Pith. sign in

REVIEW 3 major objections 6 minor 39 references

Wnuan: Staged Post-Training for Question Answering over Proprietary Enterprise Knowledge

T0 review · 3 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Three-stage post-training (document-to-QA, replay SFT, residual-error RL) raises closed-book enterprise QA from 52.76% to 91.51% acceptable answers.

desk verdict Credible, transparent empirical recipe; the residual-error sampling result is the new part, but the shared judge for selection and evaluation is the load-bearing caveat. read the letter →

arxiv 2608.01862 v1 pith:Q2QH7G3C submitted 2026-08-03 cs.AI

Xiaofeng Shi , Xiaosong Qiu , Wenxin Ma , Qian Kou , Yiming Pan , Longbin Yu , Ying Liu , Haiping Wang
show 1 more author
Hua Zhou
This is my paper
classification cs.AI
keywords enterprisequestionansweringpost-trainingpipelinedocument-to-QAconstructiongeneral-datareplayGRPOresidual-errorsamplingacceptable-answerrateinstruction-followingretention
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

To teach a general 32B language model a proprietary document set without discarding general ability, the paper proposes a three-stage recipe: document-to-QA supervision, supervised fine-tuning with general-data replay, and reinforcement learning (GRPO) applied only to the examples the SFT model still gets wrong. On a 707-question in-house benchmark, acceptable-answer rate rises from 52.76% at base to 80.06% after SFT and 91.51% after RL. The paper further claims that under a matched 100-update budget, residual-error sampling beats full-pool and size-matched random sampling by about three points, with source-cluster bootstrap intervals above zero. The general-capability cost is a 5.17-point average drop, concentrated in instruction following. If these results generalize, the recipe is a practical way to build closed-book enterprise QA on top of an open-weights base model.

What carries the argument

The key mechanism is the residual-error selection rule (Equation 2 in the paper): after SFT, the model answers every question in the selection pool; the examples judged incorrect by the automatic correctness ensemble form the RL training set. The paper's controlled comparison then holds the update count, optimizer, prompt, and reward signal fixed across three arms (residual errors, full pool, size-matched random) to isolate the effect of concentrating the budget on errors. The pipeline as a whole is the unit of value: document-to-QA construction, replay-SFT, and residual-GRPO each have a distinct role, and the paper is explicit that only the Stage-III comparison is a matched causal contrast.

What would settle it

Have two independent domain experts, who were not involved in calibration, label the RL checkpoint's answers on a stratified sample of the 707 benchmark questions, then compare their labels to the automatic ensemble. If binary agreement falls materially below the 90.5% reported for the SFT checkpoint, or if the residual-vs-random AAR gap reverses under the experts' labels, the central claim is not confirmed.

Watch

Extended reading notes

Core claim

The central claim is that residual-error selection is an effective data policy in the RL stage of enterprise post-training. After SFT fixes the easy cases, the paper selects the 56,147 questions that the SFT checkpoint still answers incorrectly, according to the same automatic three-judge correctness ensemble used for evaluation, and trains with GRPO using a semantic reward. In a controlled experiment with 100 updates per arm, this residual pool reaches 89.39% AAR versus 86.28% for full-pool sampling and 86.42% for size-matched random sampling; both contrasts exceed 2.9 points and have source-cluster bootstrap intervals above zero, and the ordering is preserved on a separate same-domain vali

Load-bearing premise

The load-bearing premise is that the automatic three-judge correctness ensemble produces trustworthy labels: every headline AAR number and every residual-vs-baseline contrast is measured by that judge, and the calibration is limited to 147 responses of the SFT checkpoint, with hallucination-detection recall of only 0.657; if the judge rewards style or template compliance rather than content, both the absolute gains and the 3-point contrasts could be artifacts of the rubric.

