Pith. sign in

REVIEW 3 major objections 5 minor 296 references

The paper derives a threshold rule for when supervised fine-tuning outperforms in-context learning, and shows that offering both never cuts platform profit.

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 02:18 UTC pith:7TXSIHRX

load-bearing objection A serious theory paper on SFT vs ICL with real formal content; the core threshold and platform theorems hold, but the non-monotonicity results rest on an unvalidated approximation and the empirics are qualitative. the 3 major comments →

arxiv 2607.14371 v1 pith:7TXSIHRX submitted 2026-07-15 cs.LG econ.THstat.ML

Supervised Fine-Tuning vs. In-Context Learning: An Equilibrium Analysis of LLM Personalization under Congestion

Fengzhuo Zhang , Zhuoran Yang , Dirk Bergemann This is my paper
classification cs.LG econ.THstat.ML MSC 91A1662J0568T07
keywords LLM personalizationin-context learningsupervised fine-tuningcongestion gamesmean-field equilibriumplatform pricingpretraining coveragesignal-to-noise ratio
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper builds a tractable model of LLM personalization with two options: in-context learning (ICL), which only adjusts the prompt, and supervised fine-tuning (SFT), which updates parameters at higher compute cost. It claims that which method is statistically better is governed by a threshold in pretraining coverage and the signal-to-noise ratio of user data, with SFT dominating once coverage passes a critical value. In a shared-resource setting with congestion, it proves that an equilibrium exists and that the equilibrium congestion level is unique, but that congestion responds non-monotonically to pretraining quality, coverage, task noise, and SFT cost. Finally, it shows that offering both SFT and ICL never lowers the platform's maximal profit relative to an ICL-only menu. If correct, user algorithm choice, congestion levels, and platform menu design reduce to computable functions of pretraining coverage, task dimension, noise, and per-sample compute costs.

Core claim

Under aligned Gaussian linear models, the paper derives closed-form mean-squared errors for ICL and SFT. ICL can only reduce error within the subspace covered by pretraining and pays a fixed residual bias in the uncovered directions, whereas SFT can learn all directions but risks large variance when data are noisy. With optimally regularized SFT, the comparison reduces to a threshold: SFT achieves weakly lower error than ICL if and only if the coverage ratio R=r/(d-r) at least equals R_crit, a closed-form function of the null-space SNR κ=sτ/(2σ̃²), pretraining precision, and task variance. The paper then embeds these statistical errors in a continuum congestion game, proves the equilibrium c

What carries the argument

The central objects are the coverage ratio R=r/(d-r) and the null-space signal-to-noise ratio κ=sτ/(2σ̃²), together with the critical threshold R_crit derived from the exact error formulas. The threshold comes from comparing ICL's conservatism in uncovered directions—a fixed cost τ per uncovered dimension—with SFT's variance there, which scales as 2σ̃²/s. The equilibrium analysis uses a monotone congestion mapping: each user's best-response resource demand decreases as conjectured congestion rises, which yields a unique fixed point for aggregate demand and supports the subsequent price and menu comparative statics.

Load-bearing premise

The headline congestion and profit results are proven under simplified error curves E_ICL(N)=2σ̃²r/(N^α+2σ̃²/ζ)+(d-r)τ and E_SFT(N)=2σ̃²d/N^α, which are chosen to match the exact formulas only as N approaches 0 or infinity; the paper provides no bound on the approximation error at intermediate sample counts, so if the true error curves deviate there, the non-monotone congestion patterns and the 'adding SFT never hurts profit' result could be artifacts of the approximation.

What would settle it

Compute the exact errors from equation (4.1) at intermediate sample counts and re-solve the equilibrium congestion and platform profit for a concrete parameter grid; if R* as a function of r, σ̃, or R_SFT loses its non-monotone shape, or if the maximal profit with SFT available drops below the ICL-only profit for any price, the central claims fail. A complementary experiment would measure ICL and SFT errors on linear regression with sample sizes between, say, 10 and 200, where ICL may not yet have plateaued, and check whether the predicted ranking flips match the threshold.

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

If this is right

  • Users with low pretraining coverage or noisy personalization data should prefer ICL; users with high coverage and informative data should prefer SFT, and congestion shifts the balance toward ICL.
  • Equilibrium congestion follows an increase-decrease-increase pattern as pretraining coverage expands when compute is cheap, and becomes monotone when prices or ICL costs are high.
  • Raising the platform's unit resource price always reduces equilibrium congestion, and the profit-maximizing price is finite under stated regularity conditions.
  • Platforms that add SFT to an ICL-only menu do not reduce their maximal profit, despite higher aggregate compute load.
  • Improving pretraining precision always lowers equilibrium congestion, while broader coverage and harder tasks can sometimes increase it.

