Pith. sign in

REVIEW 4 major objections 4 minor 41 references

Iterative preference optimization on a model's own residual errors lifts table entity linking to 94.50% on TableInstruct and 0.882 CEA on MammoTab V2.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Iterative preference learning on a model's own residual errors improves table entity linking (88.2% on MammoTab V2, surpassing the listed 86% baseline) and, with length-normalized regularization, lifts reasoning-based linking from 79.09% to 81.85%.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Useful empirical recipe for table EL with consistent internal gains, but the 'refresh' mechanism is confounded with extra training and the leaderboard SOTA claim is uncontrolled. the 4 major comments →

arxiv 2607.28680 v1 pith:H35C2W36 submitted 2026-07-29 cs.CL cs.LG

TELLER: Dual-Path Iterative Preference Optimization for Table Entity Linking

Yixin Peng , Kehao Li , Stefan Decker This is my paper
classification cs.CL cs.LG
keywords table entity linkingcell entity annotationiterative preference optimizationdirect preference optimizationchain-of-thought reasoninglength-normalized preference optimizationWikidata candidate retrievaltable understanding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

TELLER sets out to show that table entity linking—matching short cell mentions to knowledge-base entities—gets better when preference optimization is iterated on the model's current mistakes instead of being trained once on fixed chosen–rejected pairs. On a compact, candidate-conditioned prompt, two training paths are compared: a direct-answer path using iterative direct preference optimization, and a reasoning path using filtered chain-of-thought supervision followed by iterative length-normalized regularized preference optimization (L-RPO). The claimed payoff is concrete: direct answers reach 94.50% on TableInstruct and 88.20% (CEA 0.882) on MammoTab V2, and the reasoning path raises MammoTab V2 accuracy from 79.09% to 81.85% while keeping complete rationales in 91.86% of outputs. If these numbers hold, the contribution is that residual-error-driven iteration and length-aware regularization—not larger models or more data—are what unlock further gains from an instruction-tuned 8-billion-parameter model on structured entity-linking tasks.

Core claim

The paper's central claim, on its own terms, is that the bottleneck in generative table entity linking is stale supervision. After retrieving and ranking up to 20 Wikidata candidates and serializing only the target cell, its row/column, headers, and caption into the prompt, TELLER refines a supervised model through iterative preference learning. In the direct-answer path, each DPO round pairs the gold entity with the current model's incorrect predictions, so the second round sees only errors the first round did not resolve. In the reasoning path, filtered teacher rationales provide CoT supervision, and iterative L-RPO then optimizes a length-normalized preference margin while also pulling up

What carries the argument

The load-bearing mechanism is the residual-error refresh: at each round, the updated model regenerates responses, and the preference pairs for the next round are built only from its remaining mistakes. The other central object is L-RPO, the reasoning-path objective that (1) replaces total sequence log-probability with per-token average log-probability in both the trainable and frozen reference models, so long rationales do not dominate the preference margin, and (2) adds a supervised term that directly raises the likelihood of the chosen, gold-consistent rationale. The paper also contributes an offline candidate pipeline—alias matching plus BM25 over a local Wikidata index, rescored by a wei

Load-bearing premise

The headline comparison assumes that TELLER's exact-match accuracy over a 20-candidate shortlist is numerically comparable to the public MammoTab V2 CEA scores of the earlier systems the paper cites, which were posted under different or unspecified candidate-retrieval constraints; if the protocols do not match, the 0.882 versus 0.86 margin is not established.

What would settle it

Run the cited baselines on TELLER's exact 20-candidate prompt set with the same exact-match scorer, or run TELLER under the baselines' original candidate/protocol conditions. If a cited baseline also reaches or exceeds 0.882 under the matched protocol, or if TELLER drops below 0.86 without its 20-candidate shortlist, then the claimed leaderboard improvement is a protocol artifact. Separately, train the direct path with round 1's preference pairs kept fixed for round 2 (no refresh): if static-pair DPO matches 88.20%, the residual-error refresh is not the cause.