Editorial extensions

If this is right

  • If the primary trajectory is reproducible, a 32B model with about 91% AAR on this enterprise corpus is attainable without retrieval at inference time.
  • Residual-error sampling under a fixed budget gains about 3 AAR points over both full-pool and size-matched random selection, and the ordering holds on a same-domain validation set.
  • The general-capability cost is mostly instruction following: IFEval falls 6.52 points from Inst to RL while MMLU rises 0.61 and C-Eval rises 2.29, so the trade-off is not uniform across capabilities.
  • Adding fixed retrieval context is non-additive after adaptation: it helps the base model but hurts the SFT and RL checkpoints, which argues for confidence-gated retrieval rather than unconditional concatenation.
  • A further residual-focused RL continuation from the RL checkpoint does not improve AAR and worsens hallucination and IFEval, so residual selection is not universally the right policy at every stage.

Reading between the lines

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

  • If the judge's calibration transfers from the SFT checkpoint to the RL checkpoint, the residual-vs-random gap would likely be reproduced with other base models or corpora; a direct way to check is to audit a stratified sample of RL responses with expert labels.
  • The paper's fixed-prediction target-sensitivity audit shows that changing reference targets moves about 9,000 rows in and out of the residual set even though the aggregate rate changes only -0.86 points; this suggests the policy is stable at the aggregate level but individual rows are fragile, so practitioners should monitor membership churn rather than assuming set stability.
  • The concentrated IFEval drop points to a reward-design fix: adding an instruction-following reward term or explicit replay of instruction data during RL could recover the general loss; the paper shows such a continuation without that change does not work.
  • The retrieval non-additivity, if it generalizes, means RAG and post-training should be optimized jointly or gated by confidence; a testable extension is training a model on residual errors with retrieved context to see whether the pipeline can learn when to use external evidence.
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 presents Wnuan, a three-stage post-training pipeline for closed-book enterprise question answering. Stage I converts proprietary documents into QA supervision with target-aligned answer rewriting; Stage II applies full-parameter SFT with general-data replay; Stage III runs GRPO on examples that Wnuan-Inst still answers incorrectly, using a residual-error selection pool. On the 707-question WnuanBench, the 32B route improves acceptable-answer rate (AAR) from 52.76% (base) to 80.06% (SFT) to 91.51% (RL). In a matched 100-update protocol, residual-error sampling beats full-pool and size-matched random sampling by 3.11 and 2.97 AAR points, with source-cluster bootstrap intervals above zero. The paper also reports a same-domain validation replication, a general-benchmark cost concentrated in IFEval, and a train-free retrieval diagnostic. The evaluation uses a three-model automatic ensemble calibrated against one domain expert on 147 Wnuan-Inst responses.

Significance. If the headline numbers are valid, the paper provides a useful, carefully evaluated recipe for enterprise-domain post-training: document-to-QA supervision, replay-SFT, and residual-error RL. The statistical apparatus is unusually thorough for this genre: question-level paired bootstraps, source-cluster resampling, Holm-adjusted McNemar tests, a same-domain validation set, a reference-target sensitivity audit, and release of a synthetic evaluation fixture and analysis scripts. These strengths make the paper's within-protocol comparisons credible as measurements of the automatic judge's labels. The main weakness is that the judge used for residual selection is the same judge used for final evaluation, and human calibration does not cover the RL checkpoint; this is a load-bearing limitation for the absolute AAR levels and for the relative contrasts, because the training signal and the measurement share a common rubric. The paper is transparent about many of its limitations, which helps, but transparency does not by itself resolve the threat to the central claims.