Where Pith is reading between the lines

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

  • Editorial extension: the threshold formula suggests a calibration recipe for real platforms—estimate r, d, τ, s, and σ̃ from task data to decide when to market SFT versus ICL and how to price compute, but the paper does not propose an estimation procedure.
  • Editorial extension: the 'adding SFT never hurts profit' result is proven for a single platform; in a competitive multi-platform setting, users could migrate platforms, so the profit guarantee may fail without additional assumptions.
  • Editorial extension: the non-monotonicity results are proven for simplified error curves that match the exact formulas only at extreme sample counts; re-running the equilibrium analysis with exact errors at intermediate sample sizes would test whether the qualitative conclusions survive.
  • Editorial extension: the model implies a substitution between offline pretraining compute and online serving compute—improving pretraining precision reduces inference-time congestion—which could inform infrastructure planning beyond pricing.

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

3 major / 5 minor

Summary. The paper proposes a continuum-user congestion game for LLM personalization, in which users choose between ICL and SFT and choose how many personalization samples to use, while a platform sets a per-unit resource price. The statistical component derives closed-form error formulas for ICL and SFT under spectral-alignment and isotropicity assumptions, yielding a threshold condition (Proposition 3.6) for when SFT with optimally tuned regularization outperforms ICL. The game-theoretic component proves existence of an equilibrium and uniqueness of the equilibrium congestion level (Theorem 4.1), derives comparative statics of equilibrium congestion with respect to pretraining precision, coverage, task noise, and resource costs (Propositions 4.4–4.6), and characterizes anchoring in a two-type model (Proposition 4.8). Finally, the platform-level analysis shows that equilibrium congestion decreases in the price (Theorem 5.2), that an optimal finite price exists under regularity conditions (Proposition 5.3), and that adding SFT to an ICL-only menu never reduces maximal profit (Theorem 5.5). The paper includes GPT-2 experiments on linear regression that validate qualitative predictions about ICL error plateaus, linear growth in uncovered dimensions, and the sample-size-dependent ranking of SFT and ICL, plus a survey of 21 AI platforms documenting the spread of SFT APIs.

Significance. If the main results hold, the paper makes a valuable contribution by connecting statistical primitives of LLM personalization — pretraining coverage, feature alignment, signal-to-noise ratios, and per-sample compute costs — to equilibrium congestion and platform pricing. Proposition 3.6 gives a closed-form, non-fitted threshold that separates regimes where ICL's conservatism is preferable to SFT's ability to learn outside the pretrained subspace. Theorem 4.1's uniqueness of the equilibrium congestion level is a useful mean-field-type result because it makes platform profit and congestion predictions single-valued even when equilibrium policies are not unique. The GPT-2 experiments provide independent support for the statistical model's qualitative predictions, since the constants in the theory were not fitted to these experiments. However, the paper's second main finding — non-monotone comparative statics of congestion — is proved only under an unvalidated simplified error model, which currently limits the support for those conclusions. The platform menu theorem is conditional on a strong uniformity assumption that is not reflected in the abstract's unqualified statement.

