Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Predictive Monte Carlo recovers the implicit prior and posterior a Bayes-filtered transformer has internalized, using only the model's own next-token generation.

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

2026-08-01 19:07 UTC pith:EHKPLABO

load-bearing objection Worth sending to review: a clear, honest paper that moves ICL evaluation into latent space with PMC, but the central interpretability claim is conditional on unverified representation-theorem conditions that the authors explicitly flag. the 3 major comments →

arxiv 2607.17060 v1 pith:EHKPLABO submitted 2026-07-19 cs.LG

What does a Bayes-filtered transformer believe? A predictive Monte Carlo approach

Afiq Abdillah Effiezal Aswadi , Haotong Ma , Susan Wei This is my paper
classification cs.LG MSC 62F1560G0968T07
keywords Bayes-filtered transformerpredictive Monte Carloin-context learninglatent-space interpretabilityexchangeabilityMarkov exchangeabilitytask-diversity thresholdtransient generalization
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.

This paper is trying to establish that a Bayes-filtered transformer — a transformer trained on sequences generated by first drawing a latent task from a prior and then sampling observations conditional on that task — internalizes a definite prior and posterior over that latent task, and that predictive Monte Carlo (PMC) can recover them using nothing but the model's own next-token generation. The motivation is that the usual way of interpreting such models, by comparing their predictions against candidate reference posteriors, is ambiguous: distinct priors and posteriors can yield identical prediction means, so the interpretive question is moved into latent space, where the beliefs actually live. The method rolls out sequences from the trained model, reads the latent task off each completed rollout (empirical frequencies, a regression fit, or transition counts), and treats the empirical distribution of these readings as the model's implicit prior (from an empty prompt) or implicit posterior (from a given prompt). Applied to three task families covering both exchangeable and Markov-exchangeable processes (the paper's 0-Markov and 1-Markov cases), the paper reproduces the task-diversity threshold and transient generalization directly in latent space, with the recovered prior and posterior matching closed-form references on in-distribution prompts.

Core claim

A trained Bayes-filtered transformer is, in the ideal limit, the Bayesian posterior predictive of its pretraining model; real training breaks that ideal. The paper's central claim: predictive Monte Carlo can read off what prior and posterior this particular model has actually internalized. Generate continuations using only the model's next-token probabilities; read the latent task off each completed continuation (empirical frequencies, OLS fit, transition counts); the empirical distribution of those readings is the implicit prior (empty prompt) or implicit posterior (given prompt). Representation theorems — for exchangeable and Markov-exchangeable sequences — make the latent task a determini

What carries the argument

The load-bearing mechanism is the pairing of a representation theorem with a rollout estimator. For the exchangeable families (balls-and-urns, linear regression), the representation theorem for exchangeable sequences makes the latent task — urn proportions, regression vector — a deterministic function of the infinite sample path, realized as empirical frequencies or an ordinary least squares fit. For the Markov-chain family, the analogous theorem for Markov-exchangeable chains makes the transition matrix such a function, given a recurrence condition. Predictive Monte Carlo replaces the infinite path with a finite rollout drawn autoregressively from the trained transformer using forward passe

Load-bearing premise

The entire latent-space interpretation rests on the unverified premise that the trained transformer's own sequence distribution satisfies the relevant representation theorem (exchangeability for the 0-Markov families; Markov exchangeability plus recurrence for Markov chains); the paper states explicitly in its appendix that it does not verify these conditions for the models it studies.

What would settle it

Roll out one very long sequence from the trained linear-regression BFT and test whether its sequence of one-step predictive distributions converges almost surely to a random measure (equivalently, whether the predictives form a martingale). If they do not converge, PMC samples lose their status as draws from an implicit posterior, and the paper's unexplained out-of-distribution mismatch is a symptom of invalidity rather than of genuinely non-Bayesian belief. A complementary check: build two priors with identical posterior means but different shapes and see whether PMC, unlike prediction-space

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

If this is right

  • The task-diversity threshold — where a model shifts from memorizing to generalizing behavior as the number of pretraining tasks grows — is a real shift in the model's latent beliefs, not an artifact of the measurement space.
  • Transient generalization at intermediate task diversity — the model first approaching the generalizing posterior and later drifting to the memorizing one — shows up in the recovered posterior, so it is a belief-level phenomenon.
  • Because PMC needs only forward passes, it extends to any Bayes-filtered transformer whose next-token distribution is available, including models too large or too opaque for closed-form posterior analysis.
  • Where the PMC-recovered distribution matches a closed-form reference (the training prior; the memorizing posterior on in-distribution prompts), the match itself certifies the recovery without requiring verification of the abstract sufficient conditions.
  • Unifying the three task families as k-Markov exchangeable Bayes-filtered transformers gives a common vocabulary for studying in-context learning phenomena across different generative mechanisms.