major comments (3)
  1. [§5.1, Table 1; Appendices B.1–B.3] The central absolute and relative results are statements about the automatic three-model ensemble. The same protocol (J_sel, Appendix B.3) selects Stage-III residual errors and produces the formal AAR labels. Calibration against a domain expert is limited to 147 Wnuan-Inst responses; no Wnuan-RL responses appear in the calibration sample, despite the RL checkpoint's different response format (reasoning+tags, Appendix D.2). Hallucination-detection recall is 0.657, so the instrument that reports the largest RL gain (Halluc. 65.91→15.70) misses more than a third of expert-identified hallucinations. Because the RL training reward uses a disjoint judge but the selection and evaluation judge is identical, the pipeline can learn to satisfy the ensemble's idiosyncratic rubric, and the same judge then certifies the improvement. This is not merely a disclosure issue: it directly undermines the qua
  2. [§5.2, Appendix F.4] The paired bootstrap intervals capture only question-sampling variability, not retraining variability. Each of the three data-selection arms is a single run with one seed (the random arm uses seed 42; the residual and full arms do not state a seed but are described as one completed run per configuration). With one training trajectory per arm, the reported 95% CIs (e.g., residual−full +3.11, CI 0.71–5.66) do not reflect uncertainty in the optimization process. The paper acknowledges this indirectly in §5.3 ('correlations are descriptive rather than estimates over retraining variability') but the primary residual-versus-full and residual-versus-random contrasts are presented without this caveat. Multiple seeds per arm, or at least an explicit statement that the intervals are conditional on the single training run, are needed to support the data-policy claim.
  3. [§5.2 and §4.2] The residual-versus-full contrast conflates selection with repetition. The residual pool contains 56,147 rows and the full pool 230,183 rows, both trained for 100 updates, so residual examples are revisited many more times per update than full-pool examples on average. The size-matched random arm controls pool size but not repetition; the full arm controls update count but not per-example exposure. The paper acknowledges the unequal exposure in §4.2 but in §5.2 interprets the gain as 'sampling efficiency' and 'concentrating a fixed update budget on remaining errors.' As stated, the 3.11-point contrast cannot separate the value of selecting errors from the value of repeatedly training on a smaller pool. A repetition-matched control (e.g., sampling from the full pool with replacement to the same per-example exposure) would isolate the selection effect; absent that, the conclusion should be
minor comments (6)
  1. [Appendix B.1] The calibration sample is stratified by the automatic label (50 correct, 50 partial, 50 incorrect) and labeled by one expert. The paper reports post-stratified agreement (90.4%) but does not specify the stratification weights; please provide them so the post-stratified estimate can be reproduced.
  2. [§5.3, Figure 3] The stopping rule is described as 'diminishing increments and compute budget determine the practical stop.' It would help to state explicitly that the reported endpoint at update 327 was chosen without re-running the evaluation, and whether the validation reward was the only online selection signal. The current text is clear that the validation reward was used for monitoring, but the link to the final checkpoint selection could be stated once in one place.
  3. [§2 and References] The related-work section places Wnuan as 'a recipe, not a new synthesis method.' Given that framing, a comparison table of existing pipeline components (document-to-QA, replay, residual-error RL) would help readers locate the incremental contribution. The text is adequate, but the absence of such a table makes the contribution harder to assess.
  4. [Throughout] The paper alternates between 'AAR' and 'Accuracy' for the stored field (e.g., Appendix A.3 says evaluation exports name this field Accuracy). Consider using AAR in all table headers and figure labels; the current usage is consistent but the store-name note may confuse readers who see 'Accuracy' in the raw exports.
  5. [Appendix I, Figure 10] The RAG diagnostic is clearly described as non-causal, but the phrase 'non-additive after SFT and RL' in §5.6 could be read as a general conclusion. Since only one retriever and one concatenation pipeline are tested, I suggest adding 'under this pipeline' to the main-text sentence, as the authors do in the appendix.
  6. [§5.5 and Appendix H] The domain analysis is appropriately labeled descriptive. The statement 'All eight business-domain slices improve numerically' is correct, but Figure 9(a) shows visible differences in effect sizes; adding confidence intervals for these slices would prevent overinterpretation of small-sample improvements.

Circularity Check

0 steps flagged · score 2.0 of 10

Mostly self-contained pipeline paper; the shared judge for residual selection and final evaluation is a disclosed validity limitation, not a derivation-level circularity, and the one overlapping-author citation for the reward design is not load-bearing.