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

If this is right

  • Iterative DPO gives monotonic direct-answer gains on both benchmarks: 94.35% → 94.40% → 94.50% on TableInstruct and 87.59% → 88.16% → 88.20% on MammoTab V2.
  • Iterative L-RPO recovers most of the cost of explicit reasoning: MammoTab V2 accuracy rises from 79.09% (CoT-SFT) to 81.85%, with complete-rationale rate rising from 87.09% to 91.86%.
  • Length normalization is load-bearing: removing it in L-RPO round 1 cuts MammoTab accuracy from 79.37% to 69.53% and rationale completion from 89.63% to 77.06%; round 2 shows a smaller but similar drop.
  • The chosen-response regularization is also load-bearing: removing it in round 1 cuts accuracy to 68.52% and completion to 74.30%.
  • Direct-answer models remain more accurate than reasoning-based models on both datasets, indicating a persistent accuracy–interpretability trade-off.

Where Pith is reading between the lines

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

  • A controlled test the paper does not run: train the same DPO objective for the same number of steps but keep round 1's preference pairs fixed for round 2. If static pairs match the 88.20% result, the residual-error refresh is not the active ingredient; if they fall short, the refresh is doing the work.
  • The head-to-head claim against the public 0.86 CEA is only as strong as protocol comparability; a fair public comparison would require scoring all systems under identical retrieval and candidate-set conditions, not just exact-match within TELLER's fixed 20-candidate shortlist.
  • The paper stops at two rounds; a natural extension is to iterate until convergence and measure whether gains plateau or reverse, giving an empirical picture of when residual-error refresh has saturated.
  • Because every training example guarantees the gold entity among the candidates, the model has never learned NIL behavior; adding explicitly unlinkable mentions and measuring whether the pipeline can learn to abstain would test the generalizability of the approach to real tables.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper presents TELLER, a two-path training framework for table entity linking. An offline pipeline retrieves and ranks Wikidata candidates and serializes compact table context into prompts; a Llama-3.1-8B model is then trained with LoRA. The direct-answer path runs SFT followed by two rounds of iterative DPO, where each round rebuilds preference pairs from the current model's incorrect predictions. The reasoning path runs CoT-SFT on filtered and compressed teacher rationales, followed by iterative L-RPO, a length-normalized DPO variant with a chosen-response likelihood loss. Experiments on TableInstruct's entity-linking subset and MammoTab V2 report progressive accuracy gains, and the best MammoTab V2 result (0.882 CEA) is claimed to exceed the public leaderboard's best scored baseline (0.86).

Significance. If the core claims hold, the paper makes a useful contribution: the compact serialization and candidate-conditioned formulation are sensible, L-RPO directly targets a known length-bias problem in preference optimization, and the paired error analysis is a strength. The release of data, code, and rationales is also commendable. However, the headline leaderboard comparison is not adequately controlled, and the central causal claim that refreshing preference data with residual errors drives the round-2 gains is not isolated from the effect of continued training. Small accuracy differences are reported without variance, so the quantitative claims need strengthening before the conclusions can be accepted.