Where Pith is reading between the lines

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

  • Beyond the paper's claims: the unexplained mismatch between the PMC posterior and both reference posteriors on out-of-distribution linear-regression prompts is itself testable — if the validity conditions can be verified, that mismatch would indicate the model's latent beliefs genuinely depart from both baselines when extrapolating, which would be a substantive finding about how these models gener
  • The paper's suggested persona-tracking experiment generalizes: any suspected belief drift during finetuning — for example misalignment induced by narrow finetuning — could be monitored by running PMC before, during, and after the intervention, turning a qualitative suspicion into a measurable shift in a latent distribution.
  • The weaker match in the Markov-chain family (a high-dimensional latent task) suggests a practical scaling rule: the number of rollouts needed for a faithful latent-space readout grows with the latent dimension, which practitioners can calibrate by comparing PMC samples to known references.
  • Because PMC produces a target distribution without needing ground truth, it offers a natural validation target for latent probes of the residual stream, connecting predictive-level beliefs to internal representations — a link the paper leaves as future work.

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. This paper proposes predictive Monte Carlo (PMC) as a black-box interpretability tool for Bayes-filtered transformers (BFTs), claiming to recover the implicit prior and posterior over the latent task using only next-token generation. The authors train BFTs on three synthetic families (balls-and-urns, linear regression, Markov chains), apply PMC, and compare the recovered latent distributions against memorizing (empirical prior) and generalizing (population prior) baselines using energy distance and sliced Wasserstein distance. They report that previously observed phenomena (task-diversity threshold, transient generalization) appear in latent space. The central validity requirement—that the transformer-induced joint law P_phi satisfies exchangeability or Markov exchangeability/recurrence—is explicitly not verified, and an out-of-distribution linear-regression mismatch is left unexplained.

Significance. If the central claim holds, the paper makes a useful methodological contribution: it unifies several synthetic ICL studies as k-Markov exchangeable BFTs, moves evaluation from prediction space to latent space, and offers a model-agnostic, non-circular way to probe internalized beliefs, with external analytic references as checks. The experiments are carefully specified, the appendices are thorough, code is provided, and the paper is unusually honest about its limitations. However, the significance is heavily conditional on PMC validity, which the paper does not establish. The 'agreement certifies PMC output' argument covers only regimes with known references, not the regime where the tool is most needed. The explicit statement that 'We do not verify any of these conditions for the BFTs studied here' is a central caveat, not a peripheral one, and it should gate the strength of the claims.

