Crowd-SFT: Crowdsourcing for LLM Alignment

Alex Sotiropoulos; Bhaskar Krishnamachari; Jared Coleman; Linus Lei; Sulyab Thottungal Valapu

REVIEW 4 major objections 6 minor 12 references

Crowd-SFT: Crowdsourcing for LLM Alignment

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Competitive multi-model selection cuts fine-tuning distance by up to 55% and a point reward scheme tracks Shapley contributions.

desk verdict A plausible proof-of-concept for crowdsourced SFT, but the reward-fairness evidence lives in a linear toy model never tied to real fine-tuning. read the letter →

arxiv 2506.04063 v1 pith:L7T35TI3 submitted 2025-06-04 cs.HC cs.DCcs.LG

Alex Sotiropoulos , Sulyab Thottungal Valapu , Linus Lei , Jared Coleman , Bhaskar Krishnamachari This is my paper

classification cs.HCcs.DCcs.LG

keywords LLMalignmentsupervisedfine-tuningcrowdsourcingtournament-basedselectionShapleyvaluespoint-basedrewardsmulti-modelvector-spacemodeling

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 proposes that LLM alignment via supervised fine-tuning can be opened up to a crowd by running an iterative tournament: at each round, groups of users fine-tune separate copies of the model, the copy whose outputs are closest to a target is kept as the next base model, and everyone earns points proportional to their group's rank. The authors claim this competitive structure accelerates convergence relative to single-model fine-tuning, reducing the embedding-space distance to an emotion target by up to 55% after three iterations in their experiments, and that a simple point-based reward correlates well with Shapley values, so participants can be credited fairly without expensive per-person contribution computation. If true, this offers a path toward cheaper, larger-scale, and more diverse human feedback for fine-tuning, with fairness tracked by design.

What carries the argument

The carrying mechanism is the linear update rule in a preference vector space, $M_{\text{candidate},g}(t) = M_t + \delta(C_g(t) - M_t)$, where $C_g(t)$ is the point-weighted centroid of group $g$'s user preference vectors at round $t$. The paper treats both model outputs and user preferences as points in an $n$-dimensional embedding space, defines 'alignment' as Euclidean (or other metric) distance to an expert target, and iterates: group, update each clone toward its centroid, select the clone closest to the target, reward its group by rank. This turns fine-tuning into a competitive, centroid-pulling process that the paper uses to show both convergence and contribution tracking.

What would settle it

Run the same three-iteration multi-model tournament on an actual open LLM using the same fine-tuning transcripts, measure the embedding-space distance of generated responses to the target centroid, and check whether the winning model's distance drops by the predicted 20–55% relative to single-model SFT; if the observed movement does not track the linear-centroid model, the framework's convergence claim is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that introducing selection pressure among multiple fine-tuned variants makes alignment more reliable: with the same 100-sample feedback pool, competitive selection of the best of three clones per iteration reduced mean distance to the target centroid by 20.04% (disappointment), 5.24% (nervousness), and 55.37% (neutral) relative to single-model SFT, and improved all 28 emotions in the corpus. In a separate preference-space simulation using movie-genre preference data, accumulated points from a tournament rank reward correlate with approximate Shapley values, with random and interleaved grouping producing higher correlation than deterministic grouping, meaning the reward system can approximate fair contribution tracking. The paper also reports a trade-off: L2-norm evaluation gives the best convergence distance, while dot-product evaluation gives the best Shapley correlation, and reward fairness degrades as the number of users grows.

Load-bearing premise

All convergence and fairness numbers come from a simulation where fine-tuning is modeled as a straight-line step toward a group's weighted preference centroid, with a step size $\delta$ that is never reported; if real SFT does not move model weights that way, the measured 20–55% improvements and Shapley correlations may not transfer to actual LLMs.

Editorial extensions

If this is right

Multi-model selection reduces target distance in all 28 emotions in the text experiment, not just the three highlighted cases.
A rank-based point reward can stand in for Shapley values, avoiding the combinatorial cost of exact Shapley computation.
Injecting randomness into user grouping improves both convergence and reward fairness compared to deterministic grouping.
The method choice involves a trade-off: L2 norm minimizes convergence distance, while dot product maximizes contribution-tracking accuracy.
Scaling to larger user populations weakens reward fairness, so the point mechanism needs revision before open deployment at scale.

