Pith.

Artificial Intelligence · arXiv:2611.17072 v1 · submitted 12 Nov 2026 · 24 references

Reasoning Supervision and Capability Retention in LoRA Domain Adaptation

Saleh Ashkboos, Dhruv Desai, Lavinia Ghita

The paper says

“Fine-tuning foundation models on financial data has a wide range of applications, including identifying event types from news, monitoring risk and compliance language, and answering questions over filings and earnings calls.”

more of the abstract

In this setting, the main challenge is to improve task performance without degrading the generalization capabilities of the model. In this paper, we present an extensive set of ablations for fine-tuning Qwen3-8B with low-rank adaptation (LoRA). We compare three training regimes: (1) response-only LoRA, where the supervised response is the label alone; (2) single-adapter reasoning, where reasoning traces from a teacher model are included alongside the labels in the training loss; and (3) dual-adapter composite training, where separate LoRA adapters are trained on traces and labels. Response-only LoRA improves in-domain performance and leaves general capabilities near those of the frozen model. The in-domain change is largest for calibration and argmax accuracy. Under single-adapter reasoning, the in-domain metrics remain behind those of the response-only regime, while MMLU-Pro can rise slightly and the other retention benchmarks stay within ordinary variation. Dual-adapter composite training remains on the same side of this comparison and yields no further gain. Finally, we compress the teacher traces by summarizing them before training, and show that shorter summaries move calibration toward the response-only regime without reaching it.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A careful, well-scoped ablation study showing that for this financial classification task, response-only LoRA beats trace-supervised variants on in-domain calibration, and that the retention claims are partly an artifact of the selection rule. This line is wrong

3 major objections · 7 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 24 references say.

The paper says, at §3.6, §4.3

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The selection criterion (a run is kept only if it outperforms the frozen baseline on all tasks) conflates the retention claim with the admission rule. The headline claim that response-only LoRA 'leaves general capabilities near those of the frozen model' is partly an artifact of discarding runs that degraded on any task.

more of objection 1

The paper does not report how many runs were discarded at each stage, nor the distribution of retention scores before filtering. Without this information, the reader cannot assess whether the retention claim is robust or fragile. This is a load-bearing concern because the paper's main conclusion depends on it.

The paper says, at §3.1, §4.1

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

Teacher-generated labels are treated as ground truth for the classification task. Section 3.1 states the 200K examples use labels and reasoning traces generated by Qwen3-32B, and all in-domain metrics measure agreement with those labels on synthetic FinHeadlineMix headlines; no human-annotation validation is provided.

more of objection 2

If the teacher's labels are noisy or systematically different from human event types, the large calibration improvements (NLL from 5.70 to 0.48) mean the model becomes well-calibrated to the teacher's label distribution rather than to true event types, and the ranking of regimes could change. This premise is disclosed as a data-generation choice but is never checked, and the headline claims are only as strong as it.

The paper says, at §4.4, Table 5

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The paper states that only two runs in Stage 2 outperform the frozen model on all tasks (bold squares in Figure 3). However, Table 5 shows that many runs have retention metrics that are within one standard deviation of the frozen baseline, yet they are excluded because they do not strictly outperform on every task.

more of objection 3

The selection criterion is applied before any claim is made, but the central claim about capability retention is stated as a general property of response-only LoRA when it is actually a property of the subset of configurations that passed a multi-task threshold. The paper should report the number of runs discarded and the distribution of retention scores before filtering.

7 minor objections

The paper says, at §3.3, Eq. (1)

Pith says: minor objection 1 of 7

The region-balanced loss uses per-token cross-entropy averaged within each region. The paper should clarify whether the loss is averaged over tokens in each region separately (so each region contributes equally regardless of length) or whether it is a weighted average. The text says 'each bar denotes the mean cross-entropy over the tokens in that region', which suggests the former, but this could be stated more explicitly.

