Pith. sign in

REVIEW 3 major objections 3 minor 6 cited by

Competition Dynamics Shape Algorithmic Phases of In-Context Learning

T0 review · 3 major / 3 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read Four simple algorithms compete to explain in-context learning, and the winner depends on training and context.

desk verdict A genuinely useful synthetic testbed and a decomposition tool (LIA) that mostly works; the four-algorithm basis is asserted rather than derived, and the acknowledged OOD bias hints at something missing, but the paper is a serious contribution worth engaging. read the letter →

arxiv 2412.01003 v4 pith:AQGRVWJI submitted 2024-12-01 cs.LG cs.CL

classification cs.LGcs.CL
keywords in-contextlearningMarkovchainsalgorithmicphasestransformerlinearinterpolationofalgorithmsout-of-distributiongeneralizationcompetitiondynamicstransientICL
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that in-context learning (ICL) should not be treated as a single capability but as a competition between a small set of simple algorithms, and it builds a unified toy setting to show this: training a transformer to autoregressively simulate a finite mixture of Markov chains. In that setting, the model's next-token distribution is almost exactly a nonnegative linear combination of four algorithms, which pair fuzzy retrieval versus direct inference with unigram versus bigram statistics of the context. Which algorithm dominates changes with training steps, data diversity, and context size, sometimes sharply, producing an algorithmic phase diagram. This competition explains known ICL phenomena, most notably the transient nature of ICL: a retrieval algorithm that is better on the training distribution slowly replaces a better-generalizing inference algorithm, degrading out-of-distribution performance. A sympathetic reader would take the paper's central claim to be that ICL phenomena are phase-dependent outcomes of this competition rather than universal properties of a monolithic mechanism.

What carries the argument

The load-bearing object is the Linear Interpolation of Algorithms (LIA), the constrained regression in Eq. 7 that expresses the model's next-token distribution as a convex combination of four hand-defined algorithmic distributions. The four algorithms are: Uni-Ret (unigram retrieval), which weights training-set chains by how well their stationary distributions match the context's token histogram; Bi-Ret (bigram retrieval), which weights by observed transitions; Uni-Inf (unigram inference), which predicts next tokens from the context's token frequencies alone; and Bi-Inf (bigram inference), which estimates the transition matrix from observed transitions in the context. LIA's fitted weights, together with two probes (shuffling the context to test bigram use, and comparing predictions' proximity to seen versus random transition matrices to test retrieval), produce the paper's phase diagram and explain the out-of-distribution generalization dynamics.

What would settle it

Train the same transformer on a mixture of order-2 Markov chains, where the next state depends on the two previous states. If the four-algorithm LIA fits with the same near-zero residual as in the paper, the decomposition is insensitive to a genuinely different algorithm; if the residual is large or a fifth algorithm is needed, the completeness of the basis fails.

Watch

Extended reading notes

Core claim

The central discovery is a decomposition: for transformers trained on finite Markov mixtures, the trained model's predicted next-token distribution is well approximated by $p_{\text{model}} \approx \sum_{a} w_a p_a$, where $a$ ranges over Uni-Ret, Bi-Ret, Uni-Inf, and Bi-Inf, and the weights are nonnegative and sum to one (Eq. 7). The weights fitted on in-distribution sequences predict the model's out-of-distribution KL, including its non-monotonic rise and fall during training. The paper reads this as evidence that the four algorithms compete for control of the model, with experimental conditions determining the winner; for example, at moderate data diversity the model first adopts bigram inference (Bi-Inf), which generalizes well, and then is slowly overtaken by bigram retrieval (Bi-Ret), which is better on the training set but worse out of distribution, explaining the transient nature of ICL.

Load-bearing premise

The four hand-defined algorithms are assumed to be a complete basis for whatever the model actually computes, so that every next-token distribution lies approximately in their nonnegative span.