major comments (4)
  1. [§4.2, §5.1, Table 3] The claim that iterative DPO round 2 'exceeds the best baseline result of 0.86 on the public MammoTab V2 evaluation page' is not evidence-based as stated. TELLER scores are exact-match accuracy on a 20-candidate shortlist produced by the paper's own retrieval pipeline, while the public scores for TURL, Avogadro, and TableLlama were obtained under their own candidate/protocol constraints, which the paper does not document as identical. The sentence 'Since each target mention receives one prediction, our exact-match accuracy is the CEA score' conflates the metric definition with the evaluation protocol. The high-recall candidate set can be a major advantage or disadvantage relative to the baselines. Please either evaluate under the official harness with the same candidate sets, or report the public baselines re-run under the candidate-conditioned protocol; otherwise remove the leaderboard
  2. [§3.4, Table 2] The iterative-refresh claim is not isolated from continued training. In §3.4, DPO round k is initialized from π_{k−1} and round-2 data are the residual errors of π_1; Table 2 then compares DPO-1 with DPO-2. But DPO-2 differs from DPO-1 in two ways at once: the preference pairs are refreshed and the model receives another full round of gradient updates starting from the DPO-1 checkpoint. No control trains a second round on the same round-1 preference pairs (or any fixed preference set) with matched epochs and checkpoint selection. The same gap affects L-RPO and Eq. (10), where the round-2 gain on MammoTab V2 (79.09→81.85) could be due to additional optimization rather than to replacing stale negatives. Please add such a control, or soften the causal language in the abstract, §1, and §3.4, where 'controlled comparison' overstates the design.
  3. [§4.3, §5.1] The paper reports that 'all reported experimental results are averaged over five repeated evaluation runs,' but no variance, confidence intervals, or significance tests are given. This matters because several headline gains are small: 94.35→94.40→94.50 on TableInstruct and 87.59→88.16→88.20 on MammoTab V2. Without run-to-run variability, the reader cannot judge whether the direct-answer improvements are distinguishable from noise. Please report standard deviations (or other uncertainty estimates) for the five runs, and ideally a paired test for the model-to-model comparisons.
  4. [§5.2, Table 4] The ablation study as presented is not a controlled comparison. The last paragraph of §5.2 states that the ablations 'use different effective batch sizes from the full runs' and that 'the full L-RPO-1 evaluation uses the alternative serialization described above.' This undermines the quantitative attributions in Table 4, e.g., the claim that removing length normalization reduces accuracy from 79.37% to 69.53% in round 1. The text already concedes this by calling the results 'diagnostic evidence rather than fully controlled estimates,' but the section heading and surrounding exposition still imply component-level causal claims. Please rerun the ablations with matched batch sizes and serialization, or explicitly present them as diagnostics and remove the strong component-level conclusions.
minor comments (4)
  1. [§1, §3.4] The phrase 'controlled comparison' is used to describe the two-round iterative setup, but as noted in the major comments, the design does not isolate data refresh from continued training. Reword to avoid implying a control that is not present.
  2. [Appendix A.2] In the preference instance example, 'model_prediction' uses the abbreviated token '[DESC]' while the gold and chosen responses use '[DESCRIPTION]'. Under the paper's exact-match evaluation, this formatting difference alone would cause a mismatch. Please clarify whether this is only a display artifact or an actual serialization inconsistency.
  3. [Table 1] The relationship between 'Sampled training data' (168,764 for TableInstruct, 838,182 for MammoTab) and the final SFT selection (35,000 and 70,000) is not explained. Clarify the sampling procedure and why the sampled counts are not used directly.
  4. [Algorithm 1] The candidate scoring weights w_i are described as 'fixed heuristic weights,' but no values or sensitivity analysis are given. Since this is a free parameter of the pipeline, a brief sensitivity check or a statement of stability would help.

Circularity Check

0 steps flagged

No significant circularity: results are measured on held-out test sets and the training pipeline does not reduce to its inputs.

full rationale

TELLER's derivation chain is self-contained. The central quantities that are claimed as predictions—test-set accuracies on TableInstruct and MammoTab V2—are measured after training on disjoint held-out evaluation sets; no test statistic appears in any loss. Preference data are constructed from the current model's errors paired with external gold labels (§3.3), and the DPO/L-RPO updates (Eqs. 6, 8–10) are standard objectives or explicit, cited combinations of known components; they are not defined in terms of the accuracy values they are used to explain. The 'residual-error refresh' is a data-construction rule, not a fitted parameter renamed as a result, so the progressive round-1/round-2 numbers do not reduce to the construction by definition. Two methodological caveats surfaced in review are real but non-circular: round 2 is initialized from the round-1 checkpoint and trained on fresh pairs without a fixed-pair, matched-compute control, so the refresh mechanism is not isolated; and the 0.882-vs-0.86 leaderboard comparison assumes the public CEA protocol is comparable to the paper's 20-candidate exact-match setting. Neither is an equation-level or citation-level circularity. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in by self-reference.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central claims rest on a handful of modeling and evaluation assumptions rather than a mathematical derivation. The main ones: (i) candidate retrieval always includes the gold entity in the construction data, so retrieval failure is never modeled; (ii) the public leaderboard scores can be compared with TELLER's candidate-restricted exact-match accuracy; (iii) filtered teacher rationales are trustworthy CoT supervision; (iv) length normalization correctly removes length bias; (v) iterating on the model's own errors supplies the right training signal. The free parameters are mostly hand-set hyperparameters and unreported candidate-scoring weights.

