Pith. sign in

REVIEW 4 major objections 4 minor 66 references

Standard learning in competitive ML markets converges to overspecialization traps; peer-model probing provably restores global competence.

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 ↗

T0 review · deepseek-v4-flash

2026-08-02 20:16 UTC pith:ZBQ6TL3W

load-bearing objection Overspecialization and probing dynamics are a real contribution, but Theorem 2's boundary case and the theory-experiment gap for preference-aware probing need attention. the 4 major comments →

arxiv 2602.23565 v2 pith:ZBQ6TL3W submitted 2026-02-27 cs.LG cs.MA

Dynamics of Learning under User Choice: Overspecialization and Peer-Model Probing

Adhyyan Narang , Sarah Dean , Lillian J Ratliff , Maryam Fazel This is my paper
classification cs.LG cs.MA MSC 68T0562L2091A80
keywords overspecializationpeer-model probinguser choicemulti-learner dynamicsknowledge distillationfull-population riskstochastic approximationecho chambers
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.

In markets where several platforms learn from the same user population and users pick the platform that serves them best, this paper shows that standard streaming gradient descent almost surely converges to a stationary point where at least one learner serves its niche perfectly while performing arbitrarily badly on the full population—even when a good global model exists. The mechanism is the overspecialization trap: optimizing for the users who already choose you makes you less attractive to everyone else, which shrinks your data further. The paper then introduces MSGD with Probing (MSGD-P), in which learners mix their organic updates with gradient steps on pseudo-labeled examples obtained by querying peer models. It proves that MSGD-P converges almost surely to stationary points of a modified potential, and that if the pseudo-labels are accurate—satisfying a mean-squared-error bound B—every stationary point has full-population risk at most O((p+1)/p ε + B plus a generalization term), where ε is the Bayes risk. Four concrete scenarios satisfy the accuracy condition, including the notable case where a learner only knows users' intrinsic preferences and needs no assumption on peer quality; experiments on MovieLens, Census, and Amazon data confirm that a few dozen probing queries close most of the gap.

Core claim

The central claim is twofold. First, for any τ ≥ 1/2 (preferences matter at least as much as quality), there exists an instance G of the user-choice market such that MSGD converges almost surely to a unique stationary point Θ̄ where some learner's full-population risk R(θ̄_i) ≥ Γ, even though some model θ* has R(θ*) ≤ ε, with ε and Γ arbitrary (Theorem 2). In the construction, two subpopulations have linear labels with slopes C and −1; the equilibrium forces each learner to fit its own niche exactly, producing global risk (1−α)(C+1)^2 for one learner while the compromise model has risk α(1−α)(C+1)^2 ≤ ε. Second, the paper's probing algorithm MSGD-P converges almost surely to stationary point

What carries the argument

The machinery is a Lyapunov-potential argument for stochastic approximation. For MSGD, the potential is f(Θ) = Σ_i E_{O_i(Θ)}[ℓ(z;θ_i)], where O_i is the mixture of the preference-conditional distribution and the loss-minimizer-conditional distribution, weighted by τ and 1−τ; the iterates track the ODE Θ̇ = −∇f, and f is a strict Lyapunov function, so convergence to stationarity is almost sure (Theorem 1). MSGD-P augments f with probing terms p(L̂_i(θ_i)+λ/2∥θ_i∥²) built from a fixed dataset of pseudo-labels produced by median aggregation over a chosen set of peers; the same stochastic-approximation structure gives convergence to stationary points of f̃ (Theorem 3). The risk bound is carried

Load-bearing premise

The load-bearing premise is that the parameter iterates remain almost surely bounded (Assumption 4), which the paper assumes but does not prove; if iterates can escape to infinity, the stochastic-approximation convergence claim collapses.

What would settle it

Run MSGD on the paper's two-subpopulation construction with τ = 1 and a constant step size, and record the full-population risk of learner 1 over many seeds; Theorem 2 predicts almost-sure convergence to the specialist (C,−1) with risk (1−α)(C+1)^2, so a trajectory that settles near a mixed equilibrium or keeps a risk below Γ would falsify the theorem. As a check on the boundedness assumption, also monitor whether ∥Θ_t∥ diverges; unboundedness would invalidate Assumption 4 and the convergence result.

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

If this is right

  • If left uncorrected, any platform optimizing only over users who select it will lock into a niche, and the worse it serves outsiders the less it learns about them—so the trap is self-reinforcing.
  • With probing, even a small fixed dataset of pseudo-labels (n ≈ 50–100 in the experiments) shrinks the gap to within a percentage point of the full-data baseline, making distillation a practical remedy.
  • The risk bound scales as (p+1)/p ε + B, so increasing the probing weight p asymptotically removes the multiplicative blow-up of the Bayes risk, while pseudo-label noise B and the 1/√n generalization term remain as ceilings.
  • The preference-aware scenario shows that knowledge of user preferences alone suffices for the guarantee, independent of peer quality—a surprising result that broadens the applicability of probing.
  • The convergence and bounds are for convex losses and linear predictors; extending to deep networks and online probing is left open, but the convex case provides a clean benchmark.

