Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Learning from Synthetic Data without Model Collapse in Iterative Instruction Tuning

T0 review · 3 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Model collapse in synthetic-data instruction tuning is really a polarization of competence, and a two-stage pipeline that targets diagnosed weak skills while selecting examples at the model's knowledge boundary sustains improvement across g

desk verdict KITE's steady gains are real but rest on an unvalidated DINA diagnostic; referee it with concrete demands. read the letter →

arxiv 2607.17043 v1 pith:EMHZVHI2 submitted 2026-07-19 cs.CL

Xiaonan Luo , Yue Huang , Kehan Guo , Ping He , Chuan Zou , Ting Hua , Xiangliang Zhang This is my paper · ORCID
classification cs.CL
keywords modelcollapsesyntheticdatainstructiontuningself-improvementpolarizationofcompetencecognitivediagnosisDINAuncertaintycuration
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

KITE's central thesis is that when an LLM is iteratively trained on its own synthetic instructions, collapse is not a smooth loss of overall ability: the data distribution drifts toward skills the model already handles, so strong skills get reinforced while weak ones are further neglected. To reverse both arms, the paper first diagnoses weak skills with a cognitive-diagnosis model (DINA) over an LLM-tagged skill matrix, then generates candidate instructions conditioned on those weaknesses and flattens token-probability ranks to force exploration. Selection then uses a likelihood-weighted semantic entropy (KBU) that prefers prompts where the model's sampled answers disagree among plausible modes—the 'knowledge boundary'—rather than easy or off-manifold prompts. Across five open LLMs and four reasoning benchmarks, KITE reports the highest average performance in every model group after five accumulated iterations, and a nine-generation run improves monotonically. A sympathetic reader would take the paper to establish that a fixed labeling budget can be spent more effectively by steering the synthetic instruction distribution toward diagnosed weaknesses and boundary uncertainty.

What carries the argument

Two mechanisms carry the argument. (1) DINA (Deterministic Inputs, Noisy 'And' gate), a cognitive-diagnosis model that turns the model's correct/incorrect responses on training items into per-skill mastery posteriors, given a binary Q-matrix of skill requirements; low-mastery skills become natural-language weakness descriptors that condition generation. This supplies granular, actionable diagnosis rather than aggregate accuracy. (2) Kernel Boundary Uncertainty (KBU), defined as −log tr(A²) where A is a unit-trace, likelihood-weighted RBF kernel over embeddings of m sampled answers; it measures the effective number of semantic modes, and the paper selects instructions with intermediate KBU, n

What would settle it

Permute the skill labels in the Q-matrix across questions while holding the candidate bank, budget, and KBU selection fixed; if KITE's five-iteration average stays at the reported level, failure-guided targeting is not the driver and the effect is carried by KBU curation alone.

Watch

Extended reading notes

Core claim

The paper's central claim, stated on its own terms, is that model collapse in this setting appears as polarization of competence: DINA-based skill-mastery profiles show synthetic fine-tuning raising skills that were already strong while lowering weak ones (e.g., Algebra and Time & Scheduling on GSM8K). KITE is the constructive claim that two design choices reverse this: failure-guided generation using rank-based noise injection q(v|c) ∝ p(v|c) r(v|c)^α, and boundary-aware curation using the Rényi-2 entropy of a likelihood-weighted RBF kernel over sampled answers, with selection restricted to a mid-range quantile band. The paper's evidence is comparative: after five iterations, KITE achieves

Load-bearing premise

The load-bearing premise is that the DINA weakness profile is true signal: the LLM-tagged Q-matrix correctly says which skills each training question requires, and the fixed slip/guess priors with a factorized skill prior yield meaningful mastery scores for a single model; if the tags are noisy or DINA fits poorly, generated candidates target the wrong skills and KITE's advantage erodes.

