prompt-level skill adaptation. The paper should report a base row with the same SkillBook procedure prepended, and ideally an adapter-only row with SkillBook removed, under the same 1:10 cost accounting; alternatively it should explicitly state that SkillBook was disabled in the Direct-SLM evaluation and confirm this in the protocol description.
  2. [§4.4, Table 3] The TAU-2 result cannot substitute for the missing matched code-generation control. Table 3 is a clean adapter-only comparison (SkillBook, routing, and fallback disabled), but it is explicitly underpowered: 14/35 → 18/35 with one-sided McNemar p=0.171875, a statement the text correctly repeats. With only 35 tasks and 10 discordant pairs, the four-task improvement is fully consistent with chance. The conclusion that 'MERA improves small-model capability rather than only routing around a fixed student' therefore rests entirely on the HumanEval+MBPP direct-pass numbers, which are subject to the SkillBook confound in the previous comment. Once the code-generation comparison is made matched, or an adapter-only row is added, the paper should state precisely which evidence supports the central claim.
  3. [§3.5, §6.2] The joint replay admission rule is applied to the held-out replay set, which creates a mild selection-on-evaluation risk for the cost-quality claims. §3.5 says an update is promoted only if replay preserves quality while reducing cost or fallback risk, and §6.2 reports 'the aggregate ratio over the held-out replay set.' The paper states that threshold calibration uses a disjoint shard and that policy results are reported on held-out task identifiers, but it does not state that the admission decisions themselves (which cycle to report, which artifacts to promote) were made without consulting the reported held-out set. If final-cycle selection used the same 582 tasks, the 88.3% pass / 60.8% cost operating point could be optimistically selected. Please specify that admission decisions were fixed on a separate validation split or describe the selection protocol.
minor comments (6)
  1. [§4.1, Table 2] The 60.8% cost figure and the 88.3% pass point depend on the assumed small:large cost ratio of 1:10; please add a brief sensitivity check for this ratio or state more prominently that the headline cost number is conditional on a single assumed ratio.
  2. [Table 4] The column header 'Train rows' is unclear; please rename it to 'Training set size' or similar, and state whether the 'Deploy cost ($/hr)' column uses the same serving setup as Table 2.
  3. [§4.2] The claim that matched GRPO exceeds SFT by 5.5–6.7 points rests on 95% paired-t intervals computed from three seeds; with n=3 the paired-t procedure is highly sensitive to distributional assumptions, so please report per-seed values or use a nonparametric summary.
  4. [§6.4] The phrase 'non-thinking templates' is undefined; please define what templates are used and why they matter for the TAU-2 comparison.
  5. [Table 2 caption] The statement that RouteLLM-/FrugalGPT-style rows use 'matched adaptations rather than official checkpoints' should be expanded; as written it is unclear what was adapted and whether the comparison is apples-to-apples with the published methods.
  6. [Title page] The PDF text contains typesetting artifacts in the author list (for example, 'Y uhang Y ao'), which should be fixed in the final version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the reported gains are external benchmark measurements with a disjoint calibration shard; the SkillBook prompt mismatch is a validity concern, not a derivation-from-input circularity.

full rationale

The central claim is an empirical measurement on held-out HumanEval+MBPP tasks with three seeds, matched cached teacher outputs, and a router threshold calibrated on a disjoint shard. No fitted parameter is substituted for the reported pass rate, and no equation is defined in terms of the claimed outcome. The routing cost-quality point is measured by replay under a verifier, not derived from in-sample training objectives. TAU-2 is explicitly reported as underpowered (one-sided McNemar p=0.171875), so the paper does not overclaim a statistically forced result. All citations are to external prior work; there is no load-bearing self-citation or imported uniqueness theorem. The paper's own Section 7 openly limits verifier coverage and replay fidelity, which further supports that the protocol is presented as an empirical system rather than a self-referential derivation. The possible evaluation confound that the 'Direct SLM' rows disable routing and fallback but not the SkillBook prompt is a controlled-comparison validity issue, not a circularity under the definitions used here: the reported number is still an external measurement and is not equivalent to its input by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central empirical claim depends primarily on the strength of the verifier and the fidelity of replay as a proxy for deployment, plus standard assumptions of iterative distillation and RL fine-tuning. No new physical entities are introduced; novel system components (SkillBook, invocation-level router, replay gate) are judged by benchmark evidence rather than by theoretical derivation. The paper lists several of these assumptions as limitations.