Where Pith is reading between the lines

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

  • An immediate unproven extension is that if all learners probe symmetrically, the market's set of equilibria may collapse to globally competent models; the paper's simultaneous-probing experiment suggests this, but the theory covers only a single probing learner.
  • The offline-probing design is what makes the bounds tractable; switching to online probing as peers adapt could break the accurate-probing bound and produce model-collapse instabilities, exactly as the paper's discussion of future work warns.
  • The preference-aware result could be turned into a practical algorithm: a platform that can estimate which users prefer it (e.g., from opt-in/opt-out data) can route probe queries to the preference-matched peer and achieve global competence without any assumption on peer quality.
  • A testable consequence for real markets is that the gap between local and full-population performance should shrink when platforms adopt distillation from a strong incumbent; if an incumbent is absent, the majority-good median rule still helps as long as more than half of peers are near-optimal.

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 studies multi-learner streaming gradient descent in a market where each arriving user selects a platform according to a mixture of an exogenous inherent preference π(z) and predictive loss (Definition 1). It proves that the standard MSGD dynamics converge almost surely to stationary points of the aggregate observed-loss potential f (Theorem 1), and constructs a two-subpopulation linear regression instance in which, for τ ≥ 1/2, MSGD converges to a unique stationary point where one learner has arbitrarily large full-population risk while a low-risk global model exists (Theorem 2). The paper then proposes MSGD-P (Algorithm 2), which mixes organic updates with gradient steps on an offline pseudo-labeled probing dataset, proves convergence to stationary points of an augmented potential (Theorem 3), and derives finite-sample bounds on the full-population risk of stationary points under an Accurate Probing condition (Theorem 4, Corollary 1). Three sufficient scenarios for accurate probing are given: majority-good peers, a known market leader, and preference-aware probing (Definitions 2–3). The theoretical results are supplemented by semi-synthetic experiments on MovieLens, Census, and Amazon Sentiment data.

Significance. If the results are correct, the paper makes a valuable contribution to performative prediction and multi-learner learning under user choice. It goes beyond prior work by analyzing the streaming dynamics of overspecialization rather than only its existence, and by proposing a concrete, analyzable mitigation mechanism with an interpretable risk decomposition: the bound in Theorem 4 separates Bayes error, probing bias, regularization bias, and finite-sample generalization. The stochastic-approximation treatment of the potential function is technically sound, and the identification of the preference-aware scenario—where probing succeeds without any globally good peer—is interesting and nontrivial. The experiments are well-motivated and the GitHub code availability supports reproducibility. However, the paper's headline negative result (Theorem 2) has a correctness gap at the boundary τ = 1/2, and the empirical validation of the preference-aware scenario does not satisfy the assumptions of the corresponding positive guarantee. These issues need to be fixed or clearly qualified before the central claims can be fully accepted.

major comments (4)
  1. [Theorem 2, Lemma 10 (Appendix C)] The statement 'for any τ ≥ 1/2' is not supported at τ = 1/2. In Lemma 10, Case 3 is ruled out using d1 > d2, with d1 − d2 proportional to (1−α)(2τ−1); at τ = 1/2 this inequality becomes an equality and the contradiction argument fails. Solving the Case 3 stationary equations at τ = 1/2 gives θ1 = θ2 = αC − (1−α), which is a distinct stationary configuration from the claimed unique point (C,−1). Moreover, at θ1 = θ2 every user is indifferent under loss-based selection, so the argmin set in Definition 1 is not a singleton and the user-choice process M(z;Θ) is not well-defined without a tie-breaking rule; no tie-breaking rule is specified anywhere in the manuscript. Consequently, Theorem 2's conclusion of almost-sure convergence to a unique bad stationary point does not follow for τ = 1/2. The theorem should either be restricted to τ > 1/2, or the case τ = 1/2 should be analyzed separately
  2. [Assumption 4, Theorems 1 and 3] Assumption 4 (almost sure boundedness of the iterates) is a load-bearing condition for both convergence theorems, but the paper provides no verification of it for the algorithms as stated. For unregularized MSGD (Algorithm 1), gradient updates can grow without bound in general; in the bad-instance construction used in Theorem 2, the proof applies Theorem 1 to conclude convergence without checking that sup_t ||Θ_t|| < ∞ almost surely. Since the theorem's conclusion is a specific limit point, boundedness is not a trivial technicality. If Assumption 4 is intended as a standing assumption, this should be stated more prominently and the content of the claim adjusted; if the intention is that the constructed instance satisfies it, a proof should be supplied. A projected or regularized variant of MSGD would make the assumption verifiable.
  3. [Definition 3, Lemma 3(iv), Algorithm 2, Section 6] The preference-aware scenario is defined under the assumption that all learners initialize at the specialized ERM solutions Θ̄ = (θ̄_1,...,θ̄_m). However, Algorithm 2 collects pseudo-labels from the actual initial parameters Θ0 (line 4), and the experiments in Section 6 initialize models randomly. Lemma 3(iv), which gives B = ε for preference-aware probing, relies on each peer θ̄_j being the ERM solution on its preference subpopulation. With a random Θ0 this inequality can fail badly, so Assumption 5 is not satisfied and Theorem 4 does not cover the reported preference-aware experiments. The paper should either change the experimental protocol to initialize at the specialized solutions, or derive a preference-aware accuracy bound that holds for general initial models, or explicitly describe the experiments as heuristics that go beyond the theorem's assumptions.
  4. [Definition 1 and throughout] The user selection rule M(z;Θ) is undefined when multiple learners tie for the minimal loss, since the argmin is then a set and no tie-breaking rule is specified. In the bad instance at τ = 1/2 this is exactly the situation at the stationary point θ1 = θ2, but the issue is general: ties occur on measure-zero sets under Assumption 3, yet at θ1 = θ2 in the constructed example the tie set has positive measure. The algorithm and the convergence analysis need an explicit tie-breaking convention, and the statement of Theorem 2 must respect it.
minor comments (4)
  1. [Theorem 3] The statement of Theorem 3 lists Assumption 5 (accurate probing) as a hypothesis, but the proof only needs the probing datasets to be fixed finite sets; the accuracy parameter B does not appear in the convergence argument. The convergence result is independent of pseudo-label quality. The assumption should be moved to the performance guarantee (Theorem 4) rather than stated as a condition for convergence.
  2. [Theorem 4] The expression '(p+1)/p ε' should be typeset as (p+1)ε/p; the current notation 'p+1 / p' is ambiguous and is used in the theorem statement and Corollary 1.
  3. [Figure 2 caption] The caption says 'random initialization (Preference-aware scenario)', but Definition 3 assumes all learners initialize at the specialized ERM solutions. This inconsistency should be clarified in the caption and in the experimental setup description.
  4. [Algorithm 2 vs Assumption 5] Algorithm 2 constructs pseudo-labels yagg,i(·, Θ0), while Assumption 5 is stated as yagg,i(x, Θ_{-i}) without specifying the parameter argument. To avoid ambiguity, state explicitly that Assumption 5 is evaluated at the initial snapshot Θ0 used by the algorithm.