full rationale

The paper's headline numbers (52.76 -> 80.06 -> 91.51 AAR; residual-vs-full 3.11; residual-vs-random 2.97) are empirical measurements on a disjoint 707-question benchmark, not quantities equal to their inputs by construction. The closest concern is Appendix B.3, where the residual-selection judge J_sel and formal evaluation share 'the same adaptive three-model correctness protocol.' This is a real validity limitation: human calibration covers only 147 Wnuan-Inst responses and no Wnuan-RL responses, and WnuanBench shares the source corpus with training. However, the GRPO reward judge is a separate locally deployed Qwen3.5-35B (also Appendix B.3), so the RL is not directly optimizing the evaluation instrument; the residual arm is selected by J_sel but trained on Qwen3.5-35B rewards. Thus the AAR improvements are not forced by construction. The only overlapping-author citation with potential load-bearing status is the adaptation of the semantic reward from MechVQA (Kou et al. 2026) in Equation 3, but the reward is empirically evaluated through the controlled arms, validation trajectories, and general-benchmark measurements; the paper does not appeal to the citation as a theorem or uniqueness result. No ansatz is smuggled in via self-citation, and no known result is renamed. The shared-judge issue is explicitly disclosed in Limitations and Appendices B.1-B.3, which further supports treating it as an evaluation-validity caveat rather than a circular derivation. Overall, the central claims are self-contained against internal and external benchmarks, and the circularity score is low.

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

The measurable claims rest on one enterprise corpus, a private 707-question benchmark, a three-model judge ensemble whose ground truth is a single expert's 147 labels, and disclosed recipe constants. The free parameters are all disclosed; the reward weights and replay ratio have sensitivity or grid evidence, and rewriting and threshold choices do not drive the central claims. The axioms are dominated by domain assumptions about automated supervision and evaluation; the weakest is transferring the Wnuan-Inst-only judge calibration to the RL checkpoint and the remaining 560 benchmark questions. No invented physical or conceptual entities are introduced; WnuanBench is a private benchmark, not a speculative construct.

free parameters (5)
  • Semantic reward weights = 0.6*racc + 0.3*rquality + 0.1*rformat (rquality = mean(rlogic, rprof, rconcise))
    Hand-set coefficients adapted from the authors' MechVQA paper (Eq. 3 and 7). They define the RL optimization target behind the 11.45-point Inst-to-RL gain. A single 0.7/0.2 sensitivity arm differs by only 0.14 points, so the weights are mildly consequential.
  • General replay ratio = 48.3% actual (106,950 of 221,294 domain examples)
    Chosen from a 0/5/25/50/100% nominal grid by maximizing the public general average (MMLU, IFEval, C-Eval), not by domain AAR. The full grid is reported; this is a disclosed, data-driven operating point that sets the paper's main trajectory.
  • GRPO constants = epsilon=0.2, dual-clip C=3, KL beta=1e-2
    Standard values taken from DeepSeekMath and PPO literature (Shao et al. 2024; Schulman et al. 2017); not fitted to these data.
  • Answer-rewriting similarity gate = MiniLM cosine >= 0.8 (74.45% of targets rewritten)
    Configuration threshold for Stage-I target rewriting. Rewriting shows no independent domain-AAR gain (-1.41 points, p=0.423), so this parameter does not drive the central claim.
  • Residual-pool incorrect threshold = ordered median of judge scores equals 0; selects 56,147 of 230,183 rows
    Definitional choice for the residual pool. A fixed-prediction audit against rewritten targets shows 91.69% membership stability and a -0.86 point change in incorrect rate, so the threshold choice is robust to target variation.
assumptions (5)
  • domain assumption The three-model judge ensemble (ordered median of gpt-oss-120b, MiniMax-M2.5, and on-disagreement DeepSeek-V3.2 votes) yields valid acceptability labels for enterprise answers.
    Every AAR number and the residual pool flow from this ensemble. It is calibrated against one domain expert on 147 Wnuan-Inst responses (90.5% binary agreement, kappa 0.796, hallucination recall 0.657) and is not calibrated on Wnuan-RL responses. Appendix B.1-B.3.
  • domain assumption The enterprise documents and the generated QA reference answers are authoritative ground truth for the represented knowledge base.
    Stage I answers are generated by DeepSeek-V3.2 from OCR'd segments and filtered by rule, referent, answerability, and quality checks; no human verification of the 221,294 SFT targets is reported (Appendix C.2). If a meaningful share of references is wrong, every stage optimizes toward those errors.
  • domain assumption WnuanBench performance, despite semantic near-duplicates of training material, is the correct measure of adaptation success.
    458 of 707 benchmark questions have a nearest retained training question at cosine >= 0.90 (median 0.9295). The paper defines the goal as mastery of the represented knowledge base, not source-held-out or cross-enterprise transfer (Section 7, Appendix A.2). This scope is explicit but bounds the claim.
  • standard math The GRPO/PPO surrogate objective, paired bootstrap, and McNemar tests are applied correctly.
    Objective in Equations 4-6 follows Shao et al. 2024 and Schulman et al. 2017 with epsilon 0.2, dual clip C=3, beta 1e-2; statistical tests follow standard practice and are described in Appendix A.4. No re-derivation is attempted.
  • ad hoc to paper Choosing the production Stage-III configuration from WnuanBench endpoints does not materially inflate the reported final contrasts.
    The residual policy for the main run was selected after the controlled arms' update-100 WnuanBench endpoints were evaluated (Table 13, Figure 3 caption), and WnuanBench is later reported as the headline outcome. The same-domain validation replication (residual minus full +3.56, p=0.008; residual minus random +2.67, p=0.068) mitigates this but was itself retrospective (Appendix F.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Wnuan: Staged Post-Training for Question Answering over Proprietary Enterprise Knowledge." pith.science (2026). https://pith.science/paper/Q2QH7G3C