Editorial extensions

If this is right

  • If KITE is right, steady self-improvement is achievable by changing which instructions are synthesized and selected, not by changing the training objective or adding more real data.
  • A fixed labeling budget buys more progress when spent on boundary examples near diagnosed weaknesses; naive higher-temperature sampling or diversity editing alone sits below the human-data ceiling.
  • Collapse should be monitored per skill: a flat aggregate accuracy can hide simultaneous gains on strong skills and losses on weak ones.
  • Because the pipeline works under accumulated real data, KITE's target failure mode is not unbounded degradation but stagnation below the human-data ceiling; KITE passes that ceiling in the reported settings.
  • Likelihood weighting matters: uncertainty from low-probability off-manifold samples overestimates the boundary, so curation should count disagreement among plausible answers only.

Reading between the lines

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

  • I would generalize the polarization claim into a testable principle: for self-improving LLMs, report per-skill deltas, because aggregate accuracy can be flat while competence splits; skill-level reporting would make collapse visible one generation earlier.
  • The KBU band-pass suggests an optimal-difficulty region for post-training data; that region could be formalized as a curriculum and may predict which benchmarks will respond to instruction tuning.
  • Rank-based noise injection is a coarse analogue of posterior flattening; one could test whether explicit entropy-regularized decoding or temperature schedules tuned per weakness achieve the same effect with fewer samples.
  • Because the Q-matrix is LLM-tagged and non-expert-validated, the method's ceiling is tied to skill-taxonomy quality; an automatic skill-inventory refinement loop could amplify the reported gains.
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 / 6 minor

Summary. The paper studies iterative instruction tuning with self-generated synthetic data and argues that model collapse in this setting appears as a 'polarization of competence'—strong skills improve while weak skills degrade—rather than uniform performance loss. To counter this, the authors propose KITE, a two-stage framework. Stage 1 builds a candidate bank by diagnosing weaknesses with the DINA cognitive model (using an LLM-tagged Q-matrix and fixed priors) and generating instructions via rank-based logit perturbation. Stage 2 curates the bank with Kernel Boundary Uncertainty (KBU), a likelihood-weighted Rényi-2 entropy over sampled answers, selecting prompts near the model's semantic knowledge boundary. Experiments on five open-source LLMs (Qwen-3-4B, Qwen-3-1.7B, Llama-3.2-3B, Llama-3-8B, Gemma-3-4B) across GSM8K, MMLU-Pro, MATH, and GPQA after five iterations of accumulated SFT show KITE best or tied-best in all 20 (model, benchmark) cells and best on average in every backbone. Ablations, a 9-generation horizon check, OOD generalization, and perplexity analysis are included.

Significance. If the empirical claims hold, KITE is a practical and composable data-construction operator: it improves iterative self-improvement under a fixed labeling budget without altering the training objective. The paper is careful in several ways: the verifier is held constant across methods, hyperparameters are shared across all datasets/backbones, the authors openly flag GPQA as non-significant, and they provide cost and statistical reliability discussions. The polarization-of-competence reframing is interesting and, if established, would be a useful diagnostic perspective. However, the central mechanism of Stage 1 rests on an unvalidated DINA heuristic, and the statistical evidence for the headline claim, while consistent, is marginal. These issues are addressable within the manuscript's scope.