Circularity Check

0 steps flagged

No significant circularity: the derivations are self-contained conditional guarantees; self-citations provide technical lemmas but do not reduce the claims to their inputs.

full rationale

The paper's central claims do not reduce to their inputs. Theorem 2 constructs an explicit instance (Example 1) and computes the risk bounds directly from that construction (Lemma 9) and from a case analysis of stationary points (Lemma 10); the bad risk bound R(θ̄1) ≥ Γ is a computed property of the constructed distribution, not a fitted parameter. Theorem 3's convergence proof uses a stochastic approximation argument with a Lyapunov function; the only reliance on prior work of overlapping authorship is Lemma 13's gradient identity, citing Lemma 4.3 of Su and Dean [47], which is a technical mathematical lemma rather than the target result, and the boundedness Assumption 4 is explicitly stated as an assumption, not derived from the conclusion. Theorem 4's performance bound is stated in terms of the assumed pseudo-label accuracy B (Assumption 5); Lemma 3 proves the boundedness of B in each scenario from explicit conditions (majority-good, market-leader, preference-aware). In the preference-aware case, B ≤ ε follows from the optimality of the initialized specialist ERM solutions θ̄_i, which is an initialization condition stated in Definition 3, not an assertion of the conclusion that probing achieves low global risk. The bound then combines this with the optimization guarantee, which is a genuine conditional derivation. The experiments are empirical validation and are not used as evidence for the theorems. The paper's own discussion acknowledges limitations (convex losses, offline probing, the unproved Assumption 4, and the mismatch between the theory's initialization and the experiments' random initialization), but these are correctness or scope gaps, not circularity. The τ=1/2 boundary issue in Lemma 10, where Case 3 is not ruled out because the inequality is an equality, is a mathematical correctness concern in the uniqueness claim, but it is not a circular dependency: the proof does not assume the uniqueness it is trying to establish. Overall, no step was found where a prediction is equivalent to its input by definition or by fitted parameter renaming.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 0 invented entities

The central results rely on standard stochastic approximation assumptions and the accurate-probing condition. No new physical or mathematical entities are introduced; the adversarial instance in Theorem 2 uses tuned constants C and α to construct the bad outcome.

free parameters (2)
  • C (slope in bad instance) = √(Γ+ε) - 1
    Chosen in Theorem 2's construction to make learner 1's risk exceed Γ while the global optimum has risk ≤ε.
  • α (mixture weight in bad instance) = ε/(C+1)²
    Chosen so that R(θ*) ≤ ε; part of the adversarial instance.
axioms (6)
  • domain assumption Assumption 1: continuous density with bounded support and bounded labels.
    Ensures the loss is β-smooth, convex, and locally Lipschitz; used throughout the analysis.
  • standard math Assumption 2: learning rates satisfy Ση_t=∞ and Ση_t²<∞.
    Standard stochastic approximation condition for convergence.
  • domain assumption Assumption 3: measure of near-tie sets tends to zero.
    Technical regularity used to prove local Lipschitzness of ai(Θ) in Lemma 14.
  • domain assumption Assumption 4: iterates Θ_t are almost surely bounded and the stationary set is compact.
    Load-bearing for the stochastic approximation convergence theorems (Theorems 1 and 3); not proven from problem data.
  • domain assumption Assumption 5: accurate probing condition, E[(y_agg - y)²] ≤ B.
    Central condition for the performance bound (Theorem 4); the paper provides sufficient scenarios in Lemma 3.
  • standard math Lemma 4.3 of [47] (envelope theorem for gradient of ai(Θ)E_Di[ℓ]).
    Used in Lemma 13 to claim the distributional derivative term vanishes; cited from Su & Dean [47].

pith-pipeline@v1.3.0-alltime-deepseek · 36913 in / 13548 out tokens · 123940 ms · 2026-08-02T20:16:27.544131+00:00 · methodology

0 comments
read the original abstract

In many economically relevant contexts where machine learning is deployed, multiple platforms obtain data from the same pool of users, each of whom selects the platform that best serves them. Prior work in this setting focuses exclusively on the "local" losses of learners on the distribution of data that they observe. We find that there exist instances where learners who use existing algorithms almost surely converge to models with arbitrarily poor global performance, even when models with low full-population loss exist. This happens through a feedback-induced mechanism, which we call the overspecialization trap: as learners optimize for users who already prefer them, they become less attractive to users outside this base, which further restricts the data they observe. Inspired by the recent use of knowledge distillation in modern ML, we propose an algorithm that allows learners to "probe" the predictions of peer models, enabling them to learn about users who do not select them. Our analysis characterizes when probing succeeds: this procedure converges almost surely to a stationary point with bounded full-population risk when probing sources are sufficiently informative, e.g., a known market leader or a majority of peers with good global performance. We verify our findings with semi-synthetic experiments on the MovieLens, Census, and Amazon Sentiment datasets.