@misc{pith2026260801862,
  author       = {Pith},
  title        = {Pith review of: Wnuan: Staged Post-Training for Question Answering over Proprietary Enterprise Knowledge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q2QH7G3C}},
  note         = {Machine review of arXiv:2608.01862}
}
read the original abstract

Enterprise question answering requires models to acquire proprietary knowledge without discarding general capabilities. We present Wnuan, a three-stage pipeline that constructs task-oriented supervision from documents, performs supervised fine-tuning with general-data replay, and applies reinforcement learning to residual errors. On the 707-question WnuanBench, the primary 32B route raises acceptable-answer rate (AAR) from 52.76% before adaptation to 80.06% after SFT and 91.51% after RL. Under a matched 100-update protocol, residual-error sampling outperforms full-pool and size-matched random sampling by 3.11 and 2.97 points, respectively. Source-cluster bootstrap intervals remain above zero for both contrasts, and a same-domain validation set preserves the ordering. The general-benchmark average decreases by 5.17 points across the route, concentrated in instruction following. The automatic evaluation ensemble agrees with an authoritative domain expert on 90.5% of a stratified Wnuan-Inst response sample. These results characterize both the gains and the general-capability cost of staged enterprise adaptation.

Figures

Figures reproduced from arXiv: 2608.01862 by the authors.

Figure 1
Figure 1. Wnuan training path, primary 32B results, contextual systems, andWnuanBench construction. (a) Enterprise documents [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Core evidence. (a) Closed-book AAR rises from Wnuan-Base to Wnuan-Inst to Wnuan-RL. Intervals resample [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Online validation accuracy-reward trajectory for the main Wnuan-RL run. After the preliminary 100-update experiment [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Design studies preceding Stage III. (a) Task-oriented Document-to-QA training substantially outperforms fixed [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Benchmark coverage and judge calibration. (a) Task-category and business-domain counts. (b) Binary confusion [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Aggregate GRPO diagnostics. (a) Main-run validation reward components. (b) Main-run entropy and absolute PPO [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Post hoc agreement between the Stage-III training monitor and final evaluation. (a) Change in the raw validation [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Capability retention and question-level transitions. (a) MMLU, IFEval, C-Eval, and their unweighted average across [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Business-domain results. (a) Base, Inst, and RL AAR and the RL-minus-Inst change. All eight changes are positive. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: RAG diagnostics. (a) AAR before and after applying the saved BM25+BGE-M3 top-5 traces. (b) Signed changes in [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 1 canonical work pages

  1. [1]

    Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, Y. K. and Wu, Y. and Guo, Daya , journal =

  2. [2]