major comments (3)
  1. [§4 Validity; Appendix C] PMC samples are interpretable as draws from the BFT's implicit posterior only if P_phi satisfies the hypotheses of de Finetti's or Diaconis–Freedman's theorem. Appendix C states: 'We do not verify any of these conditions for the BFTs studied here.' Without verification, finite-rollout empirical frequencies, OLS estimates, and transition counts are not established as draws from an implicit posterior. The agreement on prior and in-distribution prompts is encouraging but certifies only those regimes; the OOD linear-regression regime, where no reference matches, is exactly where the interpretation is needed. Please either supply diagnostics for P_phi (e.g., predictive martingale or conditional-identically-distributed checks for the trained models) or weaken the abstract and contributions to claim recovery only in validated regimes.
  2. [§5 Linear regression; Fig. 9 / App. F] On an out-of-distribution prompt, the PMC-recovered posterior matches neither the memorizing nor the generalizing baseline. The paper states: 'Two explanations are consistent with this mismatch: the trained model may genuinely depart from both reference posteriors, or the sufficient conditions for PMC validity ... may fail for P_phi. We do not know which explanation is operative.' This unresolved ambiguity is load-bearing because it leaves open that PMC samples are merely rollout statistics rather than implicit beliefs. A concrete test would be to check whether the model's one-step predictives along rollouts satisfy an asymptotic-exchangeability or martingale-type condition, or to compare PMC against an independent latent-space probe on the same OOD prompts. Without such a test, the central claim is not established beyond matched regimes.
  3. [Appendix H; Eq. (7)] The Markov-chain baselines are not the exact posterior under the generative process. Appendix H notes that 'the exact posterior under the generative process carries a stationary-probability factor that the baselines omit, one of ℓ=8 likelihood contributions.' Because both memorizing and generalizing baselines omit the P(y1|Q) factor, the latent-space distances and the reported 'looser match' may reflect baseline misspecification rather than BFT behavior. This weakens the validation of PMC in the 1-Markov family. Please quantify the effect (e.g., by also computing baselines with the stationary factor) or justify the omission specifically for the comparisons being made.
minor comments (5)
  1. [§3, Eq. (4)] The KL in Eq. (4) is written between P(Y1:n) and P_phi(Y1:n), but P_phi is initially defined as a law on Y^∞. Please clarify that the KL is taken between the n-dimensional marginals.
  2. [§5, notation] The statement 'Throughout the experiments, P denotes the joint law (2) under Π = Π∞' overloads the symbol P, which was used generically in Section 2. Consider writing P_∞ for the evaluation distribution to avoid ambiguity in Eqs. (8) and (14).
  3. [Appendix F, Fig. 8] The Autoregressive-PFN attention mask is described in words and shown in Figure 8, but it is not stated whether the first even-position token (x1,0) attends only to itself and whether its output logits are used for predicting y1. Please make this explicit.
  4. [Appendix G] The explanation for the absent prediction-space transition in balls-and-urns says that 'in early positions in the sequence, the transformer behaves closer to the memorizing predictor.' This claim is not supported by a displayed figure; please add evidence or rephrase as a hypothesis.
  5. [Appendix H] The Markov-chain PMC evaluation uses only 16 prompts per source, because the memorizing baseline likelihood scales with M. This is reasonable, but the small prompt count should be acknowledged as a source of variance in the reported latent-space distances.

Circularity Check

0 steps flagged

No significant circularity: PMC outputs are benchmarked against external analytic references and disagreements are reported; the main caveat is an unverified validity condition, which is a verification gap rather than a circular reduction.

full rationale

The central recovery step is not a fitted input called a prediction: PMC samples rollouts from the trained transformer and evaluates fixed path functionals (empirical frequency, OLS, empirical transition matrix) without tuning parameters to force agreement with the reference posteriors (Section 4). The empirical checks are genuinely falsifiable: at task diversity M=8 the linear-regression PMC prior concentrates on the training prior ΠM rather than the population prior Π∞ (Fig. 3b), and the out-of-distribution posterior is reported to match neither reference (Section 5, Fig. 9). The paper explicitly identifies the load-bearing condition for latent-space interpretation: "The recovery of θ̃ above rests on Pφ satisfying the hypotheses of the relevant representation theorem... a trained BFT only approximates those predictives, so the hypotheses must be established for Pφ separately" (Section 4, Validity). Appendix C then states honestly: "We do not verify any of these conditions for the BFTs studied here." This is a verification/assumption gap, not a circularity, because the paper does not use its conclusion as a premise and does not claim the conditions have been established. Self-citations ([16], [17], [29], sharing author Wei) appear in related work and limitations, but they are not load-bearing: the PMC procedure and representation theorems come from external references [8, 10, 14, 15], and the self-cited work is described as contextual or inconclusive. Hence score 0 for circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or mathematical entities. The latent tasks (urn probabilities, regression weights, transition matrices) are standard. The load-bearing assumptions are the unverified PMC-validity conditions on the trained model's induced stochastic process, plus the usual representation theorems.