Figures

Figures reproduced from arXiv: 2602.23565 by Adhyyan Narang, Lillian J Ratliff, Maryam Fazel, Sarah Dean.

Figure 1
Figure 1. Figure 1: Illustration of our online multi-learner problem setting. The borders of users represent [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: At each step t: 1. A user z t ∼ P arrives and selects platform M(z t ; Θt ). 2. The selected learner observes z t , incurs loss ℓ(z t ; θ t M(z t ;Θt) ), and updates its parameters. We denote an instance of this market as G = (P, ℓ, π, τ ). Learning Objective. Each learner’s goal is to minimize the full-population risk R(θ) = Ez∼P[ℓ(z; θ)]. We write θ ⋆ = arg minθ R(θ) for the population-optimal model and … view at source ↗
Figure 2
Figure 2. Figure 2: MSGD full-population performance with random initialization (Preference￾aware scenario). Left: Census test accuracy. Mid: Amazon sentiment test accuracy Right: MovieLens test loss. The dashed black line represents the performance of a baseline θ ∗ trained on the full dataset. In all cases, the hyperparameters (τ = 0.3, λ = 10−3 ) are used. Hence, probing breaks the information barrier created by user-choic… view at source ↗
Figure 3
Figure 3. Figure 3: Effect of probing on full-population performance (Preference-aware scenario). Left: Census final accuracy vs probing weight p. Mid: Amazon sentiment final accuracy vs probing weight p Right: MovieLens final loss vs p. In all cases, triangle markers indicate the probing learner. where σ is the sigmoid function. The model predicts yˆ = 1[θ ⊤x > 0]. Amazon Reviews 2023 with Logistic Loss. We use the Amazon Re… view at source ↗
Figure 4
Figure 4. Figure 4: Performance of probing learner on Census as a function of n. Error bars show one standard deviation over 10 random seeds. probing-source selection ( [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: MSGD full-population performance with random initialization (Market-leader scenario). Left: Census test accuracy. Right: MovieLens test loss. Here τ = 0.3. only modestly, indicating that probing primarily benefits the underperforming learner and mitigates overspecialization. Expt 4: Impact of noise in selection of probed labels Finally, we test robustness to noisy probing-source selection. For each probe q… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of probing on full-population performance (Market-leader scenario) Left: Final accuracy vs probing weight p on Census. Right: MovieLens final loss vs p. The triangle markers indicate Learner 4 probes the market leader, learner 1. Here τ = 0.7 [PITH_FULL_IMAGE:figures/full_fig_p050_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: MSGD full-population performance with random initialization (Majority good scenario). Left: Census test accuracy. Right: MovieLens test loss. Here τ = 0.3. 50 [PITH_FULL_IMAGE:figures/full_fig_p050_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Effect of probing on full-population performance (Majority Good Scenario) Left: Final accuracy vs probing weight p on Census. Right: MovieLens final loss vs p. Triangle markers indicate Learner 4 is probing via median aggregation over all peers. Here τ = 0.7 [PITH_FULL_IMAGE:figures/full_fig_p051_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Effect of probing-noise on full-population performance (Preference-aware scenario). Left: Census final accuracy vs probing weight p. Mid: Amazon sentiment final accuracy vs probing weight p Right: MovieLens final loss vs p. In all cases, the green learner is the probing learner. 51 [PITH_FULL_IMAGE:figures/full_fig_p051_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Effect of probing on full-population performance when multiple learners probe (Preference-aware scenario). Census final accuracy vs probing weight p. Triangle markers indicate the probing learners (Learners 2 and 3). The dashed black line denotes the full-data baseline. 52 [PITH_FULL_IMAGE:figures/full_fig_p052_10.png] 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

66 extracted references · 3 canonical work pages

  1. [1]

    Eliciting auxiliary information for cold start user recommendation: A survey.Applied Sciences, 11(20):9608, 2021

    Nor Aniza Abdullah, Rasheed Abubakar Rasheed, Mohd Hairul Nizam Nasir, and Md Mujibur Rahman. Eliciting auxiliary information for cold start user recommendation: A survey.Applied Sciences, 11(20):9608, 2021

    2021

  2. [2]

    Large scale distributed neural network training through online distillation

    Rohan Anil, Gabriel Pereyra, Alexandre Passos, Róbert Ormandi, George E Dahl, and Geof- frey E Hinton. Large scale distributed neural network training through online distillation. In International Conference on Learning Representations (ICLR), 2018

    2018

  3. [3]

    Minillm: Knowledge distillation of large language models.arXiv preprint arXiv:2306.08543, 2023

    Anonymous. Minillm: Knowledge distillation of large language models.arXiv preprint arXiv:2306.08543, 2023. URLhttps://arxiv.org/abs/2306.08543

    Pith/arXiv arXiv 2023

  4. [4]