free parameters (7)
  • Candidate scoring weights w_i (lex, prior, type, ctx, bm25) = not reported
    Algorithm 1 combines five scores with fixed heuristic weights; values never given, so the candidate shortlist—and every accuracy number—depends on unspecified hand-chosen weights.
  • L-RPO regularization coefficient α = 0.5
    Eq. (10); hand-set; ablations show the chosen-loss term is critical, but the coefficient is not tuned or derived.
  • Preference scaling β (L-RPO) = 0.3
    Used in Eq. (8); hand-set, no sensitivity analysis reported.
  • Candidate shortlist size top-k = 20
    Algorithm 1; fixed, affects task difficulty and upper bound on accuracy.
  • Number of iterative preference rounds = 2
    §3.3/§5: runs two rounds for controlled comparison; authors note it is unclear if further rounds help.
  • Training-data sampling seed = 42
    §3.3: random sample one target cell per table using seed 42; changes the composition of the 105k training set.
  • Chosen/rejected length-ratio cutoff = 6
    §4.3: pairs with length ratio >6 are removed; hand-set.
axioms (6)
  • domain assumption Gold entity is always present among the 20 retrieved candidates in the constructed dataset
    §3.3 states every example in the final dataset includes its gold entity; this removes retrieval failures from consideration and makes the task easier than the leaderboard task if applied unevenly.
  • domain assumption Exact-match accuracy on TELLER's candidate-conditioned task is equal to the official CEA score used on the public leaderboard
    §4.2/§5.1 claim 'our exact-match accuracy is the CEA score used in this comparison'; the equivalence relies on the baselines using the same candidate restriction.
  • domain assumption Teacher rationales that pass the filtering criteria are faithful and useful supervision
    §2.2/§3.3: rationales filtered by format/correctness but not verified for faithfulness to the model's decision process.
  • domain assumption Length-normalized average log-probability removes length bias without discarding quality information
    §3.4 Eq. (7); the basis for L-RPO.
  • domain assumption The current model's residual errors are the most useful preference pairs for the next update
    §3.3 iterative refresh; central to the method's motivation, borrowed from iterated RLHF/IRPO.
  • ad hoc to paper Compact serialization (keep row/column/headers/caption, drop other cells) provides sufficient context
    §3.3; chosen to reduce distraction and input length, motivated by prior attention analysis, but not derived from first principles.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TELLER: Dual-Path Iterative Preference Optimization for Table Entity Linking." pith.science (2026). https://pith.science/paper/H35C2W36