major comments (3)
  1. [Section 4.2, Eq. (4.2)] The simplified error functions E_ICL(N)=2σ̃²r/(N^α+2σ̃²/ζ)+(d−r)τ and E_SFT(N)=2σ̃²d/N^α are asserted to 'preserve the main insights' of the exact formulas (4.1), but no error bound or numerical audit at intermediate N is provided. Propositions 4.4–4.6 and 4.8, which constitute the paper's headline non-monotonicity results, are proved only for (4.2). The approximation matches the N→0 and N→∞ limits, yet the extensive-margin switching between ICL and SFT is governed by crossings of E_ICL and E_SFT, and those crossings can shift under the approximation. If the true curves differ at finite N, the increase–decrease–increase pattern in r, the hump in R_SFT, and the anchoring regions could be artifacts of (4.2). The limitations section does not flag this gap. This is load-bearing for Finding 2, even though Theorem 4.1 and Theorem 5.5 are derived from the exact (4.1).
  2. [Theorem 5.5 and Assumption 5.4] The theorem is stated as 'adding SFT cannot reduce optimal profit', and the abstract repeats this without qualification. The proof, however, relies on Assumption 5.4, which requires RSFT ≥ RICL · sup_t (rβ(t)/(d−r))^{1/α} with β(t)=1+π(τ+m²)/σ̄². This is a strong uniform per-sample compute gap that is not derived from any measurement or primitive in the paper. The theorem is therefore conditional on an assumption that may fail for platforms where SFT per-sample costs are not sufficiently larger than ICL per-sample costs. The main text should state the theorem with this condition prominently, and the abstract/introduction should not claim an unconditional result. This is a scope issue for a central claim, not merely a stylistic one.
  3. [Section 4.2, Figures 2–3] The simulation study that illustrates Propositions 4.4–4.6 computes equilibrium congestion using the simplified errors (4.2) itself. As a result, the figures do not provide evidence that the non-monotonicity survives under the exact error model (4.1). To support the approximation, the authors should either overlay the exact-model simulations in Figures 2–3 or provide analytic bounds showing that the comparative statics are preserved uniformly in N. Without this, the empirical content of the non-monotonicity results rests entirely on an unproven simplification.
minor comments (5)
  1. [Section 4.1 after Eq. (4.1)] The user type is defined as t=(d,r,σ̄,σ̃,m,π,τ), but the error formulas use s=N^α, which depends on the action N. Clarify that s is not part of the type but is determined by the action, or include α in the type.
  2. [Assumption 4.2 vs. Assumption 5.4] Assumption 4.2 only requires RSFT>RICL, while Assumption 5.4 imposes a much larger gap. The relationship between these assumptions should be discussed; the latter is not a strengthening of the former in an obvious quantitative sense and its necessity for Theorem 5.5 should be explained.
  3. [Appendix I] The robustness check with h(x)=max{0,x} is not a separate congestion function on R_+ since max{0,x}=x for x≥0. This is a linear congestion function, which is already covered by the assumptions; consider replacing it with a different nonlinearity or clarifying its role.
  4. [Figures 8–9] The labels 'Regime of MFE' are used without defining MFE; the text refers to 'regime of the equilibrium'. Use consistent terminology.
  5. [Section 6] The GPT-2 experiments test only the statistical comparisons of Section 3 (ICL plateau, linear bias in d−r, SFT/ICL ranking flip). They do not test the equilibrium non-monotonicity or the platform profit prediction. This should be stated explicitly so readers do not infer that the experiments validate the congestion-game results.

Circularity Check

0 steps flagged

No circular derivation found; the central SFT/ICL threshold and the profit theorem are derived from explicit primitives, while the main unresolved issue is an unvalidated simplifying approximation, not a circularity.

full rationale

The core derivation chain is self-contained. Theorem 3.3 derives closed-form ICL and SFT errors from the Gaussian linear model, the Bayesian-update definition of ICL, the regularized-estimator definition of SFT, and Assumptions 3.1, 3.2 and 3.5. Proposition 3.6 then computes the SFT/ICL threshold Rcrit as an explicit function of the primitives (π, s, τ, σ̃, σ̄, m, d, r); no parameter is fitted to data and no output is reused as an input. The equilibrium and platform sections are also derived in-paper: Theorem 4.1 proves existence and congestion-level uniqueness with a monotonicity lemma, Theorem 4.3 characterizes the homogeneous-user equilibrium from the error/cost functions, and Theorem 5.5 relies on the exact errors (4.1) plus Assumption 5.4 to establish R*(p) ≥ R*_ICL(p) by a valid monotonicity argument. The paper cites prior work by overlapping authors (notably Zhang et al. 2023a,b), but the load-bearing ICL-as-Bayesian-inference premise is also supported by the external Xie et al. 2021 citation and by the GPT-2 experiments, which test qualitative predictions (plateau, linear bias in d−r, ranking flip) rather than calibrating the model constants. Thus the self-citations are not load-bearing. The only notable weakness is in Section 4.2: the paper replaces the exact errors (4.1) with the simplified errors (4.2) and asserts that this 'preserves the main insights' without supplying an error bound at intermediate N. Propositions 4.4–4.6 are proved for this approximation. This is a correctness/robustness risk, not circularity, and it does not affect Theorem 5.5, whose proof is based on the exact expressions (4.1). Overall, no step reduces by construction to its own input, so the circularity score is minimal.