    Jye E. Beardow. Scroll, click, like, share, repeat: The algorithmic polarisation phenomenon. ANU Journal of Law & Technology, 2(1):153–164, 2021. Autumn 2021 issue

    2021

  5. [5]

    Best response regression.Advances in Neural Information Processing Systems, 30, 2017

    Omer Ben-Porat and Moshe Tennenholtz. Best response regression.Advances in Neural Information Processing Systems, 30, 2017

    2017

  6. [6]

    Regression equilibrium

    Omer Ben-Porat and Moshe Tennenholtz. Regression equilibrium. InProceedings of the 2019 ACM Conference on Economics and Computation, pages 173–191, 2019

    2019

  7. [7]

    The Netflix prize

    James Bennett and Stan Lanning. The Netflix prize. InProceedings of KDD Cup and Workshop. ACM, 2007

    2007

  8. [8]

    Springer, 2008

    Vivek S Borkar.Stochastic approximation: a dynamical systems viewpoint, volume 9. Springer, 2008

    2008

  9. [9]

    Initializing services in interactive ml systems for diverse users.arXiv preprint arXiv:2312.11846, 2023

    Avinandan Bose, Mihaela Curmei, Daniel L Jiang, Jamie Morgenstern, Sarah Dean, Lillian J Ratliff, and Maryam Fazel. Initializing services in interactive ml systems for diverse users.arXiv preprint arXiv:2312.11846, 2023

    Pith/arXiv arXiv 2023

  10. [10]

    What makes a good fisherman? linear regression under self-selection bias

    Yeshwanth Cherapanamjeri, Constantinos Daskalakis, Andrew Ilyas, and Manolis Zampetakis. What makes a good fisherman? linear regression under self-selection bias. InProceedings of the 55th Annual ACM Symposium on Theory of Computing, STOC 2023. Association for Computing Machinery, 2023. doi: 10.1145/3564246.3585177. URL https://doi.org/10.1145/3564246. 3585177

    arXiv 2023

  11. [11]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing GPT-4 with 90%* ChatGPT quality.https://lmsys.org/ blog/2023-03-30-vicuna/, March 2023. Accessed: 2025

    2023

  12. [12]

    Chatbot arena: An open platform for evaluating LLMs by human preference

    Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael Jordan, Joseph E Gonzalez, and Ion Stoica. Chatbot arena: An open platform for evaluating LLMs by human preference. InProceedings of the 41st International Conference on Machine Learning, 2024. URLhttps://arxiv.org/abs/ 2403.04132

    Pith/arXiv arXiv 2024

  13. [13]

    The effect of people recommenders on echo chambers and polarization

    Federico Cinus, Marco Minici, Corrado Monti, and Francesco Bonchi. The effect of people recommenders on echo chambers and polarization. InProceedings of the Sixteenth International AAAI Conference on Web and Social Media (ICWSM ’22), pages 90–101. Association for the Advancement of Artificial Intelligence (AAAI), 2022. ICWSM 2022. 16

    2022

  14. [14]

    Emergent specialization from participation dynamics and multi-learner retraining

    Sarah Dean, Mihaela Curmei, Lillian Ratliff, Jamie Morgenstern, and Maryam Fazel. Emergent specialization from participation dynamics and multi-learner retraining. InInternational Conference on Artificial Intelligence and Statistics, pages 343–351. PMLR, 2024

    2024

  15. [15]

    Retiring adult: New datasets for fair machine learning.Advances in neural information processing systems, 34:6478–6490, 2021

    Frances Ding, Moritz Hardt, John Miller, and Ludwig Schmidt. Retiring adult: New datasets for fair machine learning.Advances in neural information processing systems, 34:6478–6490, 2021

    2021

  16. [16]

    Competing ai: How does competition feedback affect machine learning? InInternational Conference on Artificial Intelligence and Statistics, pages 1693–1701

    Tony Ginart, Eva Zhang, Yongchan Kwon, and James Zou. Competing ai: How does competition feedback affect machine learning? InInternational Conference on Artificial Intelligence and Statistics, pages 1693–1701. PMLR, 2021

    2021

  17. [17]

    Santos, Gauthier Gidel, and Simon Lacoste- Julien

    António Góis, Mehrnaz Mofakhami, Fernando P. Santos, Gauthier Gidel, and Simon Lacoste- Julien. Performative prediction on games and mechanism design. InProceedings of The 28th International Conference on Artificial Intelligence and Statistics, volume 258 ofProceedings of Machine Learning Research, pages 1855–1863. PMLR, 2025

    2025

  18. [18]

    A logit model of brand choice calibrated on scanner data.Marketing Science, 2(3):203–238, 1983

    Peter M Guadagni and John DC Little. A logit model of brand choice calibrated on scanner data.Marketing Science, 2(3):203–238, 1983

    1983

  19. [19]

    The false promise of imitating proprietary LLMs

    Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, and Dawn Song. The false promise of imitating proprietary LLMs. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  20. [20]

    Strategic classification

    Moritz Hardt, Nimrod Megiddo, Christos Papadimitriou, and Mary Wootters. Strategic classification. InProceedings of the 2016 ACM Conference on Innovations in Theoretical Computer Science, pages 111–122. ACM, 2016

    2016

  21. [21]

    The movielens datasets: History and context.Acm transactions on interactive intelligent systems (tiis), 5(4):1–19, 2015

    F Maxwell Harper and Joseph A Konstan. The movielens datasets: History and context.Acm transactions on interactive intelligent systems (tiis), 5(4):1–19, 2015

    2015

  22. [22]