free parameters (4)
  • PMC rollout length N = 56 (linear regression), 255-ell (balls-and-urns), 400 (Markov chains)
    Truncates the infinite path; chosen by hand per family.
  • Number of PMC rollouts R = 100 (posterior) to 3000 (prior)
    Controls Monte Carlo error; chosen by hand.
  • Laplace smoothing coefficient = 1
    For Markov transition matrix estimator; chosen by hand.
  • Linear response discretization = 256 bins over [-10,10]
    Enables categorical output head; chosen by hand.
axioms (4)
  • domain assumption de Finetti representation (0-Markov) and Diaconis-Freedman representation (1-Markov) apply to the trained model's law P_phi
    Required for interpreting finite-rollout empirical functionals as samples of a latent task. Appendix C.
  • domain assumption P_phi satisfies sufficient conditions for PMC validity (asymptotic exchangeability / Markov-exchangeable + recurrence)
    The paper states these are not verified. Section 4 and Appendix C.
  • domain assumption Trained BFT approximates the Bayesian PPD in the ideal limit
    Definition 1 and Section 3; needed to motivate the interpretive question.
  • ad hoc to paper The baseline posteriors omit the stationary-distribution factor for Markov chains
    Appendix H notes this omission is not innocuous at prompt length 8, contributing to looser match.

pith-pipeline@v1.3.0-alltime-deepseek · 26333 in / 12248 out tokens · 111878 ms · 2026-08-01T19:07:48.785342+00:00 · methodology

0 comments
read the original abstract

A Bayes-filtered transformer (BFT) is a transformer trained on sequences that are generated in two steps: first a latent task is drawn from a prior, then observations are drawn conditional on that task. Trained under autoregressive log loss, the BFT's next-token prediction, in the idealized limit, is the Bayesian posterior predictive distribution (PPD) induced by that prior and that conditional law. In practice the trained BFT is only an approximation of this ideal PPD, raising an interpretive question: what prior and posterior over the latent task has the trained BFT actually internalized? Existing work answers this question by comparing the trained BFT's predictions against the predictions of various "reference" posteriors, each standing in for a different candidate algorithm or computation the BFT might be implementing. This prediction-space comparison is fragile: different posteriors can share the same posterior-mean predictions. We use predictive Monte Carlo (PMC) as a general interpretability tool for any BFT: using only next-token generation, PMC returns an approximation to the implicit prior and posterior over the latent task, answering the interpretive question directly in latent space. We apply PMC to three stylized task families spanning 0-Markov and 1-Markov exchangeability. The phenomena previously reported in these settings remain visible in latent space. Code is available at https://github.com/afiq-aswadi/bft-pmc

Figures

Figures reproduced from arXiv: 2607.17060 by Afiq Abdillah Effiezal Aswadi, Haotong Ma, Susan Wei.