Editorial extensions

If this is right

  • The transient nature of ICL follows directly: a training-optimal retrieval algorithm gradually replaces the OOD-optimal inference algorithm, so performance on novel distributions peaks and then decays.
  • Data diversity acts as a phase boundary: below a threshold the model stays in retrieval phases, while above it bigram inference can emerge and generalize out of distribution.
  • In-distribution measurements alone, namely the LIA weights, can predict out-of-distribution generalization trends during training.
  • Model design choices such as width, MLP size, state-space size, and tokenization shift the phase boundaries, so scaling conclusions depend on which algorithm is being promoted.
  • ICL findings are setting-dependent; no single mechanism explains behavior across all configurations.

Reading between the lines

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

  • The paper's own controls show that LIA assigns flat weights to arbitrary algorithms, but completeness of the four-algorithm basis remains open; a natural test is to fit LIA on a model trained on an order-2 Markov mixture, where a genuinely different algorithm is known to exist.
  • If the competition picture carries over to real LLMs, training interventions such as data composition, architecture, and tokenization could aim to favor the generalizing algorithm rather than minimizing loss alone, and phase diagrams might indicate when continued training becomes harmful.
  • The paper notes that LIA's predicted OOD KL is always slightly worse than the model's actual performance; understanding that residual could reveal whether the circuits are truly competitive or partially merged.
  • The real-LLM appendix suggests these algorithms may be identifiable in production models, where prompt wording acts as an experimental condition that shifts behavior across algorithmic regimes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper introduces a synthetic sequence modeling task—learning to simulate a finite mixture of Markov chains—and shows that Transformers trained on this task reproduce several well-known ICL phenomena, including the data diversity threshold, induction head emergence, transient ICL, task retrieval vs. task learning, early ascent, and bounded efficacy. It then defines four interpretable algorithms spanning a unigram/bigram axis and a retrieval/inference axis, and demonstrates via a nonnegative linear interpolation (LIA) that a trained model's next-token distribution is well approximated by a convex combination of these four algorithms on in-distribution sequences. The LIA weights fit on ID data are used to predict OOD performance, which the authors use to explain the transient nature of ICL. Additional experiments probe how model width, tokenization, and data complexity shift the algorithmic phase diagram.

Significance. If the central claims hold, this is a valuable contribution: it provides a unified synthetic testbed that reproduces a broad ICL phenomenology, and it proposes a concrete, interpretable decomposition tool (LIA) with a strong in-distribution fit. The paper is commendable for the breadth of reproduced phenomena, the direct control experiment with arbitrary algorithms (App. I.2), the preliminary mechanistic evidence from attention maps and MLP reconstructions (App. E), and the public code release. The OOD prediction using ID-fitted weights is a genuine, non-circular test. However, the load-bearing assumption that the four hand-defined algorithms form a complete basis for the model's behavior is not established, and the paper's own appendix reports a systematic OOD gap that is the expected signature of a missing component. This undermines the quantitative precision of the competition-dynamics explanation until addressed.

major comments (3)
  1. [App. K / Sec. 4.2, Eq. (13)] The claim that LIA weights fit on ID data explain OOD performance is load-bearing for the transient-ICL explanation, but the paper's own App. K states that the LIA-predicted OOD KL is systematically higher (worse) than the model's actual OOD performance. As the skeptic notes, this is exactly the signature of a missing algorithm or of context-dependent weights not captured by the fixed four-algorithm basis. The authors should either extend the set with a fifth candidate (e.g., a trigram-inference algorithm or a context-dependent switch between Bi-Inf and Bi-Ret) and show that the gap closes, or quantify the gap relative to the model's OOD KL and show it is within statistical noise. Without this, the quantitative OOD predictions in Fig. 7 do not provide evidence for completeness of the decomposition.
  2. [Sec. 3.2 / App. I.2] The phase identification protocol distinguishes unigram vs. bigram and retrieval vs. inference, but the shuffle perturbation only destroys first-order transition structure and cannot rule out trigram or higher-order strategies. The arbitrary-algorithm control in App. I.2 shows that the four proposed algorithms receive higher weight than random ones, but this is a discrimination result, not a completeness result; it does not exclude a fifth meaningful algorithm that lies outside the tested set. The paper should state this limitation explicitly and, ideally, test a trigram or other higher-order baseline in the LIA optimization to demonstrate that the four-algorithm basis is not an artifact of the chosen dictionary.
  3. [Sec. 4.1, Fig. 6-7 and App. A.3.4] The LIA weights are optimized on 300 ID chains and evaluation averages are taken over 30 transition matrices, but the reported curves in Fig. 6 and Fig. 7—and the phase diagrams in Fig. 5 and Fig. 8—do not show error bars or confidence intervals. This makes it difficult to assess whether the sharp transitions and phase boundaries are statistically robust. The authors should add error bars (e.g., via multiple seeds or bootstrap over evaluation matrices) at least for the LIA weight trajectories and the OOD prediction curves, particularly at phase boundaries where the linear fit is reported to be slightly worse (App. I.1).
minor comments (3)
  1. [App. C.1, item 1] There is a duplicated sentence fragment: "We reproduce this data diversity threshold in We reproduce this phenomena as seen in Fig. 1 (a), Fig. 3 (a,c) and Fig. 27 (a)." This needs to be edited.
  2. [App. C.2.3] The same section heading, "Effects of equi-FLOPs training while varying model size," appears twice in succession; the second occurrence should be reworded to reflect the actual content (e.g., "Effects of equi-FLOPs training across model sizes").
  3. [Fig. 3 caption] The caption refers to a "ruby vertical dashed line," but color terms may not be robust in grayscale or for color-blind readers; consider using a label or a distinct line style instead.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: LIA weights fitted on ID data are applied to independently computed OOD algorithm probabilities, and the paper explicitly reports the residual bias of its own decomposition.