    Strategic apple tasting.Adv

    Keegan Harris, Chara Podimata, and Zhiwei Steven Wu. Strategic apple tasting.Adv. Neural Inf. Process. Syst., abs/2306.06250, June 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Fairness without demographics in repeated loss minimization

    Tatsunori Hashimoto, Megha Srivastava, Hongseok Namkoong, and Percy Liang. Fairness without demographics in repeated loss minimization. InInternational Conference on Machine Learning, pages 1929–1938. PMLR, 2018

    1929

  24. [24]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. URLhttps://arxiv.org/abs/1503.02531

    Pith/arXiv arXiv 2015

  25. [25]

    A cross-platform consumer behavior analysis of large- scale mobile shopping data

    Hong Huang, Bo Zhao, Hao Zhao, Zhou Zhuang, Zhenxuan Wang, Xiaoming Yao, Xinggang Wang, Hai Jin, and Xiaoming Fu. A cross-platform consumer behavior analysis of large- scale mobile shopping data. InProceedings of the 2018 World Wide Web Conference, WWW ’18, pages 1785–1794, Lyon, France, April 2018. International World Wide Web Conferences Steering Commit...

    arXiv 2018

  26. [26]

    Marzo, Isela Mendoza, and Celso C

    Ruben Interian, Ruslán G. Marzo, Isela Mendoza, and Celso C. Ribeiro. Network polarization, filter bubbles, and echo chambers: An annotated review of measures and reduction methods. arXiv preprint, 2022. arXiv:2207.13799. 17

    Pith/arXiv arXiv 2022

  27. [27]

    Jaenicke, and Ardalan Mirshani

    Roozbeh Irani-Kermani, Edward C. Jaenicke, and Ardalan Mirshani. Accommodating het- erogeneity in brand loyalty estimation: Application to the U.S. beer retail market.Jour- nal of Marketing Analytics, 11(4):820–835, 2023. doi: 10.1057/s41270-022-00187-2. URL https://doi.org/10.1057/s41270-022-00187-2

    work page doi:10.1057/s41270-022-00187-2 2023

  28. [28]

    Competition, alignment, and equilibria in digital marketplaces.Proceedings of the AAAI Conference on Artificial Intelligence, 37(5):5689–5696, 2023

    Meena Jagadeesan, Michael I Jordan, and Nika Haghtalab. Competition, alignment, and equilibria in digital marketplaces.Proceedings of the AAAI Conference on Artificial Intelligence, 37(5):5689–5696, 2023. doi: 10.1609/aaai.v37i5.25706

    work page doi:10.1609/aaai.v37i5.25706 2023

  29. [29]

    Improved bayes risk can yield reduced social welfare under competition.arXiv preprint arXiv:2306.14670, 2023

    Meena Jagadeesan, Michael I Jordan, Jacob Steinhardt, and Nika Haghtalab. Improved bayes risk can yield reduced social welfare under competition.arXiv preprint arXiv:2306.14670, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    Participatory personalization in classification

    Hailey James, Chirag Nagpal, Katherine A Heller, and Berk Ustun. Participatory personalization in classification. InThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  31. [31]

    Social media polarization and echo chambers in the context of covid-19: Case study.JMIRx med, 2(3):e29570, 2021

    Julie Jiang, Xiang Ren, and Emilio Ferrara. Social media polarization and echo chambers in the context of covid-19: Case study.JMIRx med, 2(3):e29570, 2021. doi: 10.2196/29570

    work page doi:10.2196/29570 2021

  32. [32]

    Competition over data: how does data purchase affect users?arXiv preprint arXiv:2201.10774, 2022

    Yongchan Kwon, Antonio Ginart, and James Zou. Competition over data: how does data purchase affect users?arXiv preprint arXiv:2201.10774, 2022

    Pith/arXiv arXiv 2022

  33. [33]

    Multi-agent performative prediction with greedy deployment and consensus seeking agents

    Qiang Li, Chung-Yiu Yau, and Hoi-To Wai. Multi-agent performative prediction with greedy deployment and consensus seeking agents. InAdvances in Neural Information Processing Systems, volume 35, pages 38449–38460, 2022

    2022

  34. [34]

    Direct preference knowledge distillation for large language models.arXiv preprint arXiv:2406.19774, 2024

    Yixing Li, Yuxian Gu, Li Dong, Dequan Wang, Yu Cheng, and Furu Wei. Direct preference knowledge distillation for large language models.arXiv preprint arXiv:2406.19774, 2024. URL https://arxiv.org/abs/2406.19774

    Pith/arXiv arXiv 2024

  35. [35]

    Holistic evaluation of language models.Transactions on Machine Learning Research, 2023

    Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, Benjamin Newman, Binhang Yuan, Bobby Yan, Ce Zhang, Christian Cosgrove, Christopher D Manning, Christopher Ré, Diana Acosta-Navas, Drew A Hudson, Eric Zelikman, Esin Durmus, Faisal Ladhak, Frieda Rong, Hongyu R...