Figure 1
Figure 1. Figure 1: PMC schematic. Given a (possibly empty) prompt y1:ℓ, the Beta-Bernoulli BFT au￾toregressively generates R rollouts y (r) ℓ+1:ℓ+N of length N (panel a). For each rollout r, ˆθ (r) = 1 ℓ+N Pℓ i=1 yi + Pℓ+N i=ℓ+1 y (r) i  is the sample mean of the full rollout (including the initial prompt). Panel (b) shows one histogram of { ˆθ (r)} R r=1 obtained with an empty prompt and one obtained with a non-empty promp… view at source ↗
Figure 2
Figure 2. Figure 2: Prediction-space measurement is fragile. 2-D Bayesian linear regression: posterior over the coefficient w ∈ R 2 given a single observation (x1, y1) = ([1, 1], 1), computed under two different priors. The Gaussian and spike-and-slab priors yield posteriors that share the mean (0.5, 0.5) (marked ×) but differ in shape (left, middle). At the test query x ′ = [1, −1], the two predictive distributions over y ′ … view at source ↗
Figure 3
Figure 3. Figure 3: PMC for in-context linear regression. (a) The latent task is the regression coefficient w ∈ R d . Unlike the Beta-Bernoulli setting of [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Linear regression: task-diversity threshold and transient generalization. (a) As task diversity M grows, the PMC posterior moves from the memorizing baseline toward the generalizing baseline. (b) At intermediate task diversity M = 32, the BFT first approximates the generalizing baseline before specializing to the memorizing baseline. Rows: in-distribution prompts (top) and out-of-distribution prompts (bott… view at source ↗
Figure 5
Figure 5. Figure 5: Balls-and-urns: task-diversity threshold and transient generalization. (a) As task diversity M grows, the PMC posterior moves from the memorizing baseline toward the generalizing baseline. (b) At intermediate task diversity M = 32, the BFT first approximates the generalizing baseline before specializing to the memorizing baseline. Rows: in-distribution prompts (top) and out-of-distribution prompts (bottom)… view at source ↗
Figure 6
Figure 6. Figure 6: Markov chains: task-diversity threshold and transient generalization. (a) As task diversity M grows, the PMC posterior moves from the memorizing baseline toward the generalizing baseline. (b) At task diversity M = 8, the BFT first approximates the generalizing baseline before specializing to the memorizing baseline. Rows: in-distribution prompts (top) and out-of-distribution prompts (bottom). Columns: dist… view at source ↗
Figure 7
Figure 7. Figure 7: , shown on its own for the main text [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Autoregressive-PFN attention mask. Even-position tokens embed the covariate xi and odd-position tokens embed the pair (xi , yi). Each token attends to itself (darker cells) and to all earlier odd-position (full-pair) tokens (lighter cells). We do not use a beginning-of-sequence (BOS) token in linear regression. The discrete task family transformers (balls-and-urns and Markov chains) use a BOS token for two… view at source ↗
Figure 9
Figure 9. Figure 9: Linear regression: PMC prior and posterior samples. Marginal densities and CDFs over the first four dimensions of w at task diversity M = 8, against the training prior ΠM, the population prior Π∞, and their corresponding posteriors. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Linear regression: task diversity in the prior and posterior. As task diversity grows, the PMC-recovered prior and posterior move from the memorizing baseline toward the generalizing baseline. The transition is sharper in the posterior. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Linear regression: transient generalization. At intermediate task diversity M = 32, the transformer first approximates the generalizing baseline before specializing to the memorizing baseline. For in-distribution prompts (top) and out-of-distribution prompts (bottom), columns show prediction-space ∆MSE (left), energy distance to each baseline (middle), and sliced Wasserstein distance to each baseline (rig… view at source ↗
Figure 12
Figure 12. Figure 12: Balls-and-urns: PMC prior and posterior samples. Marginal densities and CDFs over the first four classes of p at task diversity M = 8, against the training prior ΠM, the population prior Π∞, and their corresponding posteriors. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Balls-and-urns: task diversity in the prior and posterior, all three metrics. Prediction￾space symmetrized KL gap (left), energy distance (middle), and sliced Wasserstein distance (right) to each baseline, as task diversity varies. The main-text view is [PITH_FULL_IMAGE:figures/full_fig_p030_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Balls-and-urns: transient generalization, all three metrics. At intermediate task diversity M = 32. The main-text view is [PITH_FULL_IMAGE:figures/full_fig_p030_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Markov chains: PMC prior and posterior samples. Each figure is a 4 × 4 matrix of panels mirroring the top-left block of the transition matrix: the panel in row src = a and column dst = b shows the marginal CDF of the transition probability Qa,b from state a to state b, at task diversity M = 4, against the training prior ΠM, the population prior Π∞, and their corresponding posteriors. 33 [PITH_FULL_IMAGE:… view at source ↗
Figure 16
Figure 16. Figure 16: Markov chains: task diversity in the prior and posterior, all three metrics. Prediction￾space symmetrized KL (left), energy distance (middle), and sliced Wasserstein distance (right) to each baseline, as task diversity varies. The main-text view is [PITH_FULL_IMAGE:figures/full_fig_p034_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Markov chains: transient generalization, all three metrics. At task diversity M = 8. The main-text view is [PITH_FULL_IMAGE:figures/full_fig_p034_17.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Elicitation without Backpropagation: Steering Model Behavior by Optimizing the Latent Posterior

    cs.LG 2026-07 conditional novelty 6.0

    Posterior Prefix Tuning steers a Bayes-filtered transformer's latent posterior by optimizing a tilted surrogate with prior samples, eliciting high-utility continuations without backpropagating through the transformer.

Reference graph

Works this paper leans on

41 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    What learning algorithm is in-context learning? investigations with linear models

    Ekin Akyürek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. What learning algorithm is in-context learning? investigations with linear models. InThe Eleventh Inter- national Conference on Learning Representations, 2023. URL https://openreview.net/ forum?id=0g0X4H8yN4I

    2023

  2. [2]

    Bayesian predictive inference beyond martingales,

    Marco Battiston and Lorenzo Cappello. Bayesian predictive inference beyond martingales,

  3. [3]

    Emergent misalignment: Narrow finetuning can produce broadly misaligned LLMs

    Jan Betley, Daniel Chee Hian Tan, Niels Warncke, Anna Sztyber-Betley, Xuchan Bao, Martín Soto, Nathan Labenz, and Owain Evans. Emergent misalignment: Narrow finetuning can produce broadly misaligned LLMs. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste- Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors,Proceedings of t...