    2025 , doi =

    Guo, Daya and others , journal =. 2025 , doi =

    2025

  3. [3]

    arXiv preprint arXiv:1707.06347 , year =

    Proximal Policy Optimization Algorithms , author =. arXiv preprint arXiv:1707.06347 , year =

    arXiv

  4. [4]

    Advances in Neural Information Processing Systems , volume =

    Training Language Models to Follow Instructions with Human Feedback , author =. Advances in Neural Information Processing Systems , volume =

  5. [5]

    and Zhang, Hao and Gonzalez, Joseph E

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric P. and Zhang, Hao and Gonzalez, Joseph E. and Stoica, Ion , booktitle =. Judging

  6. [6]

    2023 , doi =

    Liu, Yang and Iter, Dan and Xu, Yichong and Wang, Shuohang and Xu, Ruochen and Zhu, Chenguang , booktitle =. 2023 , doi =

    2023

  7. [7]

    Zhu, Lianghui and Wang, Xinggang and Wang, Xinlong , booktitle =

  8. [8]

    Understanding the Effects of

    Kirk, Robert and Mediratta, Ishita and Nalmpantis, Christoforos and Luketina, Jelena and Hambro, Eric and Grefenstette, Edward and Raileanu, Roberta , booktitle =. Understanding the Effects of

Show all 39 references
  1. [9]

    Mitigating the Alignment Tax of

    Lin, Yong and Lin, Hangyu and Xiong, Wei and Diao, Shizhe and Liu, Jianmeng and Zhang, Jipeng and Pan, Rui and Wang, Haoxiang and Hu, Wenbin and Zhang, Hanning and Dong, Hanze and Pi, Renjie and Zhao, Han and Jiang, Nan and Ji, Heng and Yao, Yuan and Zhang, Tong , booktitle =....

    2024

  2. [10]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and Küttler, Heinrich and Lewis, Mike and Yih, Wen-tau and Rocktäschel, Tim and Riedel, Sebastian and Kiela, Douwe , booktitle =. Retrieval-Augmented Generation f...

  3. [11]

    International Conference on Learning Representations , year =

    Adapting Large Language Models via Reading Comprehension , author =. International Conference on Learning Representations , year =

  4. [12]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics , pages =

    Instruction-Tuned Language Models Are Better Knowledge Learners , author =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics , pages =. 2024 , doi =

    2024

  5. [13]

    2025 , doi =

    Li, Haiyun and Zhang, Jixin and Shen, Hua and Cheng, Ke and Huang, Xiaofeng , journal =. 2025 , doi =

    2025

  6. [14]

    and Liu, Han , booktitle =

    Xu, Ruiyao and Samia, Noelle I. and Liu, Han , booktitle =. 2026 , doi =

    2026

  7. [15]

    Systematic Knowledge Injection into Large Language Models via Diverse Augmentation for Domain-Specific

    Bhushan, Kushagra and Nandwani, Yatin and Khandelwal, Dinesh and Gupta, Sonam and Pandey, Gaurav and Raghu, Dinesh and Joshi, Sachindra , booktitle =. Systematic Knowledge Injection into Large Language Models via Diverse Augmentation for Domain-Specific. 2025 , doi =

    2025

  8. [16]

    and Jain, Naman and Shen, Sheng and Zaharia, Matei and Stoica, Ion and Gonzalez, Joseph E

    Zhang, Tianjun and Patil, Shishir G. and Jain, Naman and Shen, Sheng and Zaharia, Matei and Stoica, Ion and Gonzalez, Joseph E. , booktitle =

  9. [17]

    Xia, Mengzhou and Malladi, Sadhika and Gururangan, Suchin and Arora, Sanjeev and Chen, Danqi , booktitle =

  10. [18]

    Yu, Qiying and Zhang, Zheng and Zhu, Ruofei and Yuan, Yufeng and Zuo, Xiaochen and Yue, Yu and Dai, Weinan and Fan, Tiantian and Liu, Gaohong and Liu, Juncai and Liu, Lingjun and Liu, Xin and Lin, Haibin and Lin, Zhiqi and Ma, Bole and Sheng, Guangming and Tong, Yuxuan and Zha...