    Pith/arXiv arXiv 2023

  36. [36]

    Ratliff, and S

    Eric Mazumdar, Lillian J. Ratliff, and S. Shankar Sastry. On gradient-based learning in continuous games.SIAM Journal on Mathematics of Data Science, 2(1):103–131, 2020. doi: 10.1137/18M1231298. URLhttps://doi.org/10.1137/18M1231298

    doi:10.1137/18m1231298 2020

  37. [37]

    Outside the echo chamber: Optimizing the performative risk

    John Miller, Juan C Perdomo, and Tijana Zrnic. Outside the echo chamber: Optimizing the performative risk. InProceedings of the 38th International Conference on Machine Learning, volume 139 ofProceedings of Machine Learning Research, pages 7710–7720. PMLR, 2021

    2021

  38. [38]

    Multiplayer performative prediction: Learning in decision-dependent games.Journal of Machine Learning Research, 24(202):1–56, 2023

    Adhyyan Narang, Evan Faulkner, Dmitriy Drusvyatskiy, Maryam Fazel, and Lillian J Ratliff. Multiplayer performative prediction: Learning in decision-dependent games.Journal of Machine Learning Research, 24(202):1–56, 2023. 18

    2023

  39. [39]

    PMLB: A large benchmark suite for machine learning evaluation and comparison

    Randal S Olson, William La Cava, Patryk Orzechowski, Ryan J Urbanowicz, and Jason H Moore. PMLB: A large benchmark suite for machine learning evaluation and comparison. BioData Mining, 10(1):36, 2017

    2017

  40. [40]

    Performative prediction

    Juan Perdomo, Tijana Zrnic, Celestine Mendler-Dünner, and Moritz Hardt. Performative prediction. InInternational Conference on Machine Learning, pages 7599–7609. PMLR, 2020

    2020

  41. [41]

    Multi-agent performative prediction: From global stability and optimality to chaos

    Georgios Piliouras and Fang-Yi Yu. Multi-agent performative prediction: From global stability and optimality to chaos. InProceedings of the 24th ACM Conference on Economics and Computation, pages 1047–1048. ACM, 2023

    2023

  42. [42]

    Fair participation via sequential policies

    Reilly Raab, Ross Boczar, Maryam Fazel, and Yang Liu. Fair participation via sequential policies. InAAAI Conference on Artificial Intelligence, 2024

    2024

  43. [43]

    ImageNet large scale visual recognition challenge.International Journal of Computer Vision, 115(3):211–252, 2015

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge.International Journal of Computer Vision, 115(3):211–252, 2015

    2015

  44. [44]

    H.J. Scudder. Probability of error of some adaptive pattern-recognition machines.IEEE Transactions on Information Theory, 11(3):363–371, July 1965

    1965

  45. [45]

    Strategic usage in a multi-learner setting

    Eliot Shekhtman and Sarah Dean. Strategic usage in a multi-learner setting. InInternational Conference on Artificial Intelligence and Statistics. PMLR, 2024

    2024

  46. [46]

    Ai models collapse when trained on recursively generated data.Nature, 631(8022):755–759, 2024

    Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal. Ai models collapse when trained on recursively generated data.Nature, 631(8022):755–759, 2024

    2024

  47. [47]

    Learning from streaming data when users choose.arXiv [cs.LG], June 2024

    Jinyan Su and Sarah Dean. Learning from streaming data when users choose.arXiv [cs.LG], June 2024

    2024

  48. [48]

    Gkd: A general knowledge distillation framework for large-scale pre-trained language model.arXiv preprint arXiv:2306.06629, 2023

    Shicheng Tan, Weng Lam Tam, Yuanchun Wang, Wenwen Gong, Yang Yang, Hongyin Tang, Keqing He, Jiahao Liu, Jingang Wang, Shu Zhao, Peng Zhang, and Jie Tang. Gkd: A general knowledge distillation framework for large-scale pre-trained language model.arXiv preprint arXiv:2306.06629, 2023. URLhttps://arxiv.org/abs/2306.06629

    Pith/arXiv arXiv 2023

  49. [49]

    Hashimoto

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford Alpaca: An instruction-following LLaMA model.https://github.com/tatsu-lab/stanford_alpaca, 2023. Accessed: 2025

    2023

  50. [50]

    Last-iterate convergence for symmetric, general-sum,2 × 2games under the exponential weights dynamic.arXiv preprint arXiv:2502.08063, 2025

    Guanghui Wang, Ioannis Panageas, Georgios Piliouras, and Fang-Yi Yu. Last-iterate convergence for symmetric, general-sum,2 × 2games under the exponential weights dynamic.arXiv preprint arXiv:2502.08063, 2025

    arXiv 2025

  51. [51]

    Network effects in performative prediction games

    Xiaolu Wang, Chung-Yiu Yau, and Hoi To Wai. Network effects in performative prediction games. InProceedings of the 40th International Conference on Machine Learning, volume 202 ofProceedings of Machine Learning Research, pages 36514–36540. PMLR, 2023

    2023

  52. [52]

    Did deepseek copy off of openai? and what is distillation? Forbes, 2025

    John Werner. Did deepseek copy off of openai? and what is distillation? Forbes, 2025. URL https://www.forbes.com/sites/johnwerner/2025/01/30/ did-deepseek-copy-off-of-openai-and-what-is-distillation/. 19

    2025

  53. [53]

    A survey on knowledge distillation of large language models.arXiv preprint arXiv:2402.13116, 2024

    Xiaohan Xu, Ming Li, Chongyang Tao, Tao Shen, Reynold Cheng, Jinyang Li, Can Xu, Dacheng Tao, and Tianyi Zhou. A survey on knowledge distillation of large language models.arXiv preprint arXiv:2402.13116, 2024. URLhttps://arxiv.org/abs/2402.13116

    Pith/arXiv arXiv 2024

  54. [54]

    Survey on knowledge distillation for large language models: Methods, evaluation, and application.arXiv preprint arXiv:2407.01885, 2024