The paper says, at §4.3, Figure 2

Pith says: minor objection 2 of 7

Figure 2 shows the distribution of top runs over hyperparameters, but the sample size is only 5 runs. The conclusions about which hyperparameters matter most (e.g., effective LoRA learning rate) are based on very few data points and should be caveated accordingly.

show 5 more

The paper says, at §4.6, Table 2

Pith says: minor objection 3 of 7

The realized compression ratios (e.g., 0.426 for target 0.50) deviate substantially from the targets. The paper notes this but does not discuss whether the deviation is systematic (e.g., the teacher model cannot compress as much as requested) or whether it affects the interpretation of the results.

The paper says, at §4.5, Figure 4

Pith says: minor objection 4 of 7

Figure 4 shows that the rank split does not separate the runs, but the y-axis ranges are very narrow (e.g., MMLU-Pro from 0.622 to 0.638). The paper should note that the differences are small and may not be practically significant.

The paper says, at §4.2, Table 1

Pith says: minor objection 5 of 7

The frozen model's GPQA score has a standard deviation of 4.3% across three runs. This high variance should be taken into account when interpreting retention results; many of the reported differences may be within noise.

The paper says, at §4.4, Table 5

Pith says: minor objection 6 of 7

The paper notes that λ=1.0 runs perform far worse than response-only training because the training inputs differ (the reasoning text is present). This is an important caveat that should be highlighted earlier, as it affects the interpretation of the blended loss experiments.

The paper says, at §5, Conclusion

Pith says: minor objection 7 of 7

The conclusion states that 'Response-only LoRA is the regime that improves in-domain performance while leaving general capabilities near the frozen model.' This claim should be qualified by noting that it applies only to the subset of configurations that passed the selection criterion.

What would settle it

Re-run the analysis without the selection criterion: report the mean and standard deviation of each retention metric (MMLU-Pro, GPQA, IFEval) across all 432 hyperparameter configurations from Stage 1, not just the 5 that beat the frozen model on every task. If the average retention across all runs is significantly below the frozen baseline, the claim that response-only LoRA 'leaves general capabilities near those of the frozen model' would need qualification.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 13 Nov 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic partly agrees 54 s · 28,311 tokens read, 713 written
  3. Editor wrote the verdict above 41 s · 29,260 tokens read, 607 written
  4. Referee major revision: 3 major objections, 7 minor 1 min 21 s · 28,700 tokens read, 1,359 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The paper is a clean ablation study comparing response-only LoRA with two ways of incorporating teacher reasoning traces (single-adapter blended loss and dual-adapter rank splits) for a financial classification task. The headline finding—that response-only LoRA improves in-domain calibration and argmax accuracy while leaving general capabilities near frozen levels, and that adding traces does not extend these gains—is supported by the tables within the stated scope. The most novel pieces are the region-balanced loss (Eq. 1), the dual-adapter rank splits under a fixed budget, and the trace-summarization curve showing calibration moving toward the response-only regime as traces shorten. These are small but honest extensions of the authors' prior work [10]. The paper is transparent about its staged protocol and reports means and stds over three seeds.

The main soft spot is the selection criterion in §3.6: a run is kept only if it beats the frozen model on every task. This means the central claim that response-only LoRA 'leaves general capabilities near those of the frozen model' is partly a consequence of the admission rule rather than an empirical finding about typical behavior. The paper acknowledges the rule and its motivation, but does not report how many runs were discarded at each stage or what the distribution of retention scores looked like before filtering. Without that, the reader cannot assess robustness. The stress-test note is correct that this is a load-bearing concern, though the paper's own framing is careful enough that a charitable reader would not call it deceptive—just incomplete.

A second concern: ground truth is teacher-generated with no human validation. The calibration improvements (NLL from 5.70 to 0.48) mean the model becomes well-calibrated to the teacher's label distribution, not necessarily to true event types. The paper discloses this as a data-generation choice but never checks it. This limits the external validity of the headline claims.

Minor issues: no significance tests, no code or trace data released, and the most novel pieces are small extensions of [10].

