Pith. sign in

REVIEW 4 major objections 5 minor 32 references

Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Cognitive biases in LLMs are set during pretraining, not finetuning.

desk verdict Solid experimental work with a genuinely reusable cross-tuning protocol, but the headline causal claim overreaches because pretraining identity is confounded with model architecture. read the letter →

arxiv 2507.07186 v2 pith:CIKQA6US submitted 2025-07-09 cs.CL cs.AIcs.LG

Itay Itzhak , Yonatan Belinkov , Gabriel Stanovsky This is my paper
classification cs.CLcs.AIcs.LG
keywords cognitivebiasespretraininginstructionfinetuningcross-tuningbiasvectorstrainingrandomnessLLMbehaviorclusteringanalysis
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

This paper asks where the systematic irrationalities known as cognitive biases come from in large language models. It argues, with a two-step causal experiment, that the largest share of these biases is established during pretraining, so models that share a pretrained base show similar bias footprints even after being tuned on very different instruction datasets. Instruction tuning and random-seed variation shift bias magnitudes in places but do not override the base pattern. The practical stake is that efforts to debias models should target pretraining data and pipeline choices, not only post-hoc alignment.

What carries the argument

Cross-tuning: a 2x2 factorial experiment in which two pretrained models are each finetuned on both instruction datasets rather than only their own. Bias vectors, one score per bias or one score per scenario, are then clustered under labels for pretraining identity, instruction data, random labels, and K-Means. The load-bearing comparison is that pretraining-labeled clusters beat instruction-labeled and unsupervised clusters on silhouette, Calinski-Harabasz, and Davies-Bouldin indices.

What would settle it

Train two models with identical architecture and tokenizer on deliberately different pretraining corpora, finetune both on the same instruction data, and cluster their bias vectors; if bias patterns cluster by architecture rather than by corpus, the paper's attribution of bias to pretraining content would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that cognitive biases in LLMs are planted in pretraining and only swayed, not set, by finetuning. The paper supports this by finetuning OLMo-7B and T5-11B on each other's instruction data (cross-tuning), repeating each run over three random seeds, and measuring 32 biases. Bias vectors group by pretrained model identity far more tightly than by instruction dataset, outperforming even unsupervised clustering, and the same pattern replicates on community-finetuned Llama2 and Mistral models. Training randomness adds noise, but averaged seed scores reproduce the bias directions of fully finetuned models, suggesting that the underlying latent bias survives the stochasticity of tuning.

Load-bearing premise

The paper assumes that grouping models by their pretrained identity isolates the effect of pretraining, but the compared models differ in architecture, tokenizer, and data all at once, so the visible grouping could come from fixed architectural traits rather than from what pretraining taught them.

Editorial extensions

If this is right

  • Instruction tuning should be viewed as adjusting the expression of pre-existing biases, not as the origin of most bias patterns.
  • Bias evaluations of finetuned models are partly evaluations of the base model; comparing tuned models across different base models conflates pretraining with tuning effects.
  • Mitigation targeted at post-training alignment alone is unlikely to remove deeply planted biases.
  • Aggregating bias scores over several random seeds recovers stable per-model bias directions.
  • Pretraining-stage interventions, such as data curation or early bias monitoring, become the more direct lever for changing bias.

Reading between the lines

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

  • If the pretraining attribution generalizes, then for a fixed architecture, models pretrained on different corpora should show bias patterns that track the corpus; this is a test the paper's design cannot fully run because the compared models differ in architecture, tokenizer, and data at once.
  • Bias-vector clustering could serve as a cheap diagnostic for how much any alignment recipe inherits from its base model before large-scale tuning is undertaken.
  • The 'planted' framing suggests a testable prediction: bias-direction profiles should be predictable from statistics of pretraining corpora, such as the prevalence of framing language or anchoring examples.
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

4 major / 5 minor

Summary. The paper proposes a two-step causal framework to attribute cognitive biases in LLMs to pretraining, instruction finetuning, or training randomness. Using OLMo-7B and T5-11B finetuned with LoRA on Flan and Tulu across three seeds, plus community-finetuned Llama2/Mistral models, it measures 32 biases, computes bias vectors, and compares clustering by pretraining identity versus instruction dataset. The authors conclude that pretraining is the primary origin of cognitive biases, with instruction tuning and seed randomness playing smaller, secondary roles.