major comments (3)
  1. [Section 1, Figure 1; Appendix 8.2.6] The 'polarization of competence' claim is a central motivation, but the evidence in Figure 1 and Appendix 8.2.6 consists of DINA mastery point estimates with no uncertainty intervals or statistical tests. The differences shown (e.g., -0.076, +0.110) are small, and the estimates come from a single response vector with fixed slip/guess priors and an LLM-tagged Q-matrix; the paper itself labels them 'a structured heuristic' (Limitations). This does not establish that collapse reliably takes the form of polarization. I recommend adding bootstrap confidence intervals or a sensitivity analysis over Q-matrix/priors, or reframing Figure 1 as an illustrative observation rather than an empirical finding.
  2. [Section 3.2, Appendix 8.3, Table 2] Stage 1's weakness profiling evaluates the model on its own training set T, which becomes in-sample after iterative SFT, and the Q-matrix is LLM-tagged without expert validation. The ablation in Table 2 shows that removing weakness profiling costs 2.0 average points (43.2 vs 45.2), which is larger than KITE's margin over the best baseline in Table 1 for Llama-3-8B (45.23 vs 44.57). The paper asserts in Appendix 8.3 that 'downstream gains do not require the diagnosis to be exact,' but no robustness evidence is provided—e.g., a noised/corrupted Q-matrix, different slip/guess priors, or a held-out diagnostic set. Because the headline improvement depends on this component, a sensitivity analysis is needed to show KITE is not an artifact of the diagnostic heuristic.
  3. [Section 4.2, Appendix 8.5, Table 1] The statistical backing for 'more stable improvement' is a one-sided sign test over five backbones (p≈0.031). These five are not independent trials: they share the same evaluation protocol, hyperparameters, and judge model. Moreover, the test is performed against the strongest baseline per family, which is selected post hoc; a more conservative comparison against a fixed baseline (e.g., CDS) or a permutation test over (model, benchmark) cells would be more convincing. The paper's own Appendix 8.5 acknowledges that GPQA is within noise and MMLU-Pro is directional, which is honest, but the abstract's unqualified 'more stable improvement' should be tempered or supported by a stronger aggregate test.
minor comments (6)
  1. [Equation (1)] The notation r_t(v|c) is introduced as a rank (an integer), but in q^{(α)}_t(v|c) ∝ q_t(v|c) r_t(v|c)^α it is treated like a probability-like quantity. This is mathematically fine, but the paper should explicitly state that r_t(v|c) is the rank to avoid confusion.
  2. [Section 4.5, Figure 3] In the wikitext-2 experiment, KITE is applied 'without weakness profiling.' This variant should be clearly named as a KBU-only baseline, since the full KITE pipeline includes Stage 1. The text implies this, but Figure 3's legend is ambiguous.
  3. [Appendix 8.6] The theoretical analysis is a standard covariance-mismatch bound for linear-Gaussian models; the connection to rank-based logit perturbation or KBU selection is only asserted. The paper already calls it a proxy, but it would be clearer to label this section as intuition rather than a formal justification of KITE, and to state explicitly what it does and does not explain.
  4. [Appendix 8.2.4] The KBU embedding uses text-embedding-3-small, a proprietary model. For reproducibility, consider also reporting results with an open-source embedding model or providing the precise version used.
  5. [Table 1] The model names in the table (e.g., 'Qwen-3-4B-Instruct') differ slightly from the main text ('Qwen3-4B-Instruct-2507'); please standardize. Also, the table caption says 'five iterations' but the long-horizon section later uses 'Generation 5' for the main-table checkpoint—clarify the indexing.
  6. [Appendix 8.3] The claim that aggregating across items 'yields a stable posterior' for each skill from a single respondent is not self-evident and should be supported by posterior variance estimates or a small simulation.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: KITE's empirical claims rest on held-out evaluation and its formal result is a standard covariance-mismatch bound. Minor self-citations are related-work context only.