    Chuanpeng Yang, Wang Lu, Yao Zhu, Yidong Wang, Qian Chen, Chenlong Gao, Bingjie Yan, and Yiqiang Chen. Survey on knowledge distillation for large language models: Methods, evaluation, and application.arXiv preprint arXiv:2407.01885, 2024. URLhttps://arxiv.org/ abs/2407.01885

    Pith/arXiv arXiv 2024

  55. [55]

    Unsupervised word sense disambiguation rivaling supervised methods

    David Yarowsky. Unsupervised word sense disambiguation rivaling supervised methods. In33rd Annual Meeting of the Association for Computational Linguistics, pages 189–196, Cambridge, Massachusetts, USA, June 1995. Association for Computational Linguistics. doi: 10.3115/ 981658.981684

    arXiv 1995

  56. [56]

    Xueru Zhang, Mohammadmahdi Khaliligarekani, Cem Tekin, et al. Group retention when using machine learning in sequential decision making: the interplay between user dynamics and fairness.Advances in Neural Information Processing Systems, 32, 2019

    2019

  57. [57]

    Deep mutual learning

    Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4320–4328, 2018

    2018

  58. [58]

    invariant set

    Zaiwei Zhu, Runyu Wan, Yingbin Cho, Haoming Luo, Zhuoran Yang, and Zhaoran Wang. Online performative gradient descent for learning nash equilibria in decision-dependent games. InAdvances in Neural Information Processing Systems, volume 36, 2023. 20 A Extended Related Work This section provides detailed comparisons with related work summarized in Section 1...

    2023

  59. [59]

    Here, {P1,P 2} are2subpopulations

    Distribution P: is defined as a mixture of subpopulations.P = αP1 + (1−α )P2. Here, {P1,P 2} are2subpopulations. For either subpopulation, the covariates are generated from the zero mean and unit-variance uniform distribution: x∼Unif([− √ 3, √ 3])for(x, y)∼ P i,(6) For each subpopulation, the response variable is generated as: y=Cxfor(x, y)∼ P 1 (7) y=−xf...

  60. [60]

    Loss function:ℓ(x, y, θ) = (y−θT x)2 is the squared loss

  61. [61]

    Lemma 9.Consider the 1-D bad-outcome family in Example 1 with mixture weightα∈(0,1)and slope parameterC >1, and letR(θ) =E (x,y)∼P [(y−θx) 2]

    Rankingπ(z) =ifor(x, y)∼ P i. Lemma 9.Consider the 1-D bad-outcome family in Example 1 with mixture weightα∈(0,1)and slope parameterC >1, and letR(θ) =E (x,y)∼P [(y−θx) 2]. (i) The least-squares predictor on the full mixture,θ⋆ =αC−(1−α), satisfies R(θ⋆) =α(1−α)(C+ 1) 2. (ii) The specialist trained onP1 is ¯θ1 =C, and its mixture risk is R(¯θ1) = (1−α)(C+...

  62. [62]

    There existsθ⋆ withR(θ ⋆)≤ϵ

  63. [63]

    Proof.The proof follows by considering Example 1

    The MSGD iterates converge to a unique stationary point¯Θ where R(¯θi) ≥ Γfor some learner i∈[m]. Proof.The proof follows by considering Example 1. Part (i)The proof follows from Lemma 9. Choosingα= ϵ (C+1)2 satisfies the condition. Characterizing the stationary pointsUnder the idealized MSGD convergence conditions used in the appendix analysis, the MSGD ...

  64. [64]

    Lemma 12(Martingale variance bound).Suppose that Assumption 1 holds and that the probing datasets {Di}i∈U are fixed finite sets

    The step sizes{a(n)}satisfy P n a(n) =∞and P n a(n)2 <∞ 30 3.{M n} is a martingale difference sequence satisfyingE[||Mn+1||2|Fn] ≤K (1 + ||xn||2)for some K >0 4.sup n ||xn||<∞almost surely Then almost surely, the sequence{xn} converges to a (possibly sample path dependent) compact connected internally chain transitive invariant set of the ODE ˙x(t) =h(x(t...

  65. [65]

    The first and third terms are each bounded byLK∥Θ−Θ ′∥by Lipschitzness

    ≤ 0. The first and third terms are each bounded byLK∥Θ−Θ ′∥by Lipschitzness. Thusℓ(x, θ1)−ℓ(x, θ2)≤2L K∥Θ−Θ ′∥. Define S= x: ℓ(x, θ1)−ℓ(x, θ2) ≤2L K∥Θ−Θ ′∥ . Since λ(S) ≤ (2LK/C)∥Θ − Θ′∥ for some constantC from Assumption 3 andX1(Θ′) \X 1(Θ) ⊂S , we get λ(X1(Θ′) \X 1(Θ)) ≤C ′∥Θ − Θ′∥. The same argument applies to the other set difference, so altogether a1...

  66. [66]

    Combining the bounds yields the result. Lemma 29(Empirical bridge on the probing batch).On the probing dataset{(˜xq i ,˜yq i , yq i )}n q=1, define bLprobe(W) = 1 n nX q=1 CE(˜yq i , qW (˜xq i )), bLtrue(W) = 1 n nX q=1 CE(yq i , qW (˜xq i )),∆ 1,n = 1 n nX q=1 ∥˜yq i −yq i ∥1. Then for anyW, bLprobe(W)− bLtrue(W) ≤R∥W∥ F ·∆ 1,n. Proof.Apply Lemma 28 term...

    2018