Significance. The paper is valuable as a case study: it releases code and models, uses an external bias benchmark (Malberg et al.), runs multiple seeds, and introduces the cross-tuning design. If the causal claim were fully supported, it would refocus mitigation efforts on pretraining. However, the central claim is weakened by the architecture/pretraining confound and by selection on outcome, so the paper's main contribution is the experimental template and the demonstration that bias patterns track base model identity within this small set of models.

major comments (4)
  1. [§3.1, Eq. (2); §5.2, Table 2] The central inference from 'models with the same pretrained backbone cluster together' to 'biases originate in pretraining data or process' is not licensed, because OLMo-7B and T5-11B differ simultaneously in architecture (decoder-only vs. encoder-decoder), tokenizer, training objective, and pretraining corpus; the same is true for Llama2-7B vs. Mistral-7B in Table 8. Fixed architectural inductive biases could reproduce the clustering even if the pretraining corpus had no causal role. The Limitations section explicitly acknowledges the architecture difference, but no matched-architecture control is run; the evidence therefore supports 'bias patterns track the base model type', not the stronger causal claim.
  2. [§3.1, Table 1] OLMo and T5 were selected because their bias trends are opposed (Eq. 2), using the same biases (certainty effect, belief valid, belief invalid) that are later used in the cross-tuning analysis; this selection on the outcome can inflate the apparent separation between pretraining groups and is not accounted for in the clustering or significance tests.
  3. [§5.2, Table 2] The 14 data points in the main clustering analysis are generated from two pretrained checkpoints via three seeds plus the original models, so they are highly non-independent; the effective sample size for the pretraining factor is two, making tight pretraining clusters almost tautological. The external validation (Table 8) has the same issue with six models from two families. The results should be reported as seed-level variability around two backbones, with cluster stability across seeds, rather than as independent replication.
  4. [§3, Figure 1] The causal graph assumes pretraining primarily affects the hypothesized 'Latent Bias' and that training randomness affects only 'Observed Bias'; Step 2 then interprets clustering by pretraining as evidence for this assumption. Because the latent construct is not measured and the assumption is not independently tested, the conclusion that pretraining is the 'leading cause' rests in part on the model it is meant to establish. The paper should either test this assumption (e.g., by comparing same-architecture models with different pretraining data) or explicitly present the results as conditional on it.
minor comments (5)
  1. [§1, §4, Figure 1] Typos should be corrected: 'belief balid' (Section 4), 'fine each model' (Section 1), 'casual effect' and 'possbile' (Figure 1 caption), and 'expended' (Section 4).
  2. [§3.1] The parenthetical 'Section 4 for details).' has an unbalanced parenthesis and the cross-reference should be made more precise.
  3. [Appendix A, Eq. (4)] Equation (4) defines k as +/-1 but does not state how k is chosen for each bias; a sentence or reference explaining the orientation adjustment would improve clarity.
  4. [Table 8] The external-model clustering table reports metrics without the permutation-test significance indicators used in Table 2; adding the same significance test would strengthen the replication claim.
  5. [§3.1, §3.2, §4] The method is written inconsistently as 'LoRa' in some places and 'LoRA' in others; the abbreviation should be standardized.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the cross-tuning and clustering evidence is independent of its inputs.

full rationale

The derivation is self-contained: the main empirical claim is that cross-tuned models group by pretrained backbone rather than by instruction dataset, and this is established by computing bias vectors from a pre-existing 30-bias benchmark (Malberg et al., 2024) plus two inherited bias tests, then comparing pretraining-labeled, instruction-labeled, random, and K-Means clusterings on data that include three fresh random seeds per condition. The cross-tuning outcome is not forced by the choice of models: although OLMo and T5 were selected partly because their original finetuned versions differ, the prediction tested is whether swapping the finetuning data changes those directions, and the observed persistence is a contingent empirical result. The Llama2/Mistral community-model replication is also independent, since those models were not chosen for bias trends. The self-citations to Itzhak et al. (2024) appear in the motivation and model selection, but the load-bearing evidence for the pretraining conclusion is the new cross-tuning and clustering analysis, so the argument does not reduce to a self-citation. The OLMo-vs-T5 architecture/pretraining confound is a real threat to the causal wording, but it is an identification problem, not a circular reduction of the result to its inputs.