@misc{pith2026260728680,
  author       = {Pith},
  title        = {Pith review of: TELLER: Dual-Path Iterative Preference Optimization for Table Entity Linking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H35C2W36}},
  note         = {Machine review of arXiv:2607.28680}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Entity linking in tables matches short and ambiguous cell mentions to their corresponding knowledge-base entities. Existing approaches typically rely on data preprocessing pipelines that retain either compact or extensive table content as contextual evidence, and then formulate entity linking as a language generation task for instruction-tuned models; recent systems further incorporate explicit reasoning to disambiguate challenging mentions. However, their training supervision is usually static: fixed preference data cannot adapt to the residual errors of an evolving model, while variations in reasoning length can bias sequence-level preference learning. To address these limitations, we present TELLER: Table Entity Linking through Learning from Errors and Reasoning. We first retrieve and rank Wikidata candidates and retain reduced table evidence in the prompt. The direct-answer path applies iterative direct preference optimization and refreshes its preference data with residual errors from the updated model. The reasoning path uses filtered and compressed chain-of-thought rationales for supervised fine-tuning, followed by our iterative length-normalized regularized preference optimization. On the TableInstruct entity-linking subset, the direct-answer path improves accuracy from 94.35\% to 94.50\%; on the MammoTab V2 evaluation set, it improves accuracy from 87.59\% to 88.20\%. The reasoning path improves accuracy from 92.90\% to 92.95\% on TableInstruct and from 79.09\% to 81.85\% on MammoTab V2, while maintaining high rates of complete reasoning generation. These results show that iterative preference learning benefits both concise entity prediction and explicit reasoning.

Figures

Figures reproduced from arXiv: 2607.28680 by Kehao Li, Stefan Decker, Yixin Peng.

Figure 1
Figure 1. Figure 1: Overview of the two training paths. The direct-answer path starts from the SFT checkpoint with the best validation performance and applies iterative DPO. The reasoning path continues the same SFT checkpoint with CoT-SFT and then applies iterative L-RPO. In both paths, the model obtained after round 𝑘 − 1 is used to collect residual errors and construct fresh preference pairs for round 𝑘. This stage teaches… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 3 linked inside Pith

  1. [1]

    M. M. Katsakioris, Y. Zhou, D. Masato, Entity linking in tabular data needs the right attention, arXiv preprint arXiv:2207.01937 (2022)

    Pith/arXiv arXiv 2022

  2. [2]

    De Cao, G

    N. De Cao, G. Izacard, S. Riedel, F. Petroni, Autoregressive entity retrieval, in: Proceedings of the 38th International Conference on Machine Learning, volume 139 ofProceedings of Machine Learning Research, PMLR, 2021, pp. 1931–1942. URL: https://proceedings.mlr.press/v139/de-cao21a.html

    2021

  3. [3]

    Z. Xiao, M. Gong, J. Wu, X. Zhang, L. Shou, D. Jiang, Instructed language models with retrievers are powerful entity linkers, in: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023, pp. 2267–2282

    2023

  4. [4]

    Y. Ding, A. Poudel, Q. Zeng, T. Weninger, B. Veeramani, S. Bhattacharya, Entgpt: Entity linking with generative large language models, arXiv preprint arXiv:2402.06738 (2024)

    Pith/arXiv arXiv 2024

  5. [5]

    Zhang, X

    T. Zhang, X. Yue, Y. Li, H. Sun, Tablellama: Towards open large generalist models for tables, in: Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 2024, pp. 6024–6044

    2024

  6. [6]

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, et al., Chain-of-thought prompting elicits reasoning in large language models, Advances in neural information processing systems 35 (2022) 24824–24837

    2022

  7. [7]

    Wiegreffe, J

    S. Wiegreffe, J. Hessel, S. Swayamdipta, M. Riedl, Y. Choi, Reframing human-ai collaboration for generating free-text explanations, in: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2022, pp. 632–658

    2022

  8. [8]

    L. H. Li, J. Hessel, Y. Yu, X. Ren, K.-W. Chang, Y. Choi, Symbolic chain-of-thought distillation: Small models can also “think” step-by-step, in: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2023, pp. 2665–2679

    2023

  9. [9]

    Rafailov, A

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, C. Finn, Direct preference optimiza- tion: Your language model is secretly a reward model, Advances in neural information processing systems 36 (2023) 53728–53741

    2023

  10. [10]

    Xiong, H

    W. Xiong, H. Dong, C. Ye, Z. Wang, H. Zhong, H. Ji, N. Jiang, T. Zhang, Iterative preference learning from human feedback: Bridging theory and practice for RLHF under KL-constraint, in: Proceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, PMLR, 2024, pp. 54715–54754. URL: https://proc...