full rationale

The paper's derivation chain is not circular in the load-bearing sense. The four algorithms are defined as closed-form Bayesian/empirical estimators (Eqs. 2-6), and the phase probes in Sec. 3.2 operationalize the same unigram/bigram and retrieval/inference axes; this is a diagnostic classification of model behavior, not a derivation of the algorithms from the phenomena they are claimed to explain. The LIA weights in Eq. 7 are fitted to in-distribution model next-token probabilities, and the OOD predictions in Sec. 4.2 use those same weights with algorithm probabilities evaluated on OOD contexts (Eq. 13). Since the model's OOD KL is not part of the fitting objective, the transient-ICL explanation is an extrapolation rather than a fitted artifact. Appendix I.2's control with arbitrary algorithms shows the fitted weights discriminate against irrelevant dictionary elements, which supports the meaningfulness of the chosen basis, though it does not prove completeness against an omitted fifth strategy; that is an empirical-modeling limitation, not circularity. Appendix K explicitly acknowledges that the predicted OOD KL is systematically slightly worse than the model's actual OOD KL, and interprets it as an imperfect decomposition rather than hiding it. The only self-citation (Bigelow et al., 2023, involving two of the present authors) appears in the intro as one of several probabilistic-automata references and is not load-bearing for any central claim. No equation-level reduction of a predicted result to a fitted input or to a self-citation chain is exhibited, so the appropriate finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the assumed completeness of the four-algorithm basis and on the soundness of the two behavioral probes. The LIA weights are fitted quantities, and the smoothing constant in Bi-Inf is chosen by hand. No new physical entities are introduced.

free parameters (2)
  • LIA mixing weights (Uni-Ret, Bi-Ret, Uni-Inf, Bi-Inf) = varies with N and step; see Fig. 6
    In Eq. 7, weights are optimized to fit the model's next-token probabilities on ID sequences; they are fitted parameters, not derived from first principles.
  • Laplace smoothing constant in Bi-Inf (Eq. 6) = 1
    The +1 in the numerator and denominator of Eq. 6 sets a uniform prior over transitions; it is chosen by hand, not fitted or derived.
assumptions (5)
  • domain assumption Sequences are generated by a finite mixture of Markov chains with rows sampled from Dirichlet(1/k).
    Defines the task; any conclusion about ICL is restricted to this DGP.
  • domain assumption The next state depends only on the current state (Markov property).
    This structure is what makes bigram statistics sufficient; real language is not strictly Markovian.
  • ad hoc to paper The four algorithms (Uni-Ret, Bi-Ret, Uni-Inf, Bi-Inf) span the model's behavior, so a nonnegative linear combination captures p_model.
    LIA in Eq. 7 assumes completeness; App. I.2 provides a control but does not prove completeness.
  • standard math Shuffling the context preserves unigram statistics while destroying bigram statistics.
    Used in Eq. 9 to measure bigram utilization; relies on exchangeability of counts.
  • domain assumption The retrieval-proximity test (Eq. 11) correctly distinguishes retrieval from inference solutions.
    Assumes KL to nearest training chain vs. random chain is a sound measure; this is plausible but not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Competition Dynamics Shape Algorithmic Phases of In-Context Learning." pith.science (2026). https://pith.science/paper/AQGRVWJI