Reading between the lines

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

If the linear-centroid update is a faithful proxy, the reported 55% improvement should be sensitive to the step size $\delta$; varying $\delta$ across simulations would map how much the headline number depends on this un-reported parameter.
The same tournament-with-points design could be lifted into RLHF by treating the reward model, not the policy, as the object being crowd-updated, an extension the paper names but does not test.
The Shapley-correlation decline with user count suggests that orthogonalization, clustering, or stratified sampling of users may be needed to keep contribution estimates fair at internet scale.
The preference-space framing is general enough that the mechanism could be tested in non-text settings, such as community-driven recommendation or content moderation, where a defined target vector exists.

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.

Desk Editor's Note

read the letter

Here's my read. The genuinely new thing is the packaging: tournament-based multi-model selection for SFT combined with a point-reward scheme, and then checking the reward scheme against Shapley values. Nobody has quite put those together in an SFT context. The Section III experiments actually fine-tune real models on GoEmotions data and show that competitive selection reduces distance to the target compared to single-model fine-tuning. That's a real empirical artifact, though the effect is uneven: the headline 55% is for the neutral emotion, while disappointment and nervousness are 20% and 5%.

The soft spots are where the reward mechanism lives. Experiment 2 is a simulation in which 'fine-tuning' is defined as a linear step toward the group's preference centroid (Section IV-B). That update rule is asserted, not derived, and δ is never reported. Real SFT is gradient descent on text, and nothing in the paper connects the MovieLens simulation to that process. So the central fairness claim—that the point system tracks contributions to LLM alignment—is not actually supported by the evidence. It is supported only in the toy model.

Also, the Shapley comparison is under-specified: the characteristic function over user subsets is never defined, so the reported Pearson correlations cannot be independently reproduced. Lacking that, the correlation could be partly built in if the Shapley game uses the same distance metric as the point ranking. The paper also reports no statistical tests, though there are 50 runs.

To be fair, the authors know some of this. Section VI explicitly says real-world validation is a critical next step and notes the assumption of honest, consistent participation. That's the right spirit.

For a reader: this is a workshop-grade proof of concept. The idea is worth thinking about, and the paper is clear about what it did and didn't do. But I wouldn't cite the simulation results as evidence about LLM alignment without a lot more work. A good referee could push them to define the Shapley game, calibrate the linear update, and run at least one small real crowdsourcing pilot. I'd send it to review, but with low expectations.

Referee Report

4 major / 6 minor

Summary. The paper proposes Crowd-SFT, an iterative crowdsourced fine-tuning framework in which users are partitioned into groups, each group fine-tunes a model instance, the best model advances via an evaluation function, and users receive point rewards based on their group's rank. Experiment 1 (Section III) measures LLM fine-tuning as movement in a GoEmotions-derived emotion embedding space and reports that a competitive multi-model selection strategy reduces mean distance to an emotion target by up to 55.37% relative to single-model fine-tuning. Experiment 2 (Section IV) replaces LLMs with a MovieLens-based simulation that updates a model point by δ toward a group's weighted preference centroid, and reports Pearson correlations between accumulated user points and estimated Shapley values across grouping and evaluation methods. The paper concludes that the framework yields fair, scalable contribution tracking and improved convergence.

Significance. If the claims held, the framework would address a real bottleneck in SFT and RLHF annotation: the cost and homogeneity of small annotator pools. The contribution-tracking idea (tournament selection plus Shapley-correlated point rewards) is timely and the authors make several good-faith empirical gestures: they use a real LLM and a real emotion classifier in Section III, they provide the code at a public repository, and Section VI acknowledges the lack of real-world validation. However, the central fairness claim is validated only inside a linear-centroid simulation that is never calibrated to actual SFT, and the headline convergence claim is not controlled for training compute. The paper is best read as a proof-of-concept whose central claims require substantial additional evidence before they can be accepted as applying to LLM alignment.

major comments (4)