full rationale

The derivation chain does not reduce to its inputs by construction. Stage 1 builds weakness descriptors from DINA posterior mastery over responses on the training set; Stage 2 selects a budgeted subset using the model's own likelihood-weighted semantic uncertainty; the evolved model is then evaluated on held-out GSM8K/MMLU-Pro/MATH/GPQA splits that Appendix 8.2.4 states are 'held strictly disjoint' from the training/evaluation split and from which seeds are 'never drawn from the evaluation split.' No fitted parameter is renamed as a prediction: the rank-noise distribution q_t^(α) ∝ q_t r_t^α is an explicit sampling choice and KBU is an explicit curation rule, not a forecast of the measured accuracies. The only formal result, Theorem 1 in Appendix 8.6, is a textbook inverse-Wishart covariance-mismatch bound, and the paper itself disclaims it as 'a theoretical justification... rather than an exact model of our implementation,' so it is not used to manufacture the empirical gains. Self-citations (Huang et al. 2025a,b) appear only in the related-work survey and are not invoked to justify KITE's mechanism, to forbid alternatives, or to import a uniqueness claim. The genuine weaknesses are empirical rather than circular: the Q-matrix is 'LLM-tagged, non-expert-validated' with fixed slip/guess priors, the diagnostic set is the training dataset, and the paper explicitly labels the mastery scores 'a structured heuristic' (Section 6; Appendix 8.3). In-sample diagnosis and possible benchmark-sensitive hyperparameter choice are correctness/robustness risks, not definitional circularity: the reported outcome variables are independent held-out accuracies, not quantities reinserted into the data-construction operator. Therefore no load-bearing circular step is identified; score 2 reflects only the presence of minor, non-load-bearing self-citations.

Assumptions & free parameters 9 free parameters · 7 assumptions · 0 invented entities

The central empirical claim relies on a set of hand-chosen hyperparameters and several domain assumptions (DINA heuristic, external verifier correctness, embedding fidelity, boundary-uncertainty hypothesis). The theory section contributes only a standard covariance-mismatch bound under Gaussian feature-space noise and does not directly justify the rank-based perturbation used in practice.

free parameters (9)
  • Rank-noise strength α_Q = 0.75
    Controls how strongly lower-ranked tokens are upweighted during candidate generation; fixed across datasets with no reported tuning/validation protocol (Appendix 8.2.4).
  • Likelihood-weighting temperature γ = 0.8
    Shapes the answer likelihood weights in KBU; no validation split disclosed.
  • KBU quantile thresholds (umin, umax) = (0.2, 0.8)
    Selects the target percentile range of uncertainty; ablations show alternative bands underperform, so the value is load-bearing.
  • Mastery threshold τ = 0.30
    Determines which skills are considered weak for generation; affects the weakness profile.
  • Slip and guess priors s_j, g_j = 0.1 each
    Fixed DINA item parameters needed because only one response vector is observed; directly affects mastery posteriors.
  • Candidate bank size N and labeling budget B = N=2000, B=500
    Scale of generation and curation; budget is held fixed across methods, but N affects coverage.
  • Number of sampled answers m = 10
    Number of model samples used to estimate KBU.
  • RBF kernel width σ² = median pairwise squared distance
    Data-dependent bandwidth, not hand-tuned, but still a choice affecting KBU.
  • LoRA rank / learning rate / epochs = rank=8, lr=1e-5, 1 epoch
    Shared training hyperparameters; not part of KITE but affect all comparisons.
assumptions (7)
  • standard math Wishart/inverse-Wishart expectation E[(XᵀX)⁻¹] = (T-d-1)⁻¹ Σ_train⁻¹ for T > d+1
    Used in the proof of Theorem 1 (Appendix 8.6); standard result.
  • domain assumption The linear-Gaussian proxy x ~ N(0, Σ), y = wᵀx + ε is a valid model of LLM training-data distributions
    Adopted from Dohmatob et al. (2024a); the paper acknowledges it is 'not an exact model of our implementation' (Appendix 8.6).
  • domain assumption DINA cognitive-diagnosis model with LLM-tagged Q-matrix, fixed slip/guess priors, and factorized skill prior yields useful weakness profiles
    Stage 1 uses posterior mastery scores to target weak skills; Limitations calls the Q-matrix 'non-expert-validated' and scores 'a structured heuristic'.
  • domain assumption The external model gpt-5-mini provides correct verified labels and correct judge decisions
    All methods use the same verifier/judge; the answer channel distills from a stronger external model (Limitations, Appendix 8.2.5).
  • domain assumption text-embedding-3-small embeddings capture answer semantic equivalence
    KBU kernel compares answer embeddings; if embeddings ignore key distinctions, uncertainty estimates are off.
  • ad hoc to paper Rank-based logit perturbation approximates feature-space covariance noise injection
    The paper states this connection informally and says the analysis 'should be understood as a theoretical justification ... rather than as an exact model of our implementation' (Appendix 8.6).
  • ad hoc to paper Intermediate KBU values indicate proximity to the semantic knowledge boundary and thus useful learning signal
    The band-pass curation hypothesis is motivated by intuition and validated only indirectly by the main results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning from Synthetic Data without Model Collapse in Iterative Instruction Tuning." pith.science (2026). https://pith.science/paper/EMHZVHI2