    2025

  4. [4]

    GPT- NeoX-20B: An open-source autoregressive language model

    Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, and Samuel Weinbach. GPT- NeoX-20B: An open-source autoregressive language model. InProceedings of the ACL Workshop on Ch...

    Pith/arXiv arXiv 2022

  5. [5]

    Sliced and Radon Wasserstein Barycenters of Measures.Journal of Mathematical Imaging and Vision, 51 (1):22–45, 2015

    Nicolas Bonneel, Julien Rabin, Gabriel Peyré, and Hanspeter Pfister. Sliced and Radon Wasserstein Barycenters of Measures.Journal of Mathematical Imaging and Vision, 51 (1):22–45, 2015. ISSN 1573-7683. doi: 10.1007/s10851-014-0506-3. URL https: //doi.org/10.1007/s10851-014-0506-3

    doi:10.1007/s10851-014-0506-3 2015

  6. [6]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhari- wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agar- wal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Ma- teusz Litwin, S...

    1901

  7. [7]

    Dynamics of Tran- sient Structure in In-Context Linear Regression Transformers, January 2025

    Liam Carroll, Jesse Hoogland, Matthew Farrugia-Roberts, and Daniel Murfet. Dynamics of Tran- sient Structure in In-Context Linear Regression Transformers, January 2025. arXiv:2501.17745

    Pith/arXiv arXiv 2025

  8. [8]

    La prévision : ses lois logiques, ses sources subjectives.Annales de l’Institut Henri Poincaré, 7(1):1–68, 1937

    Bruno de Finetti. La prévision : ses lois logiques, ses sources subjectives.Annales de l’Institut Henri Poincaré, 7(1):1–68, 1937

    1937

  9. [9]

    Language modeling is compression

    Gregoire Deletang, Anian Ruoss, Paul-Ambroise Duquenne, Elliot Catt, Tim Genewein, Christo- pher Mattern, Jordi Grau-Moya, Li Kevin Wenliang, Matthew Aitchison, Laurent Orseau, Mar- cus Hutter, and Joel Veness. Language modeling is compression. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum? id=j...

    2024

  10. [10]

    De Finetti’s theorem for Markov chains.The Annals of Probability, 8(1):115–130, 1980

    Persi Diaconis and David Freedman. De Finetti’s theorem for Markov chains.The Annals of Probability, 8(1):115–130, 1980

    1980

  11. [11]

    Edelman, Eran Malach, and Surbhi Goel

    Ezra Edelman, Nikolaos Tsilivis, Benjamin L. Edelman, Eran Malach, and Surbhi Goel. The evolution of statistical induction heads: In-context learning markov chains. InThe Thirty- eighth Annual Conference on Neural Information Processing Systems, 2024. URL https: //openreview.net/forum?id=qaRT6QTIqJ

    2024

  12. [12]

    Fabian Falck, Ziyu Wang, and Christopher C. Holmes. Is in-context learning in large language models bayesian? a martingale perspective. InForty-first International Conference on Machine Learning, 2024. URLhttps://openreview.net/forum?id=b1YQ5WKY3w

    2024

  13. [13]

    Martingale posterior distributions.Journal of the Royal Statistical Society Series B: Statistical Methodology, 85(5):1357–1391, 2023

    Edwin Fong, Chris Holmes, and Stephen G Walker. Martingale posterior distributions.Journal of the Royal Statistical Society Series B: Statistical Methodology, 85(5):1357–1391, 2023. ISSN 1369-7412, 1467-9868. doi: 10.1093/jrsssb/qkad005

    doi:10.1093/jrsssb/qkad005 2023

  14. [14]

    Sandra Fortini and Sonia Petrone. Quasi-Bayes properties of a procedure for sequential learning in mixture models.Journal of the Royal Statistical Society Series B: Statistical Methodology, 82(4):1087–1114, September 2020. ISSN 1369-7412. doi: 10.1111/rssb.12385. 12

    work page doi:10.1111/rssb.12385 2020

  15. [15]

    Exchangeability, Prediction and Predictive Modeling in Bayesian Statistics.Statistical Science, 40(1), January 2025

    Sandra Fortini and Sonia Petrone. Exchangeability, Prediction and Predictive Modeling in Bayesian Statistics.Statistical Science, 40(1), January 2025. ISSN 0883-4237. doi: 10.1214/ 24-STS965

    2025

  16. [17]

    Uncertainty decomposition for Bayes-filtered transformers via Bayesian predictive inference, 2026

    Sandra Fortini, Kenyon Ng, Sonia Petrone, Judith Rousseau, and Susan Wei. Uncertainty decomposition for Bayes-filtered transformers via Bayesian predictive inference, 2026. URL https://arxiv.org/abs/2602.04596v2. arXiv:2602.04596v2

    Pith/arXiv arXiv 2026

  17. [18]

    What can trans- formers learn in-context? a case study of simple function classes

    Shivam Garg, Dimitris Tsipras, Percy S Liang, and Gregory Valiant. What can trans- formers learn in-context? a case study of simple function classes. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neu- ral Information Processing Systems, volume 35, pages 30583–30598. Curran Associates, Inc., 2022. URL https://pro...

    2022

  18. [19]

    Memory-based meta-learning on non-stationary distributions

    Tim Genewein, Grégoire Delétang, Anian Ruoss, Li Kevin Wenliang, Elliot Catt, Vincent Dutordoir, Jordi Grau-Moya, Laurent Orseau, Marcus Hutter, and Joel Veness. Memory-based meta-learning on non-stationary distributions. InInternational Conference on Machine Learning (ICML), 2023

    2023

  19. [20]

    Learning universal predictors

    Jordi Grau-Moya, Tim Genewein, Marcus Hutter, Laurent Orseau, Gregoire Deletang, Elliot Catt, Anian Ruoss, Li Kevin Wenliang, Christopher Mattern, Matthew Aitchison, and Joel Veness. Learning universal predictors. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors,Proceedin...

    2024

  20. [21]

    TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second

    Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  21. [22]

    Accurate predictions on small data with a tabular foundation model.Nature, 637(8045):319–326, January 2025

    Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tabular foundation model.Nature, 637(8045):319–326, January 2025. ISSN 1476-4687. doi: 10.1038/s41586-024-08328-6

    doi:10.1038/s41586-024-08328-6 2025

  22. [23]

    Springer, Cham, 3 edition, 2021

    Olav Kallenberg.Foundations of Modern Probability, volume 99 ofProbability Theory and Stochastic Modelling. Springer, Cham, 3 edition, 2021

    2021

  23. [24]

    Gener- alized sliced Wasserstein distances

    Soheil Kolouri, Kimia Nadjahi, Umut Simsekli, Roland Badeau, and Gustavo Rohde. Gener- alized sliced Wasserstein distances. InAdvances in Neural Information Processing Systems, volume 32, 2019

    2019

  24. [25]

    Vladimir Mikulik, Grégoire Delétang, Tom McGrath, Tim Genewein, Miljan Martic, Shane Legg, and Pedro A. Ortega. Meta-trained agents implement Bayes-optimal agents. InAdvances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  25. [26]

    Transformers Can Do Bayesian Inference

    Samuel Müller, Noah Hollmann, Sebastian Pineda Arango, Josif Grabocka, and Frank Hut- ter. Transformers Can Do Bayesian Inference. InInternational Conference on Learning Representations (ICLR), 2022

    2022

  26. [27]

    Uncertainty quantification for prior-data fitted networks using martingale posteriors, 2025

    Thomas Nagler and David Rügamer. Uncertainty quantification for prior-data fitted networks using martingale posteriors, 2025. arXiv:2505.11325

    Pith/arXiv arXiv 2025

  27. [28]

    Transformerlens

    Neel Nanda and Joseph Bloom. Transformerlens. https://github.com/ TransformerLensOrg/TransformerLens, 2022

    2022

  28. [29]

    Frazier, Jeremias Knoblauch, and Susan Wei

    Kenyon Ng, Edwin Fong, David T. Frazier, Jeremias Knoblauch, and Susan Wei. TabMGP: Martingale posterior with TabPFN, 2026. arXiv:2510.25154. 13

    Pith/arXiv arXiv 2026

  29. [30]

    In-context learning and induction heads.Transformer Circuits Thread, 2022

    Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Scott Johnston, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dario Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, a...

    2022

  30. [31]

    Ortega, Jane X

    Pedro A. Ortega, Jane X. Wang, Mark Rowland, Tim Genewein, Zeb Kurth-Nelson, Razvan Pascanu, Nicolas Heess, Joel Veness, Alex Pritzel, Pablo Sprechmann, Siddhant M. Jayaku- mar, Tom McGrath, Kevin Miller, Mohammad Azar, Ian Osband, Neil Rabinowitz, András György, Silvia Chiappa, Simon Osindero, Yee Whye Teh, Hado van Hasselt, Nando de Fre- itas, Matthew B...

    Pith/arXiv arXiv 2019

  31. [32]

    In-context learning through the bayesian prism

    Madhur Panwar, Kabir Ahuja, and Navin Goyal. In-context learning through the bayesian prism. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=HX5ujdsSon

    2024

  32. [33]

    Competition Dynamics Shape Algorithmic Phases of In-Context Learning

    Core Francisco Park, Ekdeep Singh Lubana, Itamar Pres, and Hidenori Tanaka. Competition Dynamics Shape Algorithmic Phases of In-Context Learning. InInternational Conference on Learning Representations (ICLR), 2025. Spotlight

    2025

  33. [34]

    Pretraining task diversity and the emergence of non-Bayesian in-context learning for regression

    Allan Raventós, Mansheej Paul, Feng Chen, and Surya Ganguli. Pretraining task diversity and the emergence of non-Bayesian in-context learning for regression. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  34. [35]

    Rizzo and Gábor J

    Maria L. Rizzo and Gábor J. Székely. Energy distance.WIREs Computational Statistics, 8(1): 27–38, January 2016. ISSN 1939-5108, 1939-0068. doi: 10.1002/wics.1375

    doi:10.1002/wics.1375 2016

  35. [36]

    Shai, Sarah E

    Adam S. Shai, Sarah E. Marzen, Lucas Teixeira, Alexander Gietelink Oldenziel, and Paul M. Riechers. Transformers represent belief state geometry in their residual stream. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tom- czak, and C. Zhang, editors,Advances in Neural Information Processing Systems, volume 37, pages 75012–75034. Curran As...

    2024

  36. [37]

    Transformers learn in-context by gradient descent

    Johannes V on Oswald, Eyvind Niklasson, Ettore Randazzo, Joao Sacramento, Alexander Mordvintsev, Andrey Zhmoginov, and Max Vladymyrov. Transformers learn in-context by gradient descent. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors,Proceedings of the 40th International Conference on Mach...

    2023

  37. [38]

    Towards understanding fine-tuning mechanisms of LLMs via circuit analysis

    Xu Wang, Yan Hu, Wenyu Du, Reynold Cheng, Benyou Wang, and Difan Zou. Towards understanding fine-tuning mechanisms of LLMs via circuit analysis. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors,Proceedings of the 42nd International Conference on Machine Learning, volume ...

    2025

  38. [39]

    In-context learning strategies emerge rationally

    Daniel Wurgaft, Ekdeep S Lubana, Core Francisco Park, Hidenori Tanaka, Gautam Reddy, and Noah Goodman. In-context learning strategies emerge rationally. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, editors,Advances in Neu- ral Information Processing Systems, volume 38, pages 118425–118485. Curran Associates, Inc., 20...

    2025

  39. [40]

    An Explanation of In-context Learning as Implicit Bayesian Inference

    Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. An Explanation of In-context Learning as Implicit Bayesian Inference. InInternational Conference on Learning Representations (ICLR), 2022. 14 Appendix contents ANotation summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...

    2022

  40. [42]

    check this condition empirically for TabPFN; the Monte Carlo budgets required are so large that the verification has so far been inconclusive. A subsequent version [17] relaxes this quasi-martingale condition to a weakersignedcondition, requiring only that the signed tail sums of the conditional drift vanish, and reports positive diagnostics on a small Be...

  41. [2025]

    arXiv:2507.21874. 11

    arXiv