[Section IV-B, Eq. (1)] The model update M_candidate,g(t) = M_t + δ(C_g(t) − M_t) is asserted without derivation or calibration; δ is never reported or varied. Because all reward-fairness results (Pearson correlations in Section IV-C and Figures 6–8) are produced inside this linear-centroid model, the Shapley-correlation evidence is not evidence about LLM fine-tuning unless this update is shown to faithfully represent SFT dynamics. A concrete fix is to calibrate the simulation to the distance trajectories measured in Section III-B, or to replace the assumed update with an empirical one derived from fine-tuning checkpoints, and to report sensitivity of the Pearson correlations to δ.
[Section III-B] The claimed 20.04%, 5.24%, and 55.37% reductions are not controlled for total training compute or number of updates. The single-model baseline is one fine-tuning run of 100 samples (Section III-A), while the multi-model condition uses three clones per iteration for three iterations, removing the winner's samples after each round; this means the multi-model condition performs more fine-tuning runs and sees a different data schedule. To support the attribution to 'selection pressure' rather than to extra compute or data splitting, the authors should report epochs, total sample-updates, and include a same-compute control (e.g., selecting the best of three independent single runs without iterative advancement).
[Section IV-B, step 6] The Shapley ground truth is not defined. The authors state that 'we compute the estimated Shapley values for each user and compare them to their accumulated points,' but the characteristic function v(S) over user subsets is never specified, nor is it explained how KernelSHAP is applied to this game. Without v(S), the reported Pearson correlations in Figures 6–8 cannot be reproduced or interpreted. Moreover, because the point rewards are based on the same distance-to-expert ranking used to define the implied game, the correlation may be partly built into the setup; the authors should define v(S) explicitly and, ideally, validate the point rule against a Shapley value computed from a separate, held-out measure of model quality.
[Sections III-B and IV-C] No statistical tests are reported. The headline improvements and the grouping/evaluation method rankings are based on means over 50 prompts or 50 simulation runs, with no confidence intervals or hypothesis tests; Table I's winning percentages do not convey whether the differences are meaningful. The authors should provide significance tests and effect sizes for the central comparisons, including the claimed 55% reduction and the Dot Product versus L2 Norm trade-off.

minor comments (6)

[Abstract and Section III-B] The abstract's '55% reduction' refers to the best emotion (neutral) at the third iteration; please report the median and range across all 28 emotions and qualify the claim accordingly.
[Abstract] The phrase 'over single-model selection' should be 'over single-model fine-tuning,' since the baseline in Section III-A has no selection mechanism.
[Section IV-B] The number of users, number of groups, number of rounds, and MovieLens preprocessing (normalization, genres used) should be stated in the setup rather than only in figure captions; '50 users and 3 groups' appears only in the Figure 6 caption.
[Section IV-B, Eq. (2)] The notation Δ_g = m − argsort(E_method(M_candidate,g(t), E)) is ambiguous; define whether argsort returns an ascending or descending rank and clarify how ties are broken.
[Figure 6] The y-axis label 'Average Inverse Pearson Correlation' is undefined; if it is the reciprocal of r, it is not defined for correlations near zero. Plot the raw correlation (with error bars) or define the transformation explicitly.
[Section VI] The term 'winner-takes-most' is vague; the reward rule gives m, m−1, ..., 1 points to all groups, so describe the rule precisely (e.g., rank-based decreasing rewards) rather than using this term.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: Experiment 1 is a fixed-metric algorithmic comparison on real LLM fine-tuning, and Experiment 2's Shapley correlation is an internal consistency check of a rank-based approximation inside an explicitly synthetic model, not an independent validation presented as derived from its own outputs.

full rationale

The paper's two experiments are self-contained but disconnected. Experiment 1 (Section III) is an empirical, fixed-metric comparison: models are fine-tuned on the same GoEmotions neighbor pool, evaluated by the same distance-to-centroid metric used for selection, and single-model versus multi-model outcomes are compared. Selecting the model with minimal distance and then reporting that same distance is the algorithm being tested, not a hidden fit or a quantity defined circularly in terms of the reported improvement. The comparison could have failed; the reported 20.04%, 5.24%, and 55.37% reductions are contingent empirical results for specific emotions. Experiment 2 (Section IV-B) is explicitly synthetic: fine-tuning is defined as M_candidate,g(t) = M_t + delta(C_g(t) - M_t), points are awarded by ranking E_method(M_candidate,g(t), E), and Shapley values are approximated with KernelSHAP on the same distance-to-expert function. The resulting Pearson correlation therefore measures how well a simple rank-based rule recovers an additive attribution of the same linear distance function. This is an internal consistency check, not an independent empirical validation of fairness in real SFT, and it is not circular: the point rule is not defined in terms of Shapley values, the correlation is not 1 by construction, and the paper reports method-dependent trade-offs (Table I, Figures 6-8). The paper itself concedes the scope limitation in Section VI: 'A critical next step is validating the framework in real-world settings with human participants.' There are no load-bearing self-citations, no fitted parameters renamed as predictions, and no uniqueness theorem imported from the authors. The main risks are external validity and the under-specified Shapley characteristic function, neither of which constitutes circular derivation.