@misc{pith2026241201003,
  author       = {Pith},
  title        = {Pith review of: Competition Dynamics Shape Algorithmic Phases of In-Context Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AQGRVWJI}},
  note         = {Machine review of arXiv:2412.01003}
}
read the original abstract

In-Context Learning (ICL) has significantly expanded the general-purpose nature of large language models, allowing them to adapt to novel tasks using merely the inputted context. This has motivated a series of papers that analyze tractable synthetic domains and postulate precise mechanisms that may underlie ICL. However, the use of relatively distinct setups that often lack a sequence modeling nature to them makes it unclear how general the reported insights from such studies are. Motivated by this, we propose a synthetic sequence modeling task that involves learning to simulate a finite mixture of Markov chains. As we show, models trained on this task reproduce most well-known results on ICL, hence offering a unified setting for studying the concept. Building on this setup, we demonstrate we can explain a model's behavior by decomposing it into four broad algorithms that combine a fuzzy retrieval vs. inference approach with either unigram or bigram statistics of the context. These algorithms engage in a competition dynamics to dominate model behavior, with the precise experimental conditions dictating which algorithm ends up superseding others: e.g., we find merely varying context size or amount of training yields (at times sharp) transitions between which algorithm dictates the model behavior, revealing a mechanism that explains the transient nature of ICL. In this sense, we argue ICL is best thought of as a mixture of different algorithms, each with its own peculiarities, instead of a monolithic capability. This also implies that making general claims about ICL that hold universally across all settings may be infeasible.

Figures

Figures reproduced from arXiv: 2412.01003 by the authors.