Axiom & Free-Parameter Ledger

3 free parameters · 8 axioms · 0 invented entities

The framework is a chain: linear statistical model → closed-form errors → approximated errors → mean-field congestion game → Stackelberg pricing. Each link adds assumptions: spectral alignment and isotropicity make the error formulas tractable; the (4.2) approximation and h=R² make the equilibrium analysis closed-form; Assumption 5.4 makes the profit theorem hold. No free parameters are fitted to data in the derivations; the free choices are functional forms and the approximation. No new physical entities, forces, or conserved quantities are postulated: the user type space, congestion level R, and the platform are modeling constructs with defined operational meaning.

free parameters (3)
  • information-growth exponent α (s = N^α) = α = 1 assumed in Sections 4.2–5
    Models how personalization data strength s grows with sample count; α=1 (i.i.d. Gaussian covariates) is adopted for analytical convenience and all closed-form equilibrium results depend on it (Section 4.2).
  • simplified error functions (4.2) constants = 2σ̃²r, 2σ̃²/ζ, 2σ̃²d (chosen to match N→∞/N→0 asymptotics of (4.1))
    The error functions used in every equilibrium theorem are hand-chosen approximations; their constants are matched to asymptotics of the exact formulas (4.1) and are not derived or bounded at intermediate N (Section 4.2).
  • congestion function h(R)=R² = h(R)=R² (exp and max{0,x} checked numerically in Appendix I)
    Adopted 'for the ease of calculation' (Section 4.2); the paper verifies robustness only numerically for two alternative functional forms.
axioms (8)
  • domain assumption Assumption 3.1: X^TX, X̃^TX̃, and Σ* share a common eigenbasis V (spectral alignment)
    Reduces the error analysis to d independent one-dimensional problems; without it, the covered/uncovered decomposition that drives the ICL-vs-SFT mechanism does not hold. Invoked in Theorem 3.3.
  • domain assumption Assumption 3.2: pretraining inputs have constant Σθ-norm
    Folds input power into a single scalar σ̄² = σ²+c²; heterogeneous norms are claimed to require only an averaged σ̄² without changing the qualitative results.
  • domain assumption Assumption 3.5: isotropicity (τ_i=τ, s_i=s, π_i=π, |v_i^T μ_θ|=m)
    Needed for the closed-form threshold R_crit in Proposition 3.6; the authors call it 'a structural simplification, not a claim of truly isotropic pretrained features.'
  • ad hoc to paper Simplified errors (4.2) preserve the ICL/SFT comparison and equilibrium structure for all N
    Asserts fidelity of the approximated errors used in all of Sections 4–5; matched at N→∞ and N→0 but unproven and unbounded for intermediate N.
  • ad hoc to paper Assumption 5.4: R_SFT ≥ R_ICL · sup_t (rβ(t)/(d−r))^{1/α}
    The compute-gap condition that makes Theorem 5.5 ('adding SFT never reduces profit') go through; the abstract states the theorem without this condition.
  • domain assumption ICL performs Bayesian inference under the learned pretraining prior
    Imported from Xie et al. 2021 and Zhang et al. 2023b; the entire ICL error formula rests on the posterior interpretation (Section 2.1).
  • domain assumption Continuum of users with compact type space and regular congestion h (strictly increasing, unbounded)
    Enables the mean-field equilibrium; core to Definition 2.1 and Theorem 4.1.
  • domain assumption LLM behavior on linear tasks is broadly representative of LLM personalization
    The empirical bridge that lets GPT-2 linear-regression results count as validation; asserted in Section 1, cited to Garg et al. 2022 / Akyürek et al. 2023.

pith-pipeline@v1.3.0-alltime-deepseek · 61399 in / 29221 out tokens · 250228 ms · 2026-08-02T02:18:19.076643+00:00 · methodology

0 comments
read the original abstract

Large Language Models (LLMs) have revolutionized AI services, but a critical tension emerges: while personalization improves model performance, it consumes scarce computational resources that users must share. When should a user invest in expensive Supervised Fine-Tuning (SFT) versus lightweight In-Context Learning (ICL)? How does congestion from other users' personalization choices reshape these incentives? And what strategies should platforms adopt when offering multiple personalization algorithms? We develop a tractable framework for LLM serving that captures the statistical-economic trade-offs users face. Our analysis yields several surprising insights. First, we show that ICL and SFT dominate in different regimes, determined by an interplay between pretraining coverage and data signal-to-noise ratios, but congestion can flip these rankings. Second, equilibrium resource consumption exhibits pronounced non-monotonicity: improving pretraining precision reduces the congestion, while broader pretraining coverage and harder tasks sometimes increase it. Third, we prove that offering both personalization methods never hurts the platform's maximal profits, despite potentially increasing computational load. Experiments with GPT-2 on linear regression tasks validate our theoretical predictions about algorithm performance. Complementing these results, our review of documentation from 21 major AI platforms shows that the share offering both SFT and ICL increased from 9.5% in 2021 to 71.4% in 2025, consistent with our platform-design implications.