Who this is for: practitioners doing LoRA fine-tuning on classification tasks who want a clear default (label-only) and a matched-budget negative result on reasoning traces. The paper deserves a serious referee because the measurements are reproducible in principle and the question is practically important. I would accept peer review but recommend asking for a selection-robustness analysis and a human-label sanity check.

Verdict: CONDITIONAL accept with moderate revisions.

The strongest objection

The staged protocol (§3.6) applies a selection criterion that a run is kept only if it outperforms the frozen baseline on all tasks, both in-domain and retention. This means that any run that improved in-domain metrics but caused a small drop on, say, GPQA (which has high variance, std ≈ 4.3% per Table 1) would be excluded from further analysis. The headline claim that response-only LoRA 'leaves general capabilities near those of the frozen model' is therefore partly a consequence of the admission rule rather than an empirical finding about the typical behavior of LoRA fine-tuning. The paper acknowledges this rule and its motivation (practical usefulness), but does not report how many runs were discarded at each stage, nor what the distribution of retention scores looked like before filtering. Without that information, the reader cannot assess whether the retention claim is robust or fragile. The concern is not that the rule is invalid—it is a legitimate design choice—but that the central claim about capability retention is stated as a general property of response-only LoRA when it is actually a property of the subset of configurations that passed a multi-task threshold. This is a load-bearing concern because the paper's main conclusion (response-only LoRA preserves general capabilities) depends on it; if the discarded runs showed systematic degradation, the conclusion would be weaker.

The reader's weakest_assumption focused on teacher-generated labels as ground truth, which is a valid concern but is disclosed and partially mitigated by the paper's framing (the study measures agreement with the teacher, not with human judges). The selection-criterion issue is more directly load-bearing for the central claim about capability retention. The paper states the rule transparently (§3.6) and uses it consistently, but the headline claim ('leaves general capabilities near those of the frozen model') is presented as a finding about the response-only regime, not about the selected subset. The paper does not report how many runs were discarded, nor the retention scores of the discarded runs. This is a concrete, addressable concern that does not require re-running the full experiment—only a summary statistic that the authors likely already have. The verdict remains CONDITIONAL because the concern is addressable and does not invalidate the in-domain comparisons, which are the paper's strongest contribution. The agreement_with_reader is 'partial' because the reader identified a different weakest assumption (teacher labels) that is also valid but less central to the retention claim.

The referee's report

If the results hold, the paper provides practical guidance for practitioners fine-tuning LLMs on domain-specific tasks: response-only LoRA is sufficient and adding reasoning traces may not help and can hurt calibration. The staged experimental design with matched hyperparameters is a strength, as is the evaluation on both in-domain and retention benchmarks. However, the significance is limited by the narrow scope (one dataset, one base model, one teacher model) and the reliance on teacher-generated labels without human validation.

Pith's reading of the paper (a machine's summary, not the authors')

LoRA is a cheap way to adapt a large language model to a new task: instead of retraining all weights, you train small extra matrices. The authors wanted to know whether feeding the model, alongside the correct answer, the reasoning steps a bigger teacher model used to reach that answer, helps an 8-billion-parameter model specialize to a 35-class financial headline classification task without losing its general abilities. They ran three matched regimes on 200,000 synthetic headlines. In the first, the model saw the headline and was trained to output the label. In the second, one adapter was trained on a mix of label loss and trace loss. In the third, the rank budget was split between a trace adapter and a label adapter. Everything was repeated with three seeds and checked on three general benchmarks (MMLU-Pro, GPQA, IFEval) plus held-out in-domain metrics. The label-only regime won. It cut negative log-likelihood from about 5.70 to 0.48 and Brier score from 0.86 to 0.23, pushed argmax accuracy from 0.53 to 0.84, and left the three general benchmarks essentially unchanged. Trace supervision consistently produced worse calibration and lower argmax accuracy, whether the traces were full, summarized, or split across two adapters; its only visible effect was a small rise on MMLU-Pro. Short summaries moved calibration partway back toward the label-only numbers without reaching them. Two caveats: every label and trace came from the teacher model, and there is no human-annotation check, so correct means matches the teacher. And runs were only reported if they beat the frozen model on every metric, so capability retention was partly built into the selection rule.