Figure 1
Figure 1. Algorithmic phase diagram for a finite Markov mixtures task. We propose to study ICL phenomena through a minimal experimental system: Transformers trained on sequence data generated by a finite mixture of Markov chains. This setup turns out to be extremely rich, capturing most (if not all) known phenomenology of ICL, but still being simple enough to be amenable to theoretical modeling. We identify four distinct, int… view at source ↗
Figure 2
Figure 2. Data generation and evaluation protocol with finite Markov mixtures. (a) Data gen￾eration. We first sample a finite set Ttrain = {T1, T2, . . . , TN } of random transition matrices to define our set of Markov chains. We then randomly select a chain from this set and sample a train￾ing sequence from it. We repeat this process at every step of training, sampling a fresh batch of sequences from by randomly selecting a … view at source ↗
Figure 3
Figure 3. Finite Markov mixture setup captures rich phenomenology of in-context learning (ICL). (a) KL divergence (OOD evaluation) as a function of training steps and data diversity (Num￾ber of Training Chains). (b) As the data diversity of the training data is increased (see ruby vertical dashed line in panel (a)), we reproduce the data diversity threshold for “task learning” ICL, similar to Raventos et al. ´ (2023); Kirsch … view at source ↗
Figures from the paper (39 more)
Figure 4
Figure 4. Figure 4: Proposed algorithms for the finite Markov mixture task. (a) Unigram based Retrieval (Uni-Ret): Given a sequence, Uni-Ret involves computing a histogram of token frequencies in the sequence and then creating a new transition matrix that is a weighted average of chains i…
Figure 5
Figure 5. Figure 5: Algorithmic phases. (a) Bigram Utilization: We shuffle the order of all states in a sequence and measure the KL (App. Eq. 9) before and after the perturbation to quantify the bigram utilization of a model. The shuffling should only affect algorithms sensitive to higher…
Figure 6
Figure 6. Figure 6: Linear Interpolation of Algorithms (LIA) (a) Algorithmic Phases extracted by LIA. We color each location by combining individual colors associated to an algorithm, weighted by their weights wa. (b) LIA weights across checkpoints extracted using ID data for N = 26 . KL …
Figure 7
Figure 7. Figure 7: Algorithmic phase transitions drive non-monotonic OOD performance. Note that each algorithm’s KL in colored dotted lines has changed from [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Algorithmic Phase Diagrams Explain the effects model size, data complexity, and tokenization. (a) Phase diagram for a wider model with an embedding dimension of 256. Here, we find that the data diversity required to observe Bi-Inf is increased to 2 10. (b) A decreased …
Figure 9
Figure 9. Figure 9: shows the learning curves (loss) and ID / OOD KL divergence through training. The data is the same as in [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Explaining the Transient Nature of In-Context Learning Our setup and LIA allows a clear understanding of the transient nature of In-Context Learning. a,b) are panels from Singh et al. (2023). a) shows that the training loss slowly decreases after a initial drop. b) sh…
Figure 11
Figure 11. Figure 11: However, on further investigation, we found evidence that our results are confounded by the faster learning speed of bigger models (Kaplan et al., 2020; Hoffmann et al., 2022; Bordelon et al., 2024). We thus properly normalized the training by FLOPs, as shown in the n…
Figure 12
Figure 12. Figure 12: Algorithmic Phase diagrams depending on model width We show algorithmic phase diagrams as we increase the model’s embedding dimension from 32 to 256. Larger models seem to enhance the memorization solutions. N Training Chains Model Width Training Uni-Ret Bi-Ret Bi-Inf…
Figure 13
Figure 13. Figure 13: Algorithmic Phases spanning data diversity and model scale. We see the algorith￾mic phases of models depend on data diversity and their width. We find that smaller models form Bi-Inf solution more efficiently, when properly normalizing for FLOPs. 27 [PITH_FULL_IMAGE:…
Figure 14
Figure 14. Figure 14: High Dimensional Distances We show that the distance of a point to the nearest neighbor within a set of points increases faster than the distance to the distributional mean as the dimension increases. (a) Schematic of the distances. In 2 dimensions, it is much more in…
Figure 15
Figure 15. Figure 15: Reconstructing Markov Chains from Neuron Weights. We show that transitions from Markov chains seen in training (left column) can be directly reconstructed from neuron weights (right column). As expected, results are especially good in low data diversity settings (top …
Figure 16
Figure 16. Figure 16: Memorization as Training Progresses. Minimum KL of transitions from training and neuron outputs, averaged across 100 randomly selected transitions. a) N = 22 . b) N = 26 . c) N = 211. Lower data diversity settings display much greater degree of similarity with trainin…
Figure 17
Figure 17. Figure 17: Neural Activity Across Increasing Data Diversities. a) Sorted GeLU activations of neurons, normalized by their maximum activation, are shown for increasing levels of data diversity. In the Bi-Ret regime, significantly fewer neurons are active compared to the Bi-Inf re…
Figure 18
Figure 18. Figure 18: First layer attention maps from the N = 26 run. We visualize the first layer attention maps from different points in the checkpoint. Each query is normalized by the maximal value present in the row for visualization purposes. The title shows the number of training ste…
Figure 19
Figure 19. Figure 19 [PITH_FULL_IMAGE:figures/full_fig_p036_19.png]
Figure 20
Figure 20. Figure 20: Attention Maps At Key Checkpoints Across Training For N = 22 . a) Training checkpoint with maximal likelihood of Uni-Inf according to LIA. b) Training checkpoint with maximal likelihood of Uni-Ret according to LIA. c) Training checkpoint with maximal likelihood of Bi-…
Figure 21