free parameters (5)
  • small:large serving cost ratio = 1:10
    Hand-chosen ratio used to normalize deployment cost in Tables 2 and 4; the 60.8% cost claim is directly proportional to this ratio.
  • savings realization factor (finance) = 70%
    Conservative discount applied to estimated serving savings before break-even in Table 4; not derived from data.
  • SkillBook signature keys = humaneval, mbpp
    Hand-specified coarse dataset-level keys; signature quality and skill reuse are not measured at the invocation granularity the method claims as its unit of adaptation.
  • TAU-2 agent token cap = 1024 tokens per turn
    Inference protocol choice for the underpowered 35-task TAU-2 comparison.
  • router threshold = not reported (calibrated on disjoint shard)
    Logistic-regression router threshold chosen on a calibration shard; the router is acknowledged to be weak, so this parameter has limited bearing on the central small-model claim.
assumptions (5)
  • domain assumption The verifier detects all task-relevant failure modes for invocation outcomes.
    Routing labels, fallback decisions, and admission all use the verifier outcome; Section 7 states that incomplete verifier coverage would make replay overestimate safety.
  • domain assumption Replayed candidate runtime states accurately predict online performance after deployment.
    Joint replay is the admission gate; Section 7 acknowledges replay cannot fully capture distribution shift induced by changing the runtime policy.
  • domain assumption Execution-verified teacher demonstrations from failed student invocations are valid supervision for fine-tuning.
    The LLM update distills teacher outputs on hard slices into the student LoRA adapter; this is the mechanism behind the 28.7% to 49.7% lift.
  • standard math Paired-t and one-sided McNemar tests are appropriate and correctly applied.
    Used to claim GRPO exceeds SFT and to characterize TAU-2 as underpowered; details are minimal.
  • domain assumption The two coarse dataset-level SkillBook signatures capture reusable procedure structure.
    SkillBook entries are keyed only by humaneval and mbpp, so its contribution on held-out tasks is a fixed prompt template, not per-invocation skill retrieval.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MERA: Model Evolution and Routing with Skill Adaptation for Agentic Systems at Scale." pith.science (2026). https://pith.science/paper/DE7DTXW5

@misc{pith2026260810333,
  author       = {Pith},
  title        = {Pith review of: MERA: Model Evolution and Routing with Skill Adaptation for Agentic Systems at Scale},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DE7DTXW5}},
  note         = {Machine review of arXiv:2608.10333}
}
read the original abstract

LLM agents execute heterogeneous sequences of model calls within a single task: some invocations require careful reasoning, while others are structured steps such as formatting or tool-argument construction. Prior routing methods exploit this asymmetry by assigning easy invocations to a cheaper small model and difficult ones to a large model. Such policies reduce inference cost, but they leave the small model's capability unchanged, so attainable savings remain bounded by the work the student can already solve. MERA instead improves the small model itself, using a single model invocation as the unit of adaptation. In each cycle, MERA replays failed student invocations to obtain execution-verified teacher demonstrations, distills recurring procedures into an iteratively updated SkillBook, and fine-tunes a student LoRA adapter via supervised learning and optional GRPO. Routing serves as supporting machinery for deployment: the improved student is served behind a cost-calibrated router with verifier-backed fallback, and a candidate SkillBook, adapter, or router is admitted only when joint replay preserves task quality. Empirically, four-cycle adaptation raises Qwen2.5-Coder-1.5B from 28.7% to 49.7% pass on held-out HumanEval+MBPP. Under verifier-backed fallback, the deployed policy retains 88.3% pass at 60.8% of always-Luna cost. On TAU-2, a fine-tuned Qwen3.5-2B improves from 14/35 to 18/35 and matches an unadapted 4B model. These results indicate that verifier-backed multi-cycle adaptation can increase small-model capability, rather than only routing around a fixed student.

Figures

Figures reproduced from arXiv: 2608.10333 by the authors.

Figure 1
Figure 1. Overview of MERA. Online traces drive scheduled SkillBook, LLM-update, and router tracks; their combined state is admitted through joint replay evaluation. It should collect evidence at the granularity of individual invocations, update several candidate compo￾nents, and admit the resulting runtime state only when replay shows that the combined policy still satisfies verification constraints. MERA addresses this prob… view at source ↗
Figure 2
Figure 2. Detailed method view of MERA. Runtime routing remains input-only and verifier-protected; update tracks share traces and are admitted through joint replay. 2 Related Work LLM routing. Prior work on model routing studies how to dispatch requests across model pools under cost-quality trade-offs. FrugalGPT-style cascades, learned routers, and preference- or uncertainty-based selectors use prompt features, predicted qual… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 10 canonical work pages

  1. [1]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InThe Eleventh International Conference on Learning Representations (ICLR 2023),

    work page 2023

  2. [3]

    Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang

    URL https://papers.nips.cc/paper_files/paper/2023/hash/d842425e4bf79ba039352da0f658a90 6-Abstract-Conference.html. Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. Language agent tree search unifies reasoning, acting, and planning in language models. InProceedings of the 41st Inter- national Conference on Machine Learning, v...

    work page 2023

  3. [5]

    Shaokun Zhang, Jieyu Zhang, Jiale Liu, Linxin Song, Chi Wang, Ranjay Krishna, and Qingyun Wu

    URL https://proceedings.iclr.cc/paper_files/paper/2024/hash/28e50ee5b72e90b50e7196fde8ea26 0e-Abstract-Conference.html. Shaokun Zhang, Jieyu Zhang, Jiale Liu, Linxin Song, Chi Wang, Ranjay Krishna, and Qingyun Wu. Of- fline training of language model agents with functions as learnable weights. InProceedings of the 41st International Conference on Machine ...

    work page 2024

  4. [6]

    Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E

    URLhttps://openre view.net/forum?id=cSimKw5p6R. Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M. Waleed Kadous, and Ion Stoica. Routellm: Learning to route llms with preference data.arXiv preprint arXiv:2406.18665,

    arXiv

  5. [7]

    URLhttps://arxiv.org/abs/2406.18665

    doi: 10.48550/arXiv.2406.18665. URLhttps://arxiv.org/abs/2406.18665. Jasper Dekoninck, Maximilian Baader, and Martin Vechev. A unified approach to routing and cascading for LLMs. InProceedings of the 42nd International Conference on Machine Learning, volume 267 ofProceedings of Machine Learning Research, pages 12987–13010. PMLR,

    arXiv doi:10.48550/arxiv.2406.18665

  6. [9]

    doi: 10.186 53/v1/2025.findings-acl.873

    Association for Computational Linguistics. doi: 10.186 53/v1/2025.findings-acl.873. URLhttps://aclanthology.org/2025.findings-acl.873/. 9 MERA Haozhen Zhang, Tao Feng, and Jiaxuan You. Router-R1: Teaching LLMs multi-round routing and aggrega- tion via reinforcement learning,

    work page 2025

  7. [10]

    NeurIPS 2025 poster

    URLhttps://openreview.net/forum?id=DWf4vroKWJ. NeurIPS 2025 poster. Claudio Spiess, Mandana Vaziri, Louis Mandel, and Martin Hirzel. Autopdl: Automatic prompt opti- mization for LLM agents. InProceedings of the Fourth International Conference on Automated Machine Learning, volume 293 ofProceedings of Machine Learning Research, pages 13/1–20. PMLR,

    work page 2025

  8. [12]

    doi: 10.1 8653/v1/2023.findings-acl.507

    Association for Computational Linguistics. doi: 10.1 8653/v1/2023.findings-acl.507. URLhttps://aclanthology.org/2023.findings-acl.507/. Siyu Yuan, Zehui Chen, Zhiheng Xi, Junjie Ye, Zhengyin Du, and Jiecao Chen. Agent-r: Training language model agents to reflect via iterative self-training.arXiv preprint arXiv:2501.11425,

    arXiv 2023

Show all 17 references
  1. [13]

    URLhttps://arxiv.org/abs/2501.11425

    doi: 10.48550/a rXiv.2501.11425. URLhttps://arxiv.org/abs/2501.11425. Chenxi Wang, Zhuoyun Yu, Xin Xie, Wuguannan Yao, Runnan Fang, Shuofei Qiao, Kexin Cao, Guozhou Zheng, Xiang Qi, Peng Zhang, and Shumin Deng. Skillx: Automatically constructing skill knowledge bases for agent...

    doi arXiv

  2. [14]

    URLhttps: //arxiv.org/abs/2604.04804

    doi: 10.48550/arXiv.2604.04804. URLhttps: //arxiv.org/abs/2604.04804. Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents.arXiv preprint arXiv:2207.01206,

    doi arXiv

  3. [16]

    doi: 10.18653/v1/2025.acl-long.150

    Association for Computational Linguistics. doi: 10.18653/v1/2025.acl-long.150. URL https://aclanthology.org/2025.acl-long.150/. Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan.τ-bench: A benchmark for tool-agent- user interaction in real-world domains. InThe Thir...

    2025 doi

  4. [17]

    Carlos E

    URLhttps://neurips.cc/virtual/2024/poster/97468. Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik R. Narasimhan. Swe-bench: Can language models resolve real-world github issues? InThe Twelfth In- ternational Conference on Learning ...

    2024

  5. [2022]

    URLhttps://arxiv.org/abs/2207.01206

    doi: 10.48550/a rXiv.2207.01206. URLhttps://arxiv.org/abs/2207.01206. Junjie Ye, Zhengyin Du, Xuesong Yao, Weijian Lin, Yufei Xu, Zehui Chen, Zaiyuan Wang, Sining Zhu, Zhiheng Xi, Siyu Yuan, Tao Gui, Qi Zhang, Xuanjing Huang, and Jiecao Chen. Toolhop: A query-driven benchmark ...

    doi arXiv

  6. [2023]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettle- moyer, Nicola Cancedda, and Thomas Scialom

    URLhttps://openreview.net/forum?id=WE_vluYUL-X. Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettle- moyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. InAdvances in Neura...

    2023

  7. [2024]

    URLhttps://proceedings.mlr.press/v235/zhou24r.html. Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Lauren Hong, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong S...

    2024

  8. [2025]

    Injae Na, Keonwoong Noh, and Woohwan Jung

    URL https://proceedings.mlr.press/v267/ding25d.html. Injae Na, Keonwoong Noh, and Woohwan Jung. Automatic transmission for LLM tiers: Optimizing cost and accuracy in large language models. InFindings of the Association for Computational Linguistics: ACL 2025, pages 16987–17004...

    2025

  9. [2026]

    Cheng-Yu Hsieh, Chun-Liang Li, Chih-kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister

    URL https://openreview.net/forum?id=ZgCCDwcGwn. Cheng-Yu Hsieh, Chun-Liang Li, Chih-kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and ...

    2023

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.