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

The claims rest on a small set of hand-set simulation parameters and strong modeling assumptions: embedding distance as alignment, linear centroid updates as fine-tuning, and KernelSHAP as ground truth. These are stated or implicit in Sections III and IV, and they limit the results to a proof-of-concept simulation rather than a demonstrated deployment.

free parameters (4)

delta (update step size)
Controls how far each candidate model moves toward the group centroid in Experiment 2; its value is not reported, and all convergence results depend on it.
k (nearest-neighbor count) = 100
Chosen by hand for the target centroid and fine-tuning transcript in Experiment 1; no sensitivity analysis is reported for this choice.
selection noise rate = 0.05
Arbitrarily set expert error rate in Experiment 2; the paper lists varying this as future work.
number of clones per iteration = 3
Used in Experiment 1's multi-model selection; no sweep of clone count is reported for that section.

assumptions (4)

domain assumption Euclidean distance between output embeddings and a target centroid measures model alignment quality.
Section III defines convergence entirely through this distance, which presumes the embedding geometry is meaningful for alignment.
ad hoc to paper Fine-tuning can be represented as a linear step toward the weighted centroid of a group's preference vectors.
Section IV-B's update rule Mcandidate,g(t) = Mt + δ(Cg(t) - Mt) is a simulation device with no derivation from actual SFT.
domain assumption KernelSHAP yields valid estimates of individual user contributions in the simulated game.
Section IV-B step 6 uses KernelSHAP as ground truth but does not define the cooperative game or value function.
domain assumption Users are honest and consistent, and the expert target is a fixed preference vector.
Section IV-B assumes honest participation and a fixed expert target; Section VI leaves adversarial and inconsistent behavior to future work.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Crowd-SFT: Crowdsourcing for LLM Alignment." pith.science (2026). https://pith.science/paper/L7T35TI3

@misc{pith2026250604063,
  author       = {Pith},
  title        = {Pith review of: Crowd-SFT: Crowdsourcing for LLM Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L7T35TI3}},
  note         = {Machine review of arXiv:2506.04063}
}

read the original abstract

Large Language Models (LLMs) increasingly rely on Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) to align model responses with human preferences. While RLHF employs a reinforcement learning approach with a separate reward model, SFT uses human-curated datasets for supervised learning. Both approaches traditionally depend on small, vetted groups of annotators, making them costly, prone to bias, and limited in scalability. We propose an open, crowd-sourced fine-tuning framework that addresses these limitations by enabling broader feedback collection for SFT without extensive annotator training. Our framework promotes incentive fairness via a point-based reward system correlated with Shapley values and guides model convergence through iterative model updates. Our multi-model selection framework demonstrates up to a 55% reduction in target distance over single-model selection, enabling subsequent experiments that validate our point-based reward mechanism's close alignment with Shapley values (a well-established method for attributing individual contributions) thereby supporting fair and scalable participation.

Figures

Figures reproduced from arXiv: 2506.04063 by the authors.

**Figure 1.** Baseline single-model fine-tuning experiment. This experiment demonstrates traditional fine-tuning and visualizes model convergence towards specific emotional vector spaces across varying sample sizes. Using the GoEmotions dataset of human-annotated Reddit comments, we identified the 100 nearest neighbors for each emotion’s highest-ranking comment and finetuned independent models on 33, 66, and 100 samples per emot… view at source ↗

**Figure 3.** This figure compares the performance of single-step fine-tuning at [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 6.** Evaluation of Method Combinations in Collaborative Model Training: This figure shows simulation results for the framework outlined in Section IV-B. Using the MovieLens dataset of user preferences toward movie genres to represent humans in the loop, we randomly sample users by their preference vectors and select a random expert target. The experiment tests combinations of grouping and evaluation methods, rewarding us… view at source ↗

Figures from the paper (2 more)

**Figure 7.** Figure 7: shows Pearson correlation declining as user numbers increase, regardless of grouping or evaluation method. While our naive reward mechanism struggles to approximate Shapley values as user populations grow, meaningful correlation persists, especially in smaller settin…

**Figure 8.** Figure 8: Scaling Effects on Model Convergence: This figure examines the same scaling factors as [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 5 canonical work pages

[1]

Aligning large language models with human: A survey,

Y . Wang, W. Zhong, L. Li, F. Mi, X. Zeng, W. Huang, L. Shang, X. Jiang, and Q. Liu, “Aligning large language models with human: A survey,” 2023. [Online]. Available: https://arxiv.org/abs/2307.12966

arXiv 2023
[2]

Open problems and fundamental limitations of reinforcement learning from human feedback,

S. Casper, X. Davies, C. Shi, T. K. Gilbert, J. Scheurer, J. Rando, R. Freedman, T. Korbak, D. Lindner, P. Freire, T. Wang, S. Marks, C.-R. Segerie, M. Carroll, A. Peng, P. Christoffersen, M. Damani, S. Slocum, U. Anwar, A. Siththaranjan, M. Nadeau, E. J. Michaud, J. Pfau, D. Krasheninnikov, X. Chen, L. Langosco, P. Hase, E. Bıyık, A. Dragan, D. Krueger, ...

arXiv 2023
[3]

Training a Helpful and Harmless Assis- tant with Reinforcement Learning from Human Feedback,

Y . Bai, A. Jones, K. Ndousse, A. Askell, A. Chen, N. DasSarma, D. Drain, S. Fort, D. Ganguli, T. Henighan, N. Joseph, S. Kadavath, J. Kernion, T. Conerly, S. El-Showk, N. Elhage, Z. Hatfield-Dodds, D. Hernandez, T. Hume, S. Johnston, S. Kravec, L. Lovitt, N. Nanda, C. Olsson, D. Amodei, T. Brown, J. Clark, S. McCandlish, C. Olah, B. Mann, and J. Kaplan, ...

arXiv 2022
[4]

Training language models to follow instructions with human feedback,

L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, J. Schulman, J. Hilton, F. Kelton, L. Miller, M. Simens, A. Askell, P. Welinder, P. Christiano, J. Leike, and R. Lowe, “Training language models to follow instructions with human feedback,” 2022. [Online]. Available: https://arxiv.org/abs/2203.02155

arXiv 2022
[5]

Chatbot arena: An open platform for evaluating llms by human preference,

W.-L. Chiang, L. Zheng, Y . Sheng, A. N. Angelopoulos, T. Li, D. Li, H. Zhang, B. Zhu, M. Jordan, J. E. Gonzalez, and I. Stoica, “Chatbot arena: An open platform for evaluating llms by human preference,”

work page
[6]

Varco arena: A tournament approach to reference-free benchmarking large language models,

S. Son, J.-M. Oh, H. Jin, C. Jang, J. Jeong, and K. Kim, “Varco arena: A tournament approach to reference-free benchmarking large language models,” 2024. [Online]. Available: https://arxiv.org/abs/2411.01281

arXiv 2024
[7]

A Marketplace for Data: An Algorithmic Solution,

A. Agarwal, M. Dahleh, and T. Sarkar, “A Marketplace for Data: An Algorithmic Solution,” in Proceedings of the 2019 ACM Conference on Economics and Computation , ser. EC ’19. New York, NY , USA: Association for Computing Machinery, Jun. 2019, pp. 701–726

work page 2019
[8]

Transparent Contribution Evaluation for Secure Federated Learning on Blockchain

S. Ma, Y . Cao, and L. Xiong, “Transparent Contribution Evaluation for Secure Federated Learning on Blockchain.” IEEE Computer Society, Apr. 2021, pp. 88–91

work page 2021

Show all 12 references

[9]

Goemotions: A dataset of fine-grained emotions,

D. Demszky, D. Movshovitz-Attias, J. Ko, A. Cowen, G. Nemade, and S. Ravi, “Goemotions: A dataset of fine-grained emotions,” 2020. [Online]. Available: https://arxiv.org/abs/2005.00547

2020 arXiv
[10]

A unified approach to interpreting model predictions,

S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in Proceedings of the 31st International Conference on Neural Information Processing Systems , ser. NIPS’17. Red Hook, NY , USA: Curran Associates Inc., 2017, pp. 4768–4777

2017
[11]

Improving kernelshap: Practical shapley value estimation using linear regression,

I. Covert and S.-I. Lee, “Improving kernelshap: Practical shapley value estimation using linear regression,” in Proceedings of The 24th International Conference on Artificial Intelligence and Statistics , ser. Proceedings of Machine Learning Research, A. Banerjee and K. Fukumi...

2021
[2024]

Available: https://arxiv.org/abs/2403.04132

[Online]. Available: https://arxiv.org/abs/2403.04132

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Aligning large language models with human: A survey,

Y . Wang, W. Zhong, L. Li, F. Mi, X. Zeng, W. Huang, L. Shang, X. Jiang, and Q. Liu, “Aligning large language models with human: A survey,” 2023. [Online]. Available: https://arxiv.org/abs/2307.12966

arXiv 2023

[2] [2]

Open problems and fundamental limitations of reinforcement learning from human feedback,

S. Casper, X. Davies, C. Shi, T. K. Gilbert, J. Scheurer, J. Rando, R. Freedman, T. Korbak, D. Lindner, P. Freire, T. Wang, S. Marks, C.-R. Segerie, M. Carroll, A. Peng, P. Christoffersen, M. Damani, S. Slocum, U. Anwar, A. Siththaranjan, M. Nadeau, E. J. Michaud, J. Pfau, D. Krasheninnikov, X. Chen, L. Langosco, P. Hase, E. Bıyık, A. Dragan, D. Krueger, ...

arXiv 2023

[3] [3]

Training a Helpful and Harmless Assis- tant with Reinforcement Learning from Human Feedback,

Y . Bai, A. Jones, K. Ndousse, A. Askell, A. Chen, N. DasSarma, D. Drain, S. Fort, D. Ganguli, T. Henighan, N. Joseph, S. Kadavath, J. Kernion, T. Conerly, S. El-Showk, N. Elhage, Z. Hatfield-Dodds, D. Hernandez, T. Hume, S. Johnston, S. Kravec, L. Lovitt, N. Nanda, C. Olsson, D. Amodei, T. Brown, J. Clark, S. McCandlish, C. Olah, B. Mann, and J. Kaplan, ...

arXiv 2022

[4] [4]

Training language models to follow instructions with human feedback,

L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, J. Schulman, J. Hilton, F. Kelton, L. Miller, M. Simens, A. Askell, P. Welinder, P. Christiano, J. Leike, and R. Lowe, “Training language models to follow instructions with human feedback,” 2022. [Online]. Available: https://arxiv.org/abs/2203.02155

arXiv 2022

[5] [5]

Chatbot arena: An open platform for evaluating llms by human preference,

W.-L. Chiang, L. Zheng, Y . Sheng, A. N. Angelopoulos, T. Li, D. Li, H. Zhang, B. Zhu, M. Jordan, J. E. Gonzalez, and I. Stoica, “Chatbot arena: An open platform for evaluating llms by human preference,”

work page

[6] [6]

Varco arena: A tournament approach to reference-free benchmarking large language models,

S. Son, J.-M. Oh, H. Jin, C. Jang, J. Jeong, and K. Kim, “Varco arena: A tournament approach to reference-free benchmarking large language models,” 2024. [Online]. Available: https://arxiv.org/abs/2411.01281

arXiv 2024

[7] [7]

A Marketplace for Data: An Algorithmic Solution,

A. Agarwal, M. Dahleh, and T. Sarkar, “A Marketplace for Data: An Algorithmic Solution,” in Proceedings of the 2019 ACM Conference on Economics and Computation , ser. EC ’19. New York, NY , USA: Association for Computing Machinery, Jun. 2019, pp. 701–726

work page 2019

[8] [8]

Transparent Contribution Evaluation for Secure Federated Learning on Blockchain

S. Ma, Y . Cao, and L. Xiong, “Transparent Contribution Evaluation for Secure Federated Learning on Blockchain.” IEEE Computer Society, Apr. 2021, pp. 88–91

work page 2021