Figure 21. Figure 21: Attention Maps At Key Checkpoints Across Training For N = 26 . a) Training checkpoint with maximal likelihood of Uni-Inf according to LIA. b) Training checkpoint with maximal likelihood of Bi-Inf according to LIA. c) Training checkpoint with maximal likelihood of Bi-R…
Figure 22
Figure 22. Figure 22: Attention Maps At Key Checkpoints Across Training For N = 211 . a) Training checkpoint with maximal likelihood of Uni-Inf according to LIA. b) Training checkpoint with maximal likelihood of Bi-Inf according to LIA. A red × indicates the position after previous occurre…
Figure 23
Figure 23. Figure 23: Attention Maps At Predicted Transition Points For N = 22 . a) First predicted transition checkpoint according to LIA. b) Second predicted transition checkpoint according to LIA. (a) (b) Layer 1 Layer 2 Algorithm Probabilities Key Index Key Index Training Step Query In…
Figure 24
Figure 24. Figure 24: Attention Maps At Predicted Transition Points For N = 26 . a) First predicted transition checkpoint according to LIA. b) Second predicted transition checkpoint according to LIA. 39 [PITH_FULL_IMAGE:figures/full_fig_p039_24.png]
Figure 25
Figure 25. Figure 25: Attention Maps At Predicted Transition Points Across Training For N = 211 . (b) (c) (a) [PITH_FULL_IMAGE:figures/full_fig_p040_25.png]
Figure 26
Figure 26. Figure 26: Attention Interpolation. Manually interpolating between attention patterns correspond￾ing to different algorithms closely matches the attention patterns observed during training when the model is predicted to be transitioning between these algorithms. This supports th…
Figure 27
Figure 27. Figure 27: Subplots of [PITH_FULL_IMAGE:figures/full_fig_p041_27.png]
Figure 28
Figure 28. Figure 28: 3D heatmap of ID/OOD KL divergence. (left) KL divergence on ID sequences. (right) KL divergence on OOD sequences. In [PITH_FULL_IMAGE:figures/full_fig_p042_28.png]
Figure 29
Figure 29. Figure 29: KL divergence depending on data diversity and optimization. a) ID KL divergence. b) OOD KL divergence. c) Excess KL divergence on OOD sequences. kl_heatmap_cs1.pdf Context Size Context Size Context Size Training Steps Training Steps Training Steps a) ID b) OOD c) KL D…
Figure 30
Figure 30. Figure 30: KL divergence depending on context size and optimization steps for N = 24 . a) ID KL divergence. b) OOD KL divergence. c) Excess KL divergence on OOD sequences. that there is a much sharper evolution near step ∼ 6×102 [PITH_FULL_IMAGE:figures/full_fig_p043_30.png]
Figure 31
Figure 31. Figure 31: KL divergence depending on context size and optimization steps for N = 26 . a) ID KL divergence. b) OOD KL divergence. c) Excess KL divergence on OOD sequences. F.1.3 KL DIVERGENCE VS. {STEPS, DIVERSITY, CONTEXT} Recall that we showed model’s ID and OOD KL divergence …
Figure 32
Figure 32. Figure 32: KL vs. N at fixed step. We show the effect of data diversity at fixed number of gradient steps. We plot the model’s KL divergence with respect to ground truth for both ID and OOD chains in black. We also plot the ID and OOD KL for each of the 4 solutions. Note that wh…
Figure 33
Figure 33. Figure 33: KL vs. Steps at fixed N. We show the effect of optimization at each fixed data diversity, N. We plot the model’s KL divergence with respect to ground truth for both ID and OOD chains in black. We also plot the ID and OOD KL for each of the 4 solutions in horizontal co…
Figure 34
Figure 34. Figure 34: KL vs. context size at fixed N and step. We show the effect of context size at fixed data diversity N and gradient steps. We plot the model’s KL divergence with respect to ground truth for both ID and OOD chains. We also plot the ID and OOD KL for each of the 4 soluti…
Figure 35
Figure 35. Figure 35: KL divergence between algorithmic solutions and the model. We quantify the KL divergence of each algorithmic solution from the model estimate of the transition matrix, across optimization steps and data diversity. We find nearly identical low KL regions as in [PITH_F…
Figure 36
Figure 36. Figure 36: Algorithmic Phase diagrams for different architectures. Results with a Transformer using a) learned positional embeddings; b) a single attention head; c) using 16 attention heads; d) using an MLP fan-out ratio of 0.25; and e) using an MLP fan-out ratio of 16 [PITH_FU…
Figure 37
Figure 37. Figure 37: illustrates these results. We fed in a context consisting of a repeating pattern of 0, 1, 2 and we generate a sequence from the model. We generate the sequence with zero temperature, i.e., select the state with the highest predicted probability. We visualize the next …
Figure 38
Figure 38. Figure 38: illustrates the results of inputting a sequence from a Markov chain as tokens of digits from 0 to 9. More precisely, we drew a transition matrix T ∗ from the DGP in App. A.1, and generated 300 sequences to collect next token probabilities to construct Tˆ as in App. A.…
Figure 39
Figure 39. Figure 39: Residual L2 of LIA fit. We show the residual probability space L2 (the argument of arg min in Eq. 7) for the LIA fit shown in [PITH_FULL_IMAGE:figures/full_fig_p050_39.png]
Figure 40
Figure 40. Figure 40: KL divergence between the model and LIA fit. (Left) KL divergence between model predictions and LIA across data diversity and optimization. (Right) KL divergence between LIA and model predictions across data diversity and optimization. L2 distance between model and LI…
Figure 41
Figure 41. Figure 41: Applying LIA with 4 arbitrary solutions LIA is optimized with 4 arbitrarily chosen solutions (a) weight for the first solution, which outputs a next state probability using a frozen transition matrix drawn from the Dirichlet prior. (b) weight for the second solution, …
Figure 42
Figure 42. Figure 42: Applying LIA with the 4 algorithmic solutions and 4 arbitrary solutions (a) weight for Uni-Ret. (b) weight for Bi-Ret. (c) weight for Uni-Inf. (d) weight for Bi-Inf. (e, f, g, h) weights corresponding to, respectively, solutions (a, b, c, d) in [PITH_FULL_IMAGE:figur…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