    2024

  11. [11]

    R. Park, R. Rafailov, S. Ermon, C. Finn, Disentangling length from quality in direct preference optimization, in: Findings of the Association for Computational Linguistics: ACL 2024, 2024, pp. 4998–5017

    2024

  12. [12]

    Y. Meng, M. Xia, D. Chen, Simpo: Simple preference optimization with a reference-free reward, Advances in Neural Information Processing Systems 37 (2024) 124198–124235

    2024

  13. [13]

    Vrandečić, M

    D. Vrandečić, M. Krötzsch, Wikidata: a free collaborative knowledgebase, Communications of the ACM 57 (2014) 78–85

    2014

  14. [14]

    Z. Liu, M. Lu, S. Zhang, B. Liu, H. Guo, Y. Yang, J. Blanchet, Z. Wang, Provably mitigating overoptimization in rlhf: Your sft loss is implicitly an adversarial regularizer, Advances in Neural Information Processing Systems 37 (2024) 138663–138697

    2024

  15. [15]

    Cremaschi, F

    M. Cremaschi, F. Belotti, J. D’Souza, M. Palmonari, Mammotab 25: A large-scale dataset for semantic table interpretation-training, testing, and detecting weaknesses, in: International Semantic Web Conference, Springer, 2025, pp. 131–148

    2025

  16. [16]

    MammoTab Team, MammoTab V2: Evaluation Dataset and Results, MammoTab Documentation,

  17. [17]

    L. Wu, F. Petroni, M. Josifoski, S. Riedel, L. Zettlemoyer, Scalable zero-shot entity linking with dense entity retrieval, in: Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), 2020, pp. 6397–6407

    2020

  18. [18]

    X. Liu, Y. Liu, K. Zhang, K. Wang, Q. Liu, E. Chen, Onenet: A fine-tuning free framework for few-shot entity linking via large language model prompting, in: Proceedings of the 2024 conference on empirical methods in natural language processing, 2024, pp. 13634–13651

    2024

  19. [19]

    Y. Li, A. Galimov, M. D. Ganapaneni, P. Thejaswi, D. Meng, P. Kumar, S. Potdar, Leveraging the power of large language models in entity linking via adaptive routing and targeted reasoning, in: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, 2025, pp. 871–882

    2025

  20. [20]

    X. Deng, H. Sun, A. Lees, Y. Wu, C. Yu, TURL: Table understanding through representation learning, Proceedings of the VLDB Endowment 14 (2020) 307–319. URL: https://www.vldb.org/ pvldb/vol14/p307-deng.pdf

    2020

  21. [21]

    Y. Wang, Y. Fan, J. Guo, R. Zhang, X. Cheng, Rocel: Advancing table entity linking through distinctive row and column contexts, in: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024, pp. 15284–15298

    2024

  22. [22]

    N. Ho, L. Schmid, S.-Y. Yun, Large language models are reasoning teachers, in: Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers), 2023, pp. 14852–14882

    2023

  23. [23]

    L. C. Magister, J. Mallinson, J. Adamek, E. Malmi, A. Severyn, Teaching small language models to reason, in: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 2023, pp. 1773–1781

    2023

  24. [24]

    Hsieh, C.-L

    C.-Y. Hsieh, C.-L. Li, C.-K. Yeh, H. Nakhost, Y. Fujii, A. Ratner, R. Krishna, C.-Y. Lee, T. Pfister, Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes, in: Findings of the Association for Computational Linguistics: ACL 2023, 2023, pp. 8003–8017

    2023

  25. [25]

    H. Chen, S. Wu, X. Quan, R. Wang, M. Yan, J. Zhang, Mcc-kd: Multi-cot consistent knowledge distillation, in: Findings of the Association for Computational Linguistics: EMNLP 2023, 2023, pp. 6805–6820

    2023

  26. [26]

    P. Wang, Z. Wang, Z. Li, Y. Gao, B. Yin, X. Ren, Scott: Self-consistent chain-of-thought distillation, in: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2023, pp. 5546–5558