Core claim. Response-only LoRA fine-tuning of Qwen3-8B on 200K financial headlines improves in-domain performance (NLL from 5.70 to 0.48, Brier from 0.86 to 0.23, argmax accuracy from 0.53 to 0.84) while leaving MMLU-Pro, GPQA, and IFEval near frozen levels; adding teacher reasoning traces, whether through a single blended adapter, a dual trace/label adapter split, or summarized traces, does not extend these gains and leaves calibration behind response-only runs (Abstract [b4]-[b7], Tables 4-7).

The premise it rests on. Teacher-generated labels are treated as ground truth for the classification task. Section 3.1 states the 200K examples use the labels and reasoning traces generated by Qwen3-32B, and all in-domain metrics (NLL, Brier, greedy and argmax accuracy, macro-F1) measure agreement with those labels on synthetic FinHeadlineMix headlines; no human-annotation validation is provided. If the teacher's labels are noisy or systematically different from human event types, the large calibration improvements (NLL 5.70 to 0.48) mean the model becomes well-calibrated to the teacher's label distribution rather than to true event types, and the ranking of regimes could change. This premise is disclosed as a data-generation choice but is never checked, and the headline claims are only as strong as it.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Fine-tuning foundation models on financial data has a wide range of applications, including identifying event types from news, monitoring risk and compliance language, and answering questions over filings and earnings calls. In this setting, the main challenge is to improve task performance without degrading the generalization capabilities of the model. In this paper, we present an extensive set of ablations for fine-tuning Qwen3-8B with low-rank adaptation (LoRA). We compare three training regimes: (1) response-only LoRA, where the supervised response is the label alone; (2) single-adapter reasoning, where reasoning traces from a teacher model are included alongside the labels in the training loss; and (3) dual-adapter composite training, where separate LoRA adapters are trained on traces and labels. Response-only LoRA improves in-domain performance and leaves general capabilities near those of the frozen model. The in-domain change is largest for calibration and argmax accuracy. Under single-adapter reasoning, the in-domain metrics remain behind those of the response-only regime, while MMLU-Pro can rise slightly and the other retention benchmarks stay within ordinary variation. Dual-adapter composite training remains on the same side of this comparison and yields no further gain. Finally, we compress the teacher traces by summarizing them before training, and show that shorter summaries move calibration toward the response-only regime without reaching it.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-14

The authors' main claim, as the classifier picked it out 0.98

“Response-only LoRA improves in-domain performance and leaves general capabilities near those of the frozen model.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.03 · states limitation 0.58 · human participants 0.03 · controlled comparison 0.92 · randomized assignment 0.02 · states quantitative result 0.10 · compares with prior results 0.92 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Fine-tuning foundation models on financial data has a wide range of applications, including identifying event types from news, monitoring risk and compliance language, and answering questions over filings and earnings calls. background 1.00 0.06 0.02 asserted 0.99
In this setting, the main challenge is to improve task performance without degrading the generalization capabilities of the model. background 0.90 0.07 0.04 asserted 0.99
In this paper, we present an extensive set of ablations for fine-tuning Qwen3-8B with low-rank adaptation (LoRA). contribution 0.88 0.95 0.03 asserted 0.97
We compare three training regimes: (1) response-only LoRA, where the supervised response is the label alone; (2) single-adapter reasoning, where reasoning traces from a teacher model are included alongside the labels in the training loss; and (3) dual-adapter composite training, where separate LoRA adapters are trained on traces and labels. method 0.89 0.93 0.04 asserted 0.87
Response-only LoRA improves in-domain performance and leaves general capabilities near those of the frozen model. result 1.00 0.93 0.48 asserted 1.00
The in-domain change is largest for calibration and argmax accuracy. result 1.00 0.93 0.07 asserted 1.00
Under single-adapter reasoning, the in-domain metrics remain behind those of the response-only regime, while MMLU-Pro can rise slightly and the other retention benchmarks stay within ordinary variation. result 1.00 0.95 0.19 asserted 0.94
Dual-adapter composite training remains on the same side of this comparison and yields no further gain. result 1.00 0.94 0.14 asserted 1.00
Finally, we compress the teacher traces by summarizing them before training, and show that shorter summaries move calibration toward the response-only regime without reaching it. result 0.93 0.96 0.11 asserted 1.00