Figures

Figures reproduced from arXiv: 2607.14371 by Dirk Bergemann, Fengzhuo Zhang, Zhuoran Yang.

Figure 1
Figure 1. Figure 1: This figure illustrates the structure of the equilibrium with homogeneous users. The intersection point of the line y = H and the blue dotted curve corresponds to the equilibrium of the game. Theorem 4.3 characterizes the equilib￾rium with homogeneous users by reduc￾ing the equilibrium problem to a one￾dimensional fixed point in the effective per￾unit compute cost H = p + h(R), illustrated in [PITH_FULL_I… view at source ↗
Figure 2
Figure 2. Figure 2: The figures show how the equilibrium congestion level [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The figures show how the equilibrium congestion level [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The separator H∗ sep(t) summarizes a type’s willingness to pay in resource costs to use the more resource-intensive method. The ordering H∗ sep(t2) < H∗ sep(t1) implies that [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The figure reports |∂qR∗ | + |∂σ˜1R∗ | across different values of the type-1 task noise σ˜1 and population share q. In the dark-blue region, the near-zero sensitivity indicates that the equilibrium congestion level is pinned by the type-2 separator and is locally insensitive to changes in σ˜1 and q. regime, R∗ = 0 is a stable zero-congestion corner: both types strictly prefer ICL and choose N˜ ICL = 0. The… view at source ↗
Figure 6
Figure 6. Figure 6: The figures show the estimation error of [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Panel (a) shows the estimation error of SFT across different numbers of personal￾ization samples N˜, illustrating that SFT can asymptotically drive the estimation error to zero as the sample size increases. Panel (b) reports the fraction of platforms that offer SFT in addition to ICL, showing that support for SFT-based personalization has increased in recent years. pursue when computational resources are s… view at source ↗
Figure 8
Figure 8. Figure 8: Figures (a)-(c) show the congestion level [PITH_FULL_IMAGE:figures/full_fig_p085_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Figures (a)-(c) show the congestion level [PITH_FULL_IMAGE:figures/full_fig_p085_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: These figures illustrate how the equilibrium congestion level [PITH_FULL_IMAGE:figures/full_fig_p086_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: These figures illustrate how the equilibrium congestion level [PITH_FULL_IMAGE:figures/full_fig_p086_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: These figures illustrate how the equilibrium congestion level [PITH_FULL_IMAGE:figures/full_fig_p087_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: These figures illustrate how the equilibrium congestion level [PITH_FULL_IMAGE:figures/full_fig_p087_13.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

296 extracted references · 13 linked inside Pith

  1. [1]

    Abril and Robert Plant

    Patricia S. Abril and Robert Plant. The patent holder's dilemma: Buy, sell, or troll?. Communications of the ACM. 2007. doi:10.1145/1188913.1188915

    arXiv 2007

  2. [2]

    Deciding equivalances among conjunctive aggregate queries

    Sarah Cohen and Werner Nutt and Yehoshua Sagic. Deciding equivalances among conjunctive aggregate queries. 2007. doi:10.1145/1219092.1219093

    arXiv 2007

  3. [3]

    Special issue: Digital Libraries. 1996

    1996

  4. [4]

    Understanding Policy-Based Networking

    David Kosiur. Understanding Policy-Based Networking. 2001

    2001

  5. [7]

    The title of book two. 2008. doi:10.1007/3-540-09237-4

    doi:10.1007/3-540-09237-4 2008

  6. [8]

    Asad Z. Spector. Achieving application requirements. Distributed Systems. 1990. doi:10.1145/90417.90738

    arXiv 1990

  7. [9]

    Douglass and David Harel and Mark B

    Bruce P. Douglass and David Harel and Mark B. Trakhtenbrot. Statecarts in use: structured analysis and object-orientation. Lectures on Embedded Systems. 1998. doi:10.1007/3-540-65193-4_29

    doi:10.1007/3-540-65193-4_29 1998

  8. [10]

    Donald E. Knuth. The Art of Computer Programming, Vol. 1: Fundamental Algorithms (3rd. ed.). 1997

    1997

  9. [11]

    Donald E. Knuth. The Art of Computer Programming. 1998

    1998

  10. [12]