    2023

  27. [27]

    Zelikman, Y

    E. Zelikman, Y. Wu, J. Mu, N. Goodman, Star: Bootstrapping reasoning with reasoning, Advances in Neural Information Processing Systems 35 (2022) 15476–15488

    2022

  28. [28]

    Turpin, J

    M. Turpin, J. Michael, E. Perez, S. Bowman, Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting, Advances in Neural Information Processing Systems 36 (2023) 74952–74965

    2023

  29. [29]

    H. Chae, Y. Song, K. Ong, T. Kwon, M. Kim, Y. Yu, D. Lee, D. Kang, J. Yeo, Dialogue chain-of-thought distillation for commonsense-aware conversational agents, in: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023, pp. 5606–5632

    2023

  30. [30]

    K. Feng, C. Li, X. Zhang, J. Zhou, Y. Yuan, G. Wang, Keypoint-based progressive chain-of-thought distillation for LLMs, in: Proceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, PMLR, 2024, pp. 13241–13255. URL: https://proceedings.mlr.press/v235/feng24k.html

    2024

  31. [31]

    Wadhwa, S

    S. Wadhwa, S. Amir, B. C. Wallace, Investigating mysteries of cot-augmented distillation, in: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024, pp. 6071–6086

    2024

  32. [32]

    H. Xia, C. T. Leong, W. Wang, Y. Li, W. Li, TokenSkip: Controllable chain-of-thought compression in LLMs, in: C. Christodoulopoulos, T. Chakraborty, C. Rose, V. Peng (Eds.), Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Suzhou, China, 2025, pp. 3351–3363. URL: https://acl...

    doi:10.18653/v1/2025.emnlp-main.165 2025

  33. [33]

    R. Y. Pang, W. Yuan, K. Cho, H. He, S. Sukhbaatar, J. Weston, Iterative reasoning preference optimization, Advances in Neural Information Processing Systems 37 (2024) 116617–116637

    2024

  34. [34]

    Accessed: 2026-07-24

    Wikidata, Wikidata: Database download, https://www.wikidata.org/wiki/Wikidata:Database_ download, 2026. Accessed: 2026-07-24

    2026

  35. [35]

    Robertson, H

    S. Robertson, H. Zaragoza, The probabilistic relevance framework: BM25 and beyond, volume 4, Now Publishers Inc, 2009

    2009

  36. [36]

    A. Xu, B. Lin, B. Xue, B. Wang, B. Xu, B. Wu, B. Zhang, C. Lin, C. Dong, C. Ling, et al., Deepseek-v4: Towards highly efficient million-token context intelligence, arXiv preprint arXiv:2606.19348 (2026)

    arXiv 2026

  37. [37]

    URL: https://cdn.openai.com/pdf/3a4153c8-c748-4b71-8e31-aecbde944f8d/oai_5_2_system-card.pdf

    OpenAI, Update to the GPT-5 System Card: GPT-5.2, Technical Report, OpenAI, 2025. URL: https://cdn.openai.com/pdf/3a4153c8-c748-4b71-8e31-aecbde944f8d/oai_5_2_system-card.pdf

    2025

  38. [38]

    Grattafiori, A

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, et al., The llama 3 herd of models, arXiv preprint arXiv:2407.21783 (2024)

    Pith/arXiv arXiv 2024

  39. [39]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, et al., Lora: Low-rank adaptation of large language models., Iclr 1 (2022) 3

    2022

  40. [40]

    Avogadro, M

    R. Avogadro, M. Ciavotta, F. De Paoli, M. Palmonari, D. Roman, Estimating link confidence for human-in-the-loop table annotation, in: 2023 IEEE/WIC International Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT), IEEE, 2023, pp. 142–149. A. Data Examples This appendix presents representative serialized records from the datasets used...

    2023

  41. [2026]

    URL: https://unimib-datai.github.io/mammotab-docs/, accessed: 2026-07-23

    2026

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.