What kind of paper the classifier reports this is: empirical study 1.00 · evidence from observation or experiment 0.99

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-14: 1 request, 6,938 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (5)
Figure 1 of the paper
Figure 1. Overview of the staged protocol. The teacher (Qwen3-32B) labels 200K FinHeadlineMix headlines and generates a reasoning trace for each. We then fine-tune the student (Qwen3-8B) with LoRA in four stages: (1) response-only LoRA, trained on the label alone over a hyperparameter grid, (2) single-adapter reasoning, trained on the trace and the label with the blended loss of Eq. (1), (3) dual-adapter composite LoRA, which splits the rank budget between a trace adapter and a label adapter, and (4) summarized traces, which repeats Stages 2 and 3 with teacher-generated summaries at target…
Figure 2 of the paper
Figure 2. Distribution of the top runs over the hyperparameter grid for the 200K dataset (5 runs) for response-only LoRA (Stage 1). The last plot uses the effective step size 𝛼𝜂/𝑟 with 𝑟 = 32, where 𝜂 is the learning rate
Figure 3 of the paper
Figure 3. Single-adapter reasoning (Stage 2) across 𝜆 values, initialized from the top response-only runs of
Figure 4 of the paper
Figure 4. Dual-adapter composite stage (Stage 3). Each row corresponds to one top run from the single-adapter reasoning stage, trained with different optimizer settings, and is evaluated at the same 𝜆 as that run and at adjacent 𝜆 values. All runs share 𝛼 = 8, a learning rate of 5 × 10−5 , and a minimum learning rate of 1 × 10−5 . near the single-adapter and response-only runs. GPQA differences stay within the run-to-run variation of the frozen model. NLL is lowest for response-only training, whose objective is the label itself and so matches the in-domain task directly. Brier, argmax…
Figure 5 of the paper
Figure 5. shows training on compressed traces for the single adapter and the composite adapter. The horizontal axis is the target length. The 25% target is reached; the 50% and 75% targets come out shorter than requested (

Figures are the authors', reproduced from the paper.