Assumptions & free parameters 0 free parameters · 4 assumptions · 1 invented entities

The analysis has no fitted free parameters in the statistical sense; the main load-bearing assumptions are the causal graph, the LoRA approximation, the validity of the bias benchmarks, and the untested claim that pretraining identity is not confounded with architecture.

assumptions (4)
  • domain assumption Pretraining primarily affects a latent bias, while training randomness affects only observed bias outputs.
    Section 3 (Causal graph) explicitly assumes this split, which justifies interpreting seed variability as measurement noise and pretraining identity as the stable source.
  • domain assumption LoRA finetuning with high rank approximates full finetuning well enough to preserve bias trends.
    Section 3.1 and Section 4 use LoRA to reduce cost; they verify via MMLU closing 85% of the gap, but do not verify bias-specific equivalence.
  • domain assumption The bias scores from Malberg et al. (2024) and Itzhak et al. (2024) are valid measures of the intended cognitive biases.
    Section 4 adopts these benchmarks without independent validation of score reliability per bias.
  • ad hoc to paper Models sharing a pretrained backbone differ only in pretraining experience and not in architecture or other fixed properties.
    Section 3.1 and 5.2 interpret clustering by pretraining identity as evidence about pretraining, but OLMo vs T5 and Llama2 vs Mistral differ in architecture and tokenizer, so this assumption is untested and load-bearing.
invented entities (1)
  • Latent Bias
    purpose: Hypothesized internal bias state acquired during pretraining, used to separate stable bias from observed variation.
    Defined in Section 3 as internal behavioral patterns; not directly measured, only inferred from observed bias scores, so it has no falsifiable handle outside the model behavior it is meant to explain.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs." pith.science (2026). https://pith.science/paper/CIKQA6US

@misc{pith2026250707186,
  author       = {Pith},
  title        = {Pith review of: Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CIKQA6US}},
  note         = {Machine review of arXiv:2507.07186}
}
abstract

Large language models (LLMs) exhibit cognitive biases -- systematic tendencies of irrational decision-making, similar to those seen in humans. Prior work has found that these biases vary across models and can be amplified by instruction tuning. However, it remains unclear if these differences in biases stem from pretraining, finetuning, or even random noise due to training stochasticity. We propose a two-step causal experimental approach to disentangle these factors. First, we finetune models multiple times using different random seeds to study how training randomness affects over $30$ cognitive biases. Second, we introduce \emph{cross-tuning} -- swapping instruction datasets between models to isolate bias sources. This swap uses datasets that led to different bias patterns, directly testing whether biases are dataset-dependent. Our findings reveal that while training randomness introduces some variability, biases are mainly shaped by pretraining: models with the same pretrained backbone exhibit more similar bias patterns than those sharing only finetuning data. These insights suggest that understanding biases in finetuned models requires considering their pretraining origins beyond finetuning effects. This perspective can guide future efforts to develop principled strategies for evaluating and mitigating bias in LLMs.

Figures

Figures reproduced from arXiv: 2507.07186 by the authors.

Figure 1
Figure 1. Potential causal relationships between pretraining, instruction finetuning, and bias emergence in LMs. Blue represents possible causes (finetuning, pretraining, and training random￾ness). Purple represents the effect (observed bias). The Grey element denotes a possbile latent cause (latent bias). Dashed arrows (→) indicate possible causal effects and non-dashed indicate known casual effect. Our two-step causal analy… view at source ↗
Figure 2
Figure 2. Illustration of the cross-tuning setup and the potential causes of biases. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Variability of biases and MMLU scores across random seeds: Mean of standard devia￾tions computed across three random seeds, averaged over all 32 biases and 17 MMLU subcategories scores for each model. Bias scores consistently show slightly more variation across seeds than MMLU scores, suggesting a modest sensitivity to training randomness. conduct a small-scale hyperparameter search over learning rate, LoRa rank (64… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Clustering PCA analysis of model bias vectors. Each point represents a model, where the fill color indicates the pretrained model (pink and turquoise), the shape shows the instruction tuning type (Tulu - circle, Flan - triangle, ShareGPT - square), and the edge color r…
Figure 6
Figure 6. Figure 6: Clustering PCA analysis of model bias vectors at the bias level. Each point represents a model, where the fill color indicates the pretrained model (OLMo in pink, T5 in turquoise), the shape shows the instruction tuning type (Tulu - circle, Flan - triangle), and the ed…
Figure 7
Figure 7. Figure 7: Mean bias score per cluster grouped by pretraining model. those observed in our controlled setup ( [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 14 canonical work pages

  1. [1]

    Comparing rationality between large language models and humans: Insights and open questions

    Dana Alsagheer, Rabimba Karanjai, Nour Diallo, Weidong Shi, Yang Lu, Suha Beydoun, and Qiaoning Zhang. Comparing rationality between large language models and humans: Insights and open questions. arXiv preprint arXiv:2403.09798,

    arXiv

  2. [2]

    arXiv preprint arXiv:2311.10702,

    arXiv

  3. [3]

    Using cognitive psychology to understand GPT-3

    URL https://api.semanticscholar. org/CorpusID:271328219. Marcel Binz and Eric Schulz. Using cognitive psychology to understand gpt-3.ArXiv, abs/2206.14576,

    work page Pith review arXiv

  4. [5]

    AGR: Age Group fairness Reward for Bias Mitigation in LLMs

    Shuirong Cao, Ruoxi Cheng, and Zhiqiang Wang. Agr: Age group fairness reward for bias mitigation in llms. ArXiv, abs/2409.04340,

    work page Pith review arXiv

  5. [6]

    Vishnu Kabir Chhabra, Ding Zhu, and Mohammad Mahdi Khalili

    URL https://api.semanticscholar.org/CorpusId:272828139. Vishnu Kabir Chhabra, Ding Zhu, and Mohammad Mahdi Khalili. Neuroplasticity and corruption in model mechanisms: A case study of indirect object identification.arXiv preprint arXiv:2503.01896,

    arXiv

  6. [8]

    semanticscholar.org/CorpusId:268856702

    URL https://api. semanticscholar.org/CorpusId:268856702. Ishita Dasgupta, Andrew K Lampinen, Stephanie CY Chan, Antonia Creswell, Dharshan Kumaran, James L McClelland, and Felix Hill. Language models show human-like content effects on reasoning. arXiv preprint arXiv:2207.07051,

    arXiv

  7. [9]

    Cognitive bias in decision-making with llms

    Jessica Echterhoff, Yao Liu, Abeer Alessa, Julian McAuley, and Zexue He. Cognitive bias in decision-making with llms. arXiv preprint arXiv:2403.00811, 2024a. Jessica Maria Echterhoff, Yao Liu, Abeer Alessa, Julian McAuley, and Zexue He. Cognitive bias in decision-making with LLMs. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.), Findings of th...

    arXiv 2024

  8. [12]

    Dirk Groeneveld, Iz Beltagy, Pete Walsh, Akshita Bhagia, Rodney Kinney, Oyvind Tafjord, Ananya Harsh Jha, Hamish Ivison, Ian Magnusson, Yizhong Wang, et al

    URL https: //doi.org/10.1162/coli_a_00524. Dirk Groeneveld, Iz Beltagy, Pete Walsh, Akshita Bhagia, Rodney Kinney, Oyvind Tafjord, Ananya Harsh Jha, Hamish Ivison, Ian Magnusson, Yizhong Wang, et al. Olmo: Accelerat- ing the science of language models. arXiv preprint arXiv:2402.00838,

    arXiv

Show all 32 references
  1. [14]

    URL https:// aclanthology.org/2024.tacl-1.43/

    doi: 10.1162/tacl a 00673. URL https:// aclanthology.org/2024.tacl-1.43/. Hamish Ivison, Yizhong Wang, Valentina Pyatkin, Nathan Lambert, Matthew Peters, Pradeep Dasigi, Joel Jang, David Wadden, Noah A Smith, Iz Beltagy, et al. Camels in a changing climate: Enhancing lm adapta...

    2024 doi

  2. [16]

    org/abs/2310.06825

    URL https://arxiv. org/abs/2310.06825. Daniel Kahneman. Thinking, fast and slow. macmillan,

    arXiv

  3. [18]

    Ryan Koo, Minhwa Lee, Vipul Raheja, Jong Inn Park, Zae Myung Kim, and Dongyeop Kang

    URL https: //api.semanticscholar.org/CorpusID:267301207. Ryan Koo, Minhwa Lee, Vipul Raheja, Jong Inn Park, Zae Myung Kim, and Dongyeop Kang. Benchmarking cognitive biases in large language models as evaluators. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Findings...

    2024

  4. [19]

    doi: 10.18653/v1/2024.findings-acl.29

    Association for Computational Linguis- tics. doi: 10.18653/v1/2024.findings-acl.29. URL https://aclanthology.org/2024. findings-acl.29/. Gili Lior, Liron Nacchace, and Gabriel Stanovsky. Wildframe: Comparing framing in humans and llms on naturally occurring texts. arXiv prepri...

    2024

  5. [20]

    Cognitive debiasing large language models for decision-making

    Yougang Lyu, Shijie Ren, Yue Feng, Zihan Wang, Zhumin Chen, Zhaochun Ren, and Maarten de Rijke. Cognitive debiasing large language models for decision-making. arXiv preprint arXiv:2504.04141,

  6. [21]

    A comprehensive evaluation of cognitive biases in llms

    Simon Malberg, Roman Poletukhin, Carolin M Schuster, and Georg Groh. A comprehensive evaluation of cognitive biases in llms. arXiv preprint arXiv:2410.15413,

  7. [22]

    Pouya Pezeshkpour, Sarthak Jain, Byron C

    URL http: //dl.acm.org/citation.cfm?id=3597307. Pouya Pezeshkpour, Sarthak Jain, Byron C. Wallace, and Sameer Singh. An empirical comparison of instance attribution methods for nlp. ArXiv, abs/2104.04128,

    arXiv

  8. [24]

    Cbeval: A framework for evaluating and interpreting cognitive biases in llms

    Ammar Shaikh, Raj Abhijit Dandekar, Sreedath Panat, and Rajat Dandekar. Cbeval: A framework for evaluating and interpreting cognitive biases in llms. arXiv preprint arXiv:2412.03605,

    arXiv

  9. [25]

    Lora vs full fine-tuning: An illusion of equivalence

    Reece Shuttleworth, Jacob Andreas, Antonio Torralba, and Pratyusha Sharma. Lora vs full fine-tuning: An illusion of equivalence. arXiv preprint arXiv:2410.21228,

  10. [26]

    Gabriel Stanovsky, Tomasz Limisiewicz, David Marevcek, and Bar Iluz

    URL https: //api.semanticscholar.org/CorpusId:235097394. Gabriel Stanovsky, Tomasz Limisiewicz, David Marevcek, and Bar Iluz. Exploring the impact of training data distribution and subword tokenization on gender bias in machine translation. ArXiv, abs/2309.12491,

    arXiv

  11. [27]

    URL https://api.semanticscholar.org/CorpusId: 262217528. 13 Published as a conference paper at COLM 2025 Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: ...

    2025 arXiv

  12. [29]

    Language models are susceptible to incorrect patient self- diagnosis in medical applications

    Rojin Ziaei and Samuel Schmidgall. Language models are susceptible to incorrect patient self- diagnosis in medical applications. arXiv preprint arXiv:2309.09362,

    arXiv

  13. [30]

    (2024) and Itzhak et al

    14 Published as a conference paper at COLM 2025 A Bias Data and Bias Scores We evaluate model bias using the data and bias scores from Malberg et al. (2024) and Itzhak et al. (2024). Below, we briefly describe the biases, datasets, and evaluation metrics. For more details on d...

    2024

  14. [31]

    Which allocation level do you choose for this purpose?

    A.2 Datasets Structure The structure of the dataset by Malberg et al. (2024) is based on a modular framework designed to scale the evaluation of cognitive biases in LLMs while maintaining a consistent test logic grounded in psychological research. At its core, each test follow...

    2024

  15. [32]

    These results verify that our finetuning setting is good enough to create models that can simulate fully finetuned models, especially regarding the bias score trends

    The MMLU scores of our models close over 85% of the performance gap between the original base models (28.6 for OLMo, 23.0 for T5) and their fine-tuned counterparts. These results verify that our finetuning setting is good enough to create models that can simulate fully finetun...

    2024

  16. [1981]

    Towards understanding fine-tuning mechanisms of llms via circuit analysis

    Xu Wang, Yan Hu, Wenyu Du, Reynold Cheng, Benyou Wang, and Difan Zou. Towards understanding fine-tuning mechanisms of llms via circuit analysis. In ICLR 2025 Workshop on Building Trust in Language Models and Applications,

    2025

  17. [1983]

    Improving gender fairness of pre- trained language models without catastrophic forgetting

    Zahra Fatemi, Chen Xing, Wenhao Liu, and Caiming Xiong. Improving gender fairness of pre- trained language models without catastrophic forgetting. ArXiv, abs/2110.05367,

    arXiv

  18. [2011]

    Enhancing diagnostic accuracy through multi-agent conversations: Using large language models to mitigate cognitive bias

    12 Published as a conference paper at COLM 2025 Yuhe Ke, Rui Yang, Sui An Lie, Taylor Xin Yi Lim, Hairil Rizal Bin Abdullah, Daniel Shu Wei Ting, and Nan Liu. Enhancing diagnostic accuracy through multi-agent conversations: Using large language models to mitigate cognitive bia...

    2025 arXiv

  19. [2020]

    Addressing cognitive bias in medical language models

    Samuel Schmidgall, Carl Harris, Ime Essien, Daniel Olshvang, Tawsifur Rahman, Ji Woong Kim, Rojin Ziaei, Jason Eshraghian, Peter Abadir, and Rama Chellappa. Addressing cognitive bias in medical language models. arXiv preprint arXiv:2402.08113,

    arXiv

  20. [2021]

    Isabel O

    URL https://www.aclanthology.org/2023.acl-short.108.pdf. Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md. Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen Ahmed. Bias and fairness in large lan- guage models: A survey. Computational Linguisti...

    2023

  21. [2022]

    Explanation sensitivity to the randomness of large language models: the case of journalistic text classification

    Jeremie Bogaert, Marie-Catherine de Marneffe, Antonin Descampe, Louis Escouflaire, Cedrick Fairon, and Francois-Xavier Standaert. Explanation sensitivity to the randomness of large language models: the case of journalistic text classification. arXiv preprint arXiv:2410.05085,

    arXiv

  22. [2023]

    11 Published as a conference paper at COLM 2025 Zhibo Chu, Zichong Wang, and Wenbin Zhang

    URL https: //lmsys.org/blog/2023-03-30-vicuna/ . 11 Published as a conference paper at COLM 2025 Zhibo Chu, Zichong Wang, and Wenbin Zhang. Fairness in large language models: A taxonomic survey. ACM SIGKDD Explorations Newsletter , 26:34 – 48,

    2023

  23. [2024]

    Generalization v.s

    Antonis Antoniades, Xinyi Wang, Yanai Elazar, Alfonso Amayuelas, Alon Albalak, Kexun Zhang, and William Yang Wang. Generalization v.s. memorization: Tracing language models’ capabilities back to pretraining data. ArXiv, abs/2407.14985,

    arXiv

  24. [2025]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685,

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.