  11. [19]

    Proceedings of the 42nd International Conference on Machine Learning , series =

    Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples , author =. Proceedings of the 42nd International Conference on Machine Learning , series =. 2025 , url =

    2025

  12. [20]

    Hard Examples Are All You Need: Maximizing

    Pikus, Benjamin and Tiwari, Pratyush Ranjan and Ye, Burton , journal =. Hard Examples Are All You Need: Maximizing. 2025 , doi =

    2025

  13. [21]

    Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Gao, Chang and Huang, Chengen and Lv, Chenxu and Zheng, Chujie and Liu, Dayiheng and Zhou, Fan and Huang, Fei and Hu, Feng and Ge, Hao and Wei, Haoran and Lin, Huan an...

  14. [22]

    2026 , month = feb, howpublished =

    2026

  15. [23]

    2025 , howpublished =

    Chinese-. 2025 , howpublished =

    2025

  16. [24]

    International Conference on Learning Representations , year =

    Measuring Massive Multitask Language Understanding , author =. International Conference on Learning Representations , year =

  17. [25]

    arXiv preprint arXiv:2311.07911 , year =

    Instruction-Following Evaluation for Large Language Models , author =. arXiv preprint arXiv:2311.07911 , year =

    arXiv

  18. [26]

    2023 , url =

    Huang, Yuzhen and Bai, Yuzhuo and Zhu, Zhihao and Zhang, Junlei and Zhang, Jinghan and Su, Tangjun and Liu, Junteng and Lv, Chuancheng and Zhang, Yikai and Lei, Jiayi and Fu, Yao and Sun, Maosong and He, Junxian , booktitle =. 2023 , url =

    2023

  19. [27]

    2026 , eprint =

    Kou, Qian and Shi, Xiaofeng and Li, Yulin and Qiu, Xiaosong and Wang, Xinyang and Zhou, Hua and Cao, Dongxing , journal =. 2026 , eprint =

    2026

  20. [28]

    2508.10925 , archivePrefix =

    arXiv preprint arXiv:2508.10925 , year =. 2508.10925 , archivePrefix =

    arXiv

  21. [29]

    arXiv preprint arXiv:2605.26494 , year =

    The. arXiv preprint arXiv:2605.26494 , year =. doi:10.48550/arXiv.2605.26494 , url =. 2605.26494 , archivePrefix =

    doi arXiv

  22. [30]

    2512.02556 , archivePrefix =

    arXiv preprint arXiv:2512.02556 , year =. 2512.02556 , archivePrefix =

    arXiv

  23. [31]

    2025 , month = sep, howpublished =

    2025

  24. [32]

    2026 , howpublished =

    2026

  25. [33]

    2026 , month = jun, howpublished =

    2026

  26. [34]

    2026 , month = apr, howpublished =

    2026

  27. [35]

    2026 , month = mar, howpublished =

    Introducing. 2026 , month = mar, howpublished =

    2026

  28. [36]

    and Zaragoza, Hugo , journal =

    Robertson, Stephen E. and Zaragoza, Hugo , journal =. The Probabilistic Relevance Framework:. 2009 , doi =

    2009

  29. [37]

    2024 , eprint =

    Chen, Jianlv and Xiao, Shitao and Zhang, Peitian and Luo, Kun and Lian, Defu and Liu, Zheng , journal =. 2024 , eprint =. doi:10.48550/arXiv.2402.03216 , url =

    2024 doi arXiv

  30. [38]

    2026 , url =

    Xiaofeng Shi and Hua Zhou and Lulu Zhao , title =. 2026 , url =. doi:10.57967/hf/9611 , publisher =

    2026 doi

  31. [39]

    Mastering Complex Control in

    Ye, Deheng and Liu, Zhao and Sun, Mingfei and Shi, Bei and Zhao, Peilin and Wu, Hao and Yu, Hongsheng and Yang, Shaojie and Wu, Xipeng and Guo, Qingwei and Chen, Qiaobo and Yin, Yinyuting and Zhang, Hao and Shi, Tengfei and Wang, Liang and Fu, Qiang and Yang, Wei and Huang, La...

    2020

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.