References (24)
  1. [10] Lavinia Ghita, Dhruv Desai, and Ioana Boier. Scaling laws for task-specific LLM distillation.arXiv preprint arXiv:2606.24747, 2026. URLhttps://arxiv.org/abs/2606.24747v2 on PithPith's reviewarXiv
  2. [1] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Representations, 2022 on Pith
  3. [2] Qwen Team. Qwen3 technical report, 2025. URLhttps://arxiv.org/abs/2505.09388 on PithPith's reviewarXiv
  4. [3] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA: Efficient finetuning of quantized LLMs. InAdvances in Neural Information Processing Systems, 2023. URLhttps://arxiv.org/abs/2305.14314 on PithPith's reviewarXiv
  5. [4] Dan Biderman, Jacob Portes, Jose Javier Gonzalez Ortiz, Mansheej Paul, Philip Greengard, Connor Jennings, Daniel King, Sam Havens, Vitaliy Chiley, Jonathan Frankle, Cody Blakeney, and John Patrick Cunningham. LoRA learns less and forgets less.Transactions on Machine Learning R... on PithPith's reviewarXiv
  6. [5] Damjan Kalajdzievski. Scaling laws for forgetting when fine-tuning large language models.arXiv preprint arXiv:2401.05605, 2024. URLhttps://arxiv.org/abs/2401.05605 on PithPith's reviewarXiv
  7. [6] John Schulman and Thinking Machines Lab. LoRA without regret.Thinking Machines Lab: Connectionism, 2025. doi: 10.64434/tml.20250929. URLhttps://thinkingmachines.ai/blog/lora/ on Pithdoi:10.64434/tml.20250929
  8. [8] Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. FinGPT: Open-source financial large language models. In FinLLM Workshop at IJCAI, 2023. URLhttps://arxiv.org/abs/2306.06031 on PitharXiv
  9. [11] URLhttps://arxiv.org/abs/2306.05443 on PithPith's reviewarXiv
  10. [12] Yoon Kim and Alexander M. Rush. Sequence-level knowledge distillation. InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, 2016. URLhttps://arxiv.org/abs/1606.07947. 9 Reasoning Supervision and Capability Retention in LoRA Domain Adaptation on PithPith's reviewarXiv
  11. [13] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015. URLhttps://arxiv.org/abs/1503.02531 on PithPith's reviewarXiv
  12. [14] Xinghao Chen, Zhijing Sun, Wenjin Guo, Miaoran Zhang, Yanjun Chen, Yirong Sun, Hui Su, Yijie Pan, Dietrich Klakow, Wenjie Li, and Xiaoyu Shen. Unveiling the key factors for distilling chain-of-thought reasoning. InFindings of the Association for Computational Linguistics: ACL ... on Pithdoi:10.18653/v1/2025
  13. [15] 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! Outperforming larger language models with less training data and smaller model sizes. InFindings of the Asso... on PithPith's reviewarXiv
  14. [16] Xiang Li, Shizhu He, Jiayu Wu, Zhao Yang, Yao Xu, Jun Yang, Haifeng Liu, Kang Liu, and Jun Zhao. MoDE-CoTD: Chain-of-thought distillation for complex reasoning tasks with mixture of decoupled LoRA-experts. InProceedings of the 2024 Joint International Conference on Computation... on Pith
  15. [17] Chengwei Dai, Kun Li, Wei Zhou, and Songlin Hu. Improve student’s reasoning generalizability through cascading decomposed CoTs distillation. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15623–15643, 2024. doi: 10.18653/v1/2024... on Pithdoi:10.18653/v1/2024.emnlp-main.875
  16. [18] Dhruv Desai and Lavinia Ghita. FinHeadlineMix: Large-scale synthetic financial news headlines. https: //huggingface.co/datasets/nvidia/FinHeadlineMix, 2026. Hugging Face dataset on Pith
  17. [19] Maxime Griot, Paul Steven Scotti, and Tanishq Mathew Abraham. Compress-Distill: Reasoning trace compression for efficient knowledge distillation.arXiv preprint arXiv:2606.05988, 2026. URLhttps://arxiv.org/abs/2606.05988 on PithPith's reviewarXiv
  18. [20] Glenn W. Brier. Verification of forecasts expressed in terms of probability.Monthly Weather Review, 78(1):1–3, 1950 on Pith
  19. [21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2019 on Pith
  20. [22] David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. GPQA: A graduate-level Google-proof Q&A benchmark. InFirst Conference on Language Modeling (COLM), 2024 on Pith
show 4 more
  1. [23] Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, Tianle Li, Max Ku, Kai Wang, Alex Zhuang, Rongqi Fan, Xiang Yue, and Wenhu Chen. MMLU-Pro: A more robust and challenging multi-task language unders... on Pith
  2. [24] Diogo Almeida. Introducing system one models and Jev. https://typesafe.ai/blog/ introducing-system-one-models-and-jev, 2026. TypeSafe, September 15, 2026. 10 Reasoning Supervision and Capability Retention in LoRA Domain Adaptation A. Appendix A.1. Hyperparameter Search for Res... on Pith
  3. [25] Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. Instruction-following evaluation for large language models.arXiv preprint arXiv:2311.07911, 2023 on PithPith's reviewarXiv
  4. [2023] URLhttps://arxiv.org/abs/2303.17564 on PithPith's reviewarXiv

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment