Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read In-context meta-learning emerges through three distinct attention circuits, not one induction-head jump.

desk verdict Clean, well-controlled study of staged circuit formation in task-inference ICL; worth refereeing, though the three-phase partition rests on one unvaried threshold. read the letter →

arxiv 2505.16694 v2 pith:ND7ML7JQ submitted 2025-05-22 cs.CL cs.AI

Gouki Minegishi , Hiroki Furuta , Shohei Taniguchi , Yusuke Iwasawa , Yutaka Matsuo This is my paper
classification cs.CLcs.AI
keywords in-contextlearningmeta-learningmechanisticinterpretabilityattentioncircuitsphasesinductionheadslabelcircuitemergence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that when a transformer is trained to infer a task from few-shot examples, rather than to copy an answer already present in the context, it does not acquire in-context learning in a single abrupt transition. Instead, the model passes through three discrete phases, each implementing a distinct attention circuit: a bigram-only circuit that ignores context, a label-attention circuit that reads only the labels, and finally a full-context circuit that chunks each example pair and uses the inferred task to answer. The authors argue this staged emergence is the mechanistic counterpart of task inference, the part of in-context learning that induction heads cannot explain. If correct, the result reframes circuit emergence as a multi-phase developmental process and connects toy transformer training to phenomena like random-label robustness and smooth loss curves in large language models.

What carries the argument

The load-bearing objects are three attention metrics computed from the causal attention maps of each layer: Bigram (attention from the query token to itself), Label Attention (attention from the query to label tokens in the context), and Chunk Example (attention from each x to its paired ℓ). The paper also uses a $\Delta$-Accuracy measure, Acc(t+Δ t)−Acc(t) with Δ t=100, and a fixed threshold of 0.025 to mark phase boundaries. The controlled ICML task, with random per-task label assignments and the query absent from the context, forces task inference so that any accuracy gain beyond memorization must come from context. The metrics do the work of making otherwise invisible attention patterns measurable and testable during training.

What would settle it

Systematically sweep the ΔAccuracy threshold used to mark phases (for example, 0.01, 0.015, 0.02, 0.03, 0.04, 0.05) on the same runs shown in Figure 3; if the number or location of phases changes with the threshold, the three-phase claim is a threshold artifact rather than a property of the circuits themselves.

Watch

Extended reading notes

Core claim

This paper claims that in its proposed In-Context Meta-Learning (ICML) setting, a two-layer attention-only transformer acquires task-inference in-context learning through three ordered circuit phases, and it gives names, attention patterns, and metrics for each. Phase 1 (Non-Context Circuit, NCC) uses bigram attention in both layers, relying on memorized weights to answer the query without reading the context, which caps accuracy near 1/T. Phase 2 (Semi-Context Circuit, SCC) adds label attention in layer 1 while keeping bigram attention in layer 2, reaching roughly 75% accuracy because a label seen in the context can rule out one hypothesis. Phase 3 (Full-Context Circuit, FCC) chunks each (x, ℓ) pair into one token in layer 1 and attends to labels in layer 2, reaching 100% accuracy by abstracting the task. Using three attention metrics, the paper shows that circuit changes align with abrupt accuracy jumps; pruning experiments show each circuit alone reproduces its phase's accuracy; and in multi-head models the phases are hidden in the accuracy curve even though circuit metrics still change abruptly, with heads specializing in parallel. The same layer-wise pattern, chunk-example attention early and label attention later, appears in GPT2-XL on a sentiment classification prompt.

Load-bearing premise

The three-phase split rests on treating an accuracy jump of more than 0.025 over 100 training steps as the boundary between phases, and a different threshold could change when, or whether, the phases appear.

Editorial extensions

If this is right

  • If the three-phase picture is correct, induction heads are only one endpoint of a developmental sequence, and task-inference in-context learning should be analyzed with circuits that read labels and chunk example pairs, not just match-and-copy.
  • The SCC's label-attention mechanism provides a circuit-level explanation for why LLMs keep high in-context learning accuracy under randomized labels: the model can exploit task-relevant label statistics before it has learned any (x, ℓ) correspondence.
  • Because pruning each phase's circuit reproduces that phase's accuracy, circuit identity is causal for performance at each plateau, so interventions on attention can steer which phase the model lands in.
  • Multi-head models can hide phase transitions: smooth accuracy curves do not imply smooth circuit formation, so circuit metrics may reveal abrupt internal changes even when loss decreases smoothly.
  • Data statistics control phase skipping: few tasks, many classes, high within-class noise, or Zipfian sampling can suppress early phases, giving a testable handle on when task inference emerges.

Reading between the lines

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

  • Beyond the paper: if phase transitions are discrete at the circuit level, then curve-smoothing choices (larger batch, more heads, averaged metrics) may hide rather than remove them; monitoring the three attention metrics during training could serve as an early-warning signal for capability jumps in larger models.
  • Beyond the paper: the ordering NCC → SCC → FCC suggests a curriculum prediction—training on data that makes labels task-informative early should bring the SCC phase forward, while removing label-set statistics should delay or eliminate it.
  • Beyond the paper: the SCC's reliance on labels alone predicts that a label-set-only probe, feeding the model only the unordered labels without their paired inputs, should show the same roughly 75% accuracy plateau at the same training step; that experiment is directly runnable with the released code.
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

4 major / 5 minor

Summary. The paper studies how a two-layer, attention-only transformer acquires in-context learning (ICL) in a new task setting called In-Context Meta-Learning (ICML), where the model must infer a task from labeled examples rather than copy an answer from context. The authors report that training proceeds through three discrete phases—Non-Context Circuit (NCC), Semi-Context Circuit (SCC), and Full-Context Circuit (FCC)—each characterized by a distinct combination of attention patterns. They introduce