    Journal of Urban Economics , volume=

    Road congestion: a reconsideration of pricing theory , author=. Journal of Urban Economics , volume=. 1974 , publisher=

    1974

  11. [13]

    2024 , publisher=

    The economics of urban transportation , author=. 2024 , publisher=

    2024

  12. [14]

    Handbook of regional and urban economics , volume=

    Transportation costs and the spatial organization of economic activity , author=. Handbook of regional and urban economics , volume=. 2015 , publisher=

    2015

  13. [15]

    American Economic Review , volume=

    The fundamental law of road congestion: Evidence from US cities , author=. American Economic Review , volume=. 2011 , publisher=

    2011

  14. [16]

    2015 , publisher=

    The traffic assignment problem: models and methods , author=. 2015 , publisher=

    2015

  15. [17]

    Econometrica , volume=

    Peak-Hour Road Congestion Pricing: Experimental Evidence and Equilibrium Implications , author=. Econometrica , volume=. 2024 , publisher=

    2024

  16. [18]

    Structured Variational Inference Procedures and their Realizations (as incol)

    Dan Geiger and Christopher Meek. Structured Variational Inference Procedures and their Realizations (as incol). Proceedings of Tenth International Workshop on Artificial Intelligence and Statistics, The Barbados

  17. [19]

    Stan W. Smith. An experiment in bibliographic mark-up: Parsing metadata for XML export. Proceedings of the 3rd. annual workshop on Librarians and Computers. 2010. doi:99.9999/woot07-S422

    2010

  18. [20]

    arXiv preprint arXiv:2408.14511 , year=

    Unveiling the statistical foundations of chain-of-thought prompting methods , author=. arXiv preprint arXiv:2408.14511 , year=

    Pith/arXiv arXiv

  19. [21]

    SIAM Journal on Control and Optimization , volume=

    Probabilistic analysis of mean-field games , author=. SIAM Journal on Control and Optimization , volume=. 2013 , publisher=

    2013

  20. [22]

    2018 , publisher=

    Probabilistic theory of mean field games with applications I-II , author=. 2018 , publisher=

    2018

  21. [23]

    Advances in neural information processing systems , volume=

    Locating and editing factual associations in gpt , author=. Advances in neural information processing systems , volume=

  22. [24]

    arXiv preprint arXiv:2509.26030 , year=

    Muon outperforms Adam in tail-end associative memory learning , author=. arXiv preprint arXiv:2509.26030 , year=

    arXiv

  23. [25]

    Journal of Machine Learning Research , volume=

    Learning regularized graphon mean-field games with unknown graphons , author=. Journal of Machine Learning Research , volume=

  24. [26]

    Catch me, if you can: Evading network signatures with web-based polymorphic worms

    Matthew Van Gundy and Davide Balzarotti and Giovanni Vigna. Catch me, if you can: Evading network signatures with web-based polymorphic worms. Proceedings of the first USENIX workshop on Offensive Technologies. 2007

    2007

  25. [27]

    Catch me, if you can: Evading network signatures with web-based polymorphic worms

    Matthew Van Gundy and Davide Balzarotti and Giovanni Vigna. Catch me, if you can: Evading network signatures with web-based polymorphic worms. Proceedings of the first USENIX workshop on Offensive Technologies. 2008

    2008

  26. [28]

    Catch me, if you can: Evading network signatures with web-based polymorphic worms

    Matthew Van Gundy and Davide Balzarotti and Giovanni Vigna. Catch me, if you can: Evading network signatures with web-based polymorphic worms. Proceedings of the first USENIX workshop on Offensive Technologies. 2009

    2009

  27. [29]

    Predicate Path expressions

    Sten Andler. Predicate Path expressions. Proceedings of the 6th. ACM SIGACT-SIGPLAN symposium on Principles of Programming Languages. 1979. doi:10.1145/567752.567774

    arXiv 1979

  28. [30]

    LOGICS of Programs: AXIOMATICS and DESCRIPTIVE POWER

    David Harel. LOGICS of Programs: AXIOMATICS and DESCRIPTIVE POWER. 1978

    1978

  29. [31]

    Anisi , title =

    David A. Anisi , title =

  30. [32]

    Clarkson

    Kenneth L. Clarkson. Algorithms for Closest-Point Problems (Computational Geometry). 1985

    1985

  31. [33]

    Introduction to Bayesian Statistics

    Harry Thornburg. Introduction to Bayesian Statistics. 2001

    2001

  32. [34]