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

  1. Mechanistic Foundations of Goal-Directed Control

    cs.LG 2026-03 conditional novelty 7.0 of 10

    Context window k is the critical parameter for arbitration-gate formation in an embodied control architecture: no circuit below k≤4, resolved phase structure above k≥8, with EMA-like commitment dynamics.

  2. ICLR: In-Context Learning of Representations

    cs.CL 2024-12 conditional novelty 7.0 of 10

    As in-context examples grow, Llama-3.1-8B reorganizes its concept representations to mirror the connectivity structure of a graph defined entirely in context.

  3. Sequential Correlations Change In-Context Learning: Effective Context Length and Architectural Mismatch

    stat.ML 2026-07 conditional novelty 6.0 of 10

    Within-context token correlations reduce ICL to an effective shorter i.i.d. context length, while query–context correlations lower error and favor softmax over linear attention.

  4. Distinct Computations Emerge From Compositional Curricula in In-Context Learning

    cs.LG 2025-06 conditional novelty 6.0 of 10

    When transformer models see easy component examples before a harder combined math problem in one prompt, they solve unseen versions of the combined problem and store intermediate steps internally, unlike models traine...

  5. Decomposing Elements of Problem Solving: What "Math" Does RL Teach?

    cs.AI 2025-05 conditional novelty 6.0 of 10

    Reinforcement learning (GRPO) on math LLMs primarily increases execution robustness on already-solvable problems, not planning or coverage of new problems.

  6. The broader spectrum of in-context learning

    cs.CL 2024-12 conditional novelty 6.0 of 10

    The authors propose a broad definition of in-context learning as any sequence task where context reduces loss, and argue research should study this wider spectrum.

Reference graph

Works this paper leans on

20 extracted references · 17 canonical work pages · cited by 6 Pith papers

  1. [1]

    k: The number of states in the Markov chain, fixed to 10 unless mentioned otherwise

  2. [2]

    l: The length of the sequence generated, fixed to 512 unless mentioned otherwise

  3. [3]

    Bayesian

    URL https://arxiv.org/abs/2405.15618. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth ´ee Lacroix, Baptiste Rozi `ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkor...

  4. [4]

    task retrieval

    Task Retrieval to Task Learning transition: Researchers have classified the operation of ICL as either “task retrieval” or “task learning” (Min et al., 2022; Pan et al., 2023). To this end, such papers explore effects of, e.g., shuffling next-token predictions in few-shot tasks and finding that one can still achieve almost similar performance as the scena...

  5. [5]

    This phenomena is empirically observed in Xie et al

    Early Ascent: Early ascent describes the phenomena where the error/risk on an ICL task initially increases before decreasing. This phenomena is empirically observed in Xie et al. (2021) while Lin & Lee (2024) suggests an explanation with the linear regression setting. In this work, we show that this phenomena can be reproduced when training with an interm...

  6. [6]

    ,211} for the main experiments

    N: The number of training matrices, N ∈ {22, 23, . . . ,211} for the main experiments. The set of Markov chains used for training is constructed by drawingN transition matrices. We draw each row of each transition matrix from a Dirichlet distribution with parameter α = 1k. We index individual transition matrices via a subscript, i.e., Tn and use square br...

  7. [8]

    non-Bayesian

    Emergence of Non-Bayesian ICL with data diversity : Prior work exploring linear regression tasks shows that ICL performance drastically improves on OOD data with increase in data di- versity Ravent´os et al. (2023); Kirsch et al. (2022); Lu et al. (2024). In specific, Ravent ´os et al. (2023) shows that there is a threshold needed for a “non-Bayesian” ICL...

  8. [9]

    Optimization Limited Emergence

    Formation of induction heads and variants : Induction heads are a specialized attention head that help infer next-token predictions in a context-conditioned manner Elhage et al. (2021); Ols- son et al. (2022); Reddy (2023); Edelman et al. (2024); Aky ¨urek et al. (2024). Often, there is a sudden loss drop that correlates with induction head formation, and...

Show all 20 references
  1. [10]

    (2023); Anand et al

    Transient Nature of In-Context Learning : Recently researchers have found that ICL can be transient during pre-training Singh et al. (2023); Anand et al. (2024); Panwar et al. (2024). In this work we show that this happens when a solution performing better on the training set ...

  2. [13]

    General Purpose ICL

    Bounded Efficacy: Bounded efficacy of biased label ICL is observed empirically in Min et al. (2022) and coined as a term in Lin & Lee (2024). In this work, we show that this happens when the model’s algorithm is a superposition of a retrieval solution and an inference solution...

  3. [14]

    Random model’s ability to reconstruct a transition matrix: We take a randomly initialized model and report the KL it achieves when trying to reconstruct a transition matrix

  4. [15]

    inference

    Trained model’s ability to reconstruct an unseen transition matrix: We take a trained model, sample an unseen transition matrix, and analyze whether we can reconstruct such an unseen 3Performing this analysis over all seen matrices can be prohibitively expensive, requiring ana...

  5. [16]

    Panel (a): We observe a very robustBi-Ret solution for N = 22, a very low data diversity

  6. [17]

    Panel (c): We observe a highly non-monotonic OOD KL divergence forN = 25, a medium data diversity

  7. [18]

    Specifically, the model first finds the Bi-Inf solution and then moves to a retrieval solu- tion which harms the OOD KL

    Panels (d, e): We observe at high data diversity N = 2 6, 28 the transient nature of ICL. Specifically, the model first finds the Bi-Inf solution and then moves to a retrieval solu- tion which harms the OOD KL

  8. [19]

    memorizing

    Panel (f): At a very high data diversity ( N = 211), we observe that the Bi-Ret solution does not show up (at least not noticeably) within the compute budget. We show both the ID and OOD KL divergence of the model and different solutions at fixedN, step, and across context len...

  9. [25]

    Another question with a slightly different implication is whether the circuit supporting the solution did in fact already emerge, yet is not used

    One natural question is whether it could emerge with data diversity under 26, but is just not ob- served due to Uni-Ret and Bi-Ret being found first. Another question with a slightly different implication is whether the circuit supporting the solution did in fact already emerg...

  10. [2021]

    Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al

    https://transformer-circuits.pub/2021/framework/index.html. Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposi- tion. arXiv preprint arXiv:220...

  11. [2023]

    BAYESIAN

    instead of the default learning positional embedding, which significantly delayed the emer- gence of Bi-Inf. Further discussion about these results are in App. F.2. All matrix weights are ini- tialized as N (0, 0.02) except residual projections which are initialized as N (0, 0...

  12. [2024]

    URL https://arxiv.org/abs/2306.04891. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K ¨opf, Ed- ward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, S...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.