@misc{pith2026260717043,
  author       = {Pith},
  title        = {Pith review of: Learning from Synthetic Data without Model Collapse in Iterative Instruction Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EMHZVHI2}},
  note         = {Machine review of arXiv:2607.17043}
}
read the original abstract

Model collapse is a central challenge in learning from synthetic data: as later-generation large language models (LLMs) are trained on an increasing proportion of model-generated data, performance can degrade due to narrowed coverage and accumulated bias. Existing work mainly studies how to bound this degradation. In iterative model evolution, however, the more meaningful objective is to ensure that each successive model improves over its predecessor, which requires diagnosing collapse at a granularity that is actionable for data curation. We study this problem in synthetic data self-improving for instruction tuning. We show that collapse in this setting is not simply uniform performance degradation, but can appear as a polarization of competence, where synthetic training reinforces already strong skills while further degrading weak ones. Motivated by this observation, we propose KITE (Knowledge-boundary Instruction Tuning via Exploration), a two-stage framework that combines failure-guided data generation with boundary-aware uncertainty curation. Experiments across several datasets and multiple open-source LLMs show that KITE yields more stable improvement than strong synthetic-data baselines.

Figures

Figures reproduced from arXiv: 2607.17043 by the authors.

Figure 1
Figure 1. Selective skill-mastery profiles of Llama-3- [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Top: iterative model evolution with KITE synthesis. Bottom: the two stages of KITE, failure-guided [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Perplexity-based analysis of synthetic data under recursive training. (a) Perplexity distributions on wikitext-2 with Llama-3-8B-Instruct for human-written data, synthetic data generated with temperature=0.3, synthetic data generated with temperature=1.0, and synthetic data generated with KITE. (b) Perplexity on the wikitext-2 test set across recursive model generations when Llama-3-8B-Instruct trained on synthetic … view at source ↗
Figures from the paper (8 more)
Figure 5
Figure 5. Figure 5: shows whether KITE trained on all math subsets of the four in-domain datasets described above can generalize to out-of-distribution (OOD) mathematical reasoning benchmarks. We evaluate 40 50 60 70 Accuracy (%) OlymMATH-Easy AIME-2025 AIME-2026 Initial Model Human-Data …
Figure 6
Figure 6. Figure 6: LLM skill tagging prompt used to construct the Q-matrix for the DINA-based diagnosis pipeline. Given a [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Few-shot synthesis prompt used in the Few-shot synthesis baseline and in [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Answer comparison prompt used in the LLM-as-judge evaluation pipeline. The judge compares the model [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: Labeling prompt used for the external verifier in the training-time labeling stage. Given a selected synthetic [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 10
Figure 10. Figure 10: Skill mastery profile on MMLU-Pro-Physics estimated by DINA for the initial and fine-tuned models. [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Skill mastery profile on MATH estimated by DINA for the initial and fine-tuned models. Each [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: Skill mastery profile on GPQA estimated by DINA for the initial and fine-tuned models. Each [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Question Begets Question: Self-Evolving Curriculum for Reinforcement Fine-Tuning on Competition Mathematics

    cs.LG 2026-08 conditional novelty 5.0 of 10

    A self-evolving curriculum that retrains a language model on variants of problems it can mostly get right lifts AIME pass@1 from 5.6% to 16.5%, beating static augmentation under the same data budget.

Reference graph

Works this paper leans on

56 extracted references · 3 canonical work pages · cited by 1 Pith paper

  1. [1]

    2024 , eprint=

    The Curse of Recursion: Training on Generated Data Makes Models Forget , author=. 2024 , eprint=

    2024

  2. [2]

    Forty-first International Conference on Machine Learning , year=

    A Tale of Tails: Model Collapse as a Change of Scaling Laws , author=. Forty-first International Conference on Machine Learning , year=

  3. [3]

    First Conference on Language Modeling , year=

    Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data , author=. First Conference on Language Modeling , year=

  4. [4]

    Forty-second International Conference on Machine Learning , year=

    How to Synthesize Text Data without Model Collapse? , author=. Forty-second International Conference on Machine Learning , year=

  5. [5]

    The Thirteenth International Conference on Learning Representations , year=

    Beyond Model Collapse: Scaling Up with Synthesized Data Requires Verification , author=. The Thirteenth International Conference on Learning Representations , year=

  6. [6]

    Machine-generated text detection prevents language model collapse

    Drayson, George and Yilmaz, Emine and Lampos, Vasileios. Machine-generated text detection prevents language model collapse. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.1506

    doi:10.18653/v1/2025.emnlp-main.1506 2025

  7. [7]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Escaping Collapse: The Strength of Weak Data for Large Language Model Training , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  8. [8]

    First Conference on Language Modeling , year=

    Best Practices and Lessons Learned on Synthetic Data , author=. First Conference on Language Modeling , year=

Show all 56 references
  1. [9]

    Generating Datasets with Pretrained Language Models

    Schick, Timo and Sch. Generating Datasets with Pretrained Language Models. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021. doi:10.18653/v1/2021.emnlp-main.555

    2021 doi

  2. [10]

    Synthetic Data Augmentation for Zero-Shot Cross-Lingual Question Answering

    Riabi, Arij and Scialom, Thomas and Keraron, Rachel and Sagot, Beno \^i t and Seddah, Djam \'e and Staiano, Jacopo. Synthetic Data Augmentation for Zero-Shot Cross-Lingual Question Answering. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processin...

    2021 doi

  3. [11]

    WildChat: 1M Chat

    Wenting Zhao and Xiang Ren and Jack Hessel and Claire Cardie and Yejin Choi and Yuntian Deng , booktitle=. WildChat: 1M Chat. 2024 , url=

    2024

  4. [12]

    Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned

    Zhangchen Xu and Fengqing Jiang and Luyao Niu and Yuntian Deng and Radha Poovendran and Yejin Choi and Bill Yuchen Lin , booktitle=. Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned. 2025 , url=

    2025

  5. [13]

    2025 , eprint=

    Oasis: One Image is All You Need for Multimodal Instruction Data Synthesis , author=. 2025 , eprint=

    2025

  6. [14]

    2025 , eprint=

    SyntheT2C: Generating Synthetic Data for Fine-Tuning Large Language Models on the Text2Cypher Task , author=. 2025 , eprint=

    2025

  7. [15]

    2024 , eprint=

    Simple synthetic data reduces sycophancy in large language models , author=. 2024 , eprint=

    2024

  8. [16]

    ChemOrch: Empowering

    Yue Huang and Zhengzhe Jiang and Xiaonan Luo and Kehan Guo and Haomin Zhuang and Yujun Zhou and Zhengqing Yuan and Xiaoqi Sun and Jules Schleinitz and Yanbo Wang and Shuhao Zhang and Mihir Surve and Nitesh V Chawla and Olaf Wiest and Xiangliang Zhang , booktitle=. ChemOrch: Em...

    2025

  9. [17]

    AugGPT: Leveraging ChatGPT for Text Data Augmentation , year=

    Dai, Haixing and Liu, Zhengliang and Liao, Wenxiong and Huang, Xiaoke and Cao, Yihan and Wu, Zihao and Zhao, Lin and Xu, Shaochen and Zeng, Fang and Liu, Wei and Liu, Ninghao and Li, Sheng and Zhu, Dajiang and Cai, Hongmin and Sun, Lichao and Li, Quanzheng and Shen, Dinggang a...

  10. [18]

    MetaSynth: Meta-Prompting-Driven Agentic Scaffolds for Diverse Synthetic Data Generation , url=

    Riaz, Haris and Bhabesh, Sourav Sanjukta and Arannil, Vinayak and Ballesteros, Miguel and Horwood, Graham , year=. MetaSynth: Meta-Prompting-Driven Agentic Scaffolds for Diverse Synthetic Data Generation , url=. doi:10.18653/v1/2025.findings-acl.962 , booktitle=

    2025 doi

  11. [19]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Aligning to Thousands of Preferences via System Message Generalization , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  12. [20]

    The Thirteenth International Conference on Learning Representations , year=

    DataGen: Unified Synthetic Dataset Generation via Large Language Models , author=. The Thirteenth International Conference on Learning Representations , year=

  13. [21]

    arXiv preprint arXiv:2110.14168 , year=

    Training Verifiers to Solve Math Word Problems , author=. arXiv preprint arXiv:2110.14168 , year=

    arXiv

  14. [22]

    Journal of Educational and Behavioral Statistics , volume =

    De La Torre, Jimmy , title =. Journal of Educational and Behavioral Statistics , volume =

  15. [23]

    2024 , url=

    Yubo Wang and Xueguang Ma and Ge Zhang and Yuansheng Ni and Abhranil Chandra and Shiguang Guo and Weiming Ren and Aaran Arulraj and Xuan He and Ziyan Jiang and Tianle Li and Max Ku and Kai Wang and Alex Zhuang and Rongqi Fan and Xiang Yue and Wenhu Chen , booktitle=. 2024 , url=

    2024

  16. [24]

    2021 , eprint=

    Measuring Massive Multitask Language Understanding , author=. 2021 , eprint=

    2021

  17. [25]

    2021 , eprint=

    Measuring Mathematical Problem Solving With the MATH Dataset , author=. 2021 , eprint=

    2021

  18. [26]

    Bowman , booktitle=

    David Rein and Betty Li Hou and Asa Cooper Stickland and Jackson Petty and Richard Yuanzhe Pang and Julien Dirani and Julian Michael and Samuel R. Bowman , booktitle=. 2024 , url=

    2024

  19. [27]

    2023 , eprint=

    Self-Instruct: Aligning Language Models with Self-Generated Instructions , author=. 2023 , eprint=

    2023

  20. [28]

    Kernel Language Entropy: Fine-grained Uncertainty Quantification for

    Alexander V Nikitin and Jannik Kossen and Yarin Gal and Pekka Marttinen , booktitle=. Kernel Language Entropy: Fine-grained Uncertainty Quantification for. 2024 , url=

    2024

  21. [29]

    2026 , howpublished =

    2026

  22. [30]

    Introducing Claude Opus 4.6 , year =

  23. [31]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  24. [32]

    2024 , url =

    Llama 3 Model Card , author=. 2024 , url =

    2024

  25. [33]

    Gemma 3 , url=

    Gemma-Team , year=. Gemma 3 , url=

  26. [34]

    Forty-second International Conference on Machine Learning , year=

    Collapse or Thrive: Perils and Promises of Synthetic Data in a Self-Generating World , author=. Forty-second International Conference on Machine Learning , year=

  27. [35]

    The Thirteenth International Conference on Learning Representations , year=

    Strong Model Collapse , author=. The Thirteenth International Conference on Learning Representations , year=

  28. [36]

    2023 , eprint=

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena , author=. 2023 , eprint=

    2023

  29. [37]

    and Le, Quoc V and Firat, Orhan

    Kazemi, Mehran and Fatemi, Bahare and Bansal, Hritik and Palowitch, John and Anastasiou, Chrysovalantis and Mehta, Sanket Vaibhav and Jain, Lalit K and Aglietti, Virginia and Jindal, Disha and Chen, Peter and Dikkala, Nishanth and Tyen, Gladys and Liu, Xin and Shalit, Uri and ...

    2025

  30. [38]

    AIME Problems and Solutions , year =

  31. [39]

    2025 , eprint=

    Challenging the Boundaries of Reasoning: An Olympiad-Level Math Benchmark for Large Language Models , author=. 2025 , eprint=

    2025

  32. [40]

    L lama F actory: Unified Efficient Fine-Tuning of 100+ Language Models

    Zheng, Yaowei and Zhang, Richong and Zhang, Junhao and Ye, Yanhan and Luo, Zheyan. L lama F actory: Unified Efficient Fine-Tuning of 100+ Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)....

    2024 doi

  33. [41]

    2016 , eprint=

    Pointer Sentinel Mixture Models , author=. 2016 , eprint=

    2016

  34. [42]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Model Collapse Demystified: The Case of Regression , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  35. [43]

    CDS : Data Synthesis Method Guided by Cognitive Diagnosis Theory

    Zhao, Haokun and Han, Jinyi and Liang, Jiaqing and Xiao, Yanghua and Meng, Xiaojun and Wei, Jiansheng. CDS : Data Synthesis Method Guided by Cognitive Diagnosis Theory. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.439

    2025 doi

  36. [44]

    2024 , eprint=

    Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models , author=. 2024 , eprint=

    2024

  37. [45]

    2022 , eprint=

    STaR: Bootstrapping Reasoning With Reasoning , author=. 2022 , eprint=

    2022

  38. [46]

    2025 , eprint=

    Self-Rewarding Language Models , author=. 2025 , eprint=

    2025

  39. [47]

    2023 , eprint=

    Reinforced Self-Training (ReST) for Language Modeling , author=. 2023 , eprint=

    2023

  40. [48]

    2026 , eprint=

    Self-Generative Adversarial Fine-Tuning for Large Language Models , author=. 2026 , eprint=

    2026

  41. [49]

    2026 , eprint=

    Iterative Finetuning is Mostly Idempotent , author=. 2026 , eprint=

    2026

  42. [50]

    Language Models can Categorize System Inputs for Performance Analysis

    Sobhani, Dominic and Zhong, Ruiqi and Marrese-Taylor, Edison and Sakaguchi, Keisuke and Matsuo, Yutaka. Language Models can Categorize System Inputs for Performance Analysis. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Compu...

    2025 doi

  43. [51]

    Second Conference on Language Modeling , year=

    EvalTree: Profiling Language Model Weaknesses via Hierarchical Capability Trees , author=. Second Conference on Language Modeling , year=

  44. [52]

    Forty-second International Conference on Machine Learning , year=

    When Dynamic Data Selection Meets Data Augmentation: Achieving Enhanced Training Acceleration , author=. Forty-second International Conference on Machine Learning , year=

  45. [53]

    2024 , eprint=

    LESS: Selecting Influential Data for Targeted Instruction Tuning , author=. 2024 , eprint=

    2024

  46. [54]

    Superfiltering: Weak-to-Strong Data Filtering for Fast Instruction-Tuning

    Li, Ming and Zhang, Yong and He, Shwai and Li, Zhitao and Zhao, Hongyu and Wang, Jianzong and Cheng, Ning and Zhou, Tianyi. Superfiltering: Weak-to-Strong Data Filtering for Fast Instruction-Tuning. Proceedings of the 62nd Annual Meeting of the Association for Computational Li...

    2024 doi

  47. [55]

    The Fourteenth International Conference on Learning Representations , year=

    OptimSyn: Influence-Guided Rubrics Optimization for Synthetic Data Generation , author=. The Fourteenth International Conference on Learning Representations , year=

  48. [56]

    2026 , eprint=

    Learning from Synthetic Data: Limitations of ERM , author=. 2026 , eprint=

    2026

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.