    CLIFFORD: a Maple 11 Package for Clifford Algebra Computations, version 11

    Rafal Ablamowicz and Bertfried Fauser. CLIFFORD: a Maple 11 Package for Clifford Algebra Computations, version 11. 2007

    2007

  33. [35]

    Stats and Analysis

    Poker-Edge.Com. Stats and Analysis. 2006

    2006

  34. [36]

    A more perfect union

    Barack Obama. A more perfect union. 2008

    2008

  35. [37]

    The fountain of youth

    Joseph Scientist. The fountain of youth. 2009

    2009

  36. [38]

    Solder man

    Dave Novak. Solder man. ACM SIGGRAPH 2003 Video Review on Animation theater Program: Part I - Vol. 145 (July 27--27, 2003). 2003. doi:99.9999/woot07-S422

    2003

  37. [39]

    Interview with Bill Kinder: January 13, 2005

    Newton Lee. Interview with Bill Kinder: January 13, 2005. Comput. Entertain. 2005. doi:10.1145/1057270.1057278

    arXiv 2005

  38. [40]

    The Enabling of Digital Libraries

    Bernard Rous. The Enabling of Digital Libraries. Digital Libraries. 2008

    2008

  39. [42]

    (new) Finding minimum congestion spanning trees , journal =

    Werneck, Renato and Setubal, Jo\. (new) Finding minimum congestion spanning trees , journal =. 2000 , issn =. doi:10.1145/351827.384253 , acmid =

    arXiv 2000

  40. [44]

    and Mei, Alessandro , title =

    Conti, Mauro and Di Pietro, Roberto and Mancini, Luigi V. and Mei, Alessandro , title =. Inf. Fusion , volume =. 2009 , issn =. doi:10.1016/j.inffus.2009.01.002 , acmid =

    doi:10.1016/j.inffus.2009.01.002 2009

  41. [45]

    and Hutchful, David K

    Li, Cheng-Lun and Buyuktur, Ayse G. and Hutchful, David K. and Sant, Natasha B. and Nainwal, Satyendra K. , title =. CHI '08 extended abstracts on Human factors in computing systems , year =. doi:10.1145/1358628.1358946 , acmid =

    arXiv

  42. [46]

    , title =

    Hollis, Billy S. , title =. 1999 , isbn =

    1999

  43. [47]

    Goossens, Michel and Rahtz, S. P. and Moore, Ross and Sutor, Robert S. , title =. 1999 , isbn =

    1999

  44. [48]

    and Rosenberg, Arnold L

    Buss, Jonathan F. and Rosenberg, Arnold L. and Knott, Judson D. , title =. 1987 , source =

    1987

  45. [49]

    CHI '08: CHI '08 extended abstracts on Human factors in computing systems , year =

    , note =. CHI '08: CHI '08 extended abstracts on Human factors in computing systems , year =

  46. [50]

    Algorithms for Closest-Point Problems (Computational Geometry) , year =

    Clarkson, Kenneth Lee , advisor =. Algorithms for Closest-Point Problems (Computational Geometry) , year =

  47. [51]

    SIGCOMM Comput. Commun. Rev. , year =

  48. [52]

    2004 , isbn =

    IEEE TCSC Executive Committee , booktitle =. 2004 , isbn =. doi:http://dx.doi.org/10.1109/ICWS.2004.64 , acmid =

    doi:10.1109/icws.2004.64 2004

  49. [53]

    Distributed systems (2nd Ed.) , year =

  50. [54]

    , title =

    Petrie, Charles J. , title =. 1986 , source =

    1986

  51. [55]

    Donald E. Knuth. Seminumerical Algorithms. 1981

    1981

  52. [56]

    E-commerce and cultural values , year =

    Kong, Wei-Chang , Title =. E-commerce and cultural values , year =

  53. [57]

    E-commerce and cultural values , year =

    Kong, Wei-Chang , type =. E-commerce and cultural values , year =

  54. [58]

    Chapter 9 , booktitle =

    Kong, Wei-Chang , editor =. Chapter 9 , booktitle =. 2002 , address =

    2002

  55. [59]

    E-commerce and cultural values , editor =

    Kong, Wei-Chang , title =. E-commerce and cultural values , editor =. 2003 , isbn =

    2003

  56. [60]

    E-commerce and cultural values - (InBook-num-in-chap) , chapter =

    Kong, Wei-Chang , editor =. E-commerce and cultural values - (InBook-num-in-chap) , chapter =. 2004 , address =

    2004

  57. [61]

    E-commerce and cultural values (Inbook-text-in-chap) , chapter =

    Kong, Wei-Chang , editor =. E-commerce and cultural values (Inbook-text-in-chap) , chapter =. 2005 , address =

    2005

  58. [62]

    E-commerce and cultural values (Inbook-num chap) , chapter =

    Kong, Wei-Chang , editor =. E-commerce and cultural values (Inbook-num chap) , chapter =. 2006 , address =

    2006

  59. [63]

    Microelectron

    Mehdi Saeedi and Morteza Saheb Zamani and Mehdi Sedighi , title =. Microelectron. J. , volume =. 2010 , pages =

    2010

  60. [64]

    Mehdi Saeedi and Morteza Saheb Zamani and Mehdi Sedighi and Zahra Sasanian , title =. J. Emerg. Technol. Comput. Syst. , volume =

  61. [65]

    Kirschmer, Markus and Voight, John , title =. SIAM J. Comput. , issue_date =. 2010 , issn =. doi:http://dx.doi.org/10.1137/080734467 , acmid =

    doi:10.1137/080734467 2010

  62. [66]

    Hoare, C. A. R. , title =. Structured programming (incoll) , editor =. 1972 , isbn =

    1972

  63. [67]

    History of programming languages I (incoll) , editor =

    Lee, Jan , title =. History of programming languages I (incoll) , editor =. 1981 , isbn =. doi:http://doi.acm.org/10.1145/800025.1198348 , acmid =

    arXiv 1981

  64. [68]

    , title =

    Dijkstra, E. , title =. Classics in software engineering (incoll) , year =

  65. [69]

    , title =

    Wenzel, Elizabeth M. , title =. Multimedia interface design (incoll) , year =. doi:10.1145/146022.146089 , acmid =

    arXiv

  66. [70]

    , title =

    Mumford, E. , title =. Critical issues in information systems research (incoll) , year =

  67. [71]

    and Golden, Donald G

    McCracken, Daniel D. and Golden, Donald G. , title =. 1990 , isbn =

    1990

  68. [72]

    The analysis of linear partial differential operators

    H. The analysis of linear partial differential operators. 1985 , PAGES =

    1985

  69. [73]

    IEEE", address =

    A. Adya and P. Bahl and J. Padhye and A.Wolman and L. Zhou , title =. Proceedings of the IEEE 1st International Conference on Broadnets Networks (BroadNets'04) , publisher = "IEEE", address = "Los Alamitos, CA", year =

  70. [74]

    I. F. Akyildiz and W. Su and Y. Sankarasubramaniam and E. Cayirci , title =. Comm. ACM , volume = 38, number = "4", year =

  71. [75]

    I. F. Akyildiz and T. Melodia and K. R. Chowdhury , title =. Computer Netw. , volume = 51, number = "4", year =

  72. [76]

    ACM", address =

    P. Bahl and R. Chancre and J. Dungeon , title =. Proceeding of the 10th International Conference on Mobile Computing and Networking (MobiCom'04) , publisher = "ACM", address = "New York, NY", year =

  73. [77]

    8 (Special Issue on Sensor Networks)

    D. Culler and D. Estrin and M. Srivastava , title =. IEEE Comput. , volume = 37, number = "8 (Special Issue on Sensor Networks)", publisher = "IEEE", address = "Los Alamitos, CA", year =

  74. [78]

    Natarajan and M

    A. Natarajan and M. Motani and B. de Silva and K. Yap and K. C. Chua , title =. Network Architectures , editor =. 960935712

  75. [79]

    Tzamaloukas and J

    A. Tzamaloukas and J. J. Garcia-Luna-Aceves , title =

  76. [80]

    Zhou and J

    G. Zhou and J. Lu and C.-Y. Wan and M. D. Yarvis and J. A. Stankovic , title =

  77. [81]

    Mapping Powerlists onto Hypercubes

    Jacob Kornerup. Mapping Powerlists onto Hypercubes. 1994

    1994

  78. [82]

    Automatic Parallelization for Distributed-Memory Multiprocessing Systems

    Michael Gerndt. Automatic Parallelization for Distributed-Memory Multiprocessing Systems

  79. [83]

    J. E. Archer, Jr. and R. Conway and F. B. Schneider. User recovery and reversal in interactive systems. ACM Trans. Program. Lang. Syst

  80. [84]

    D. D. Dunlop and V. R. Basili. Generalizing specifications for uniformly implemented loops. ACM Trans. Program. Lang. Syst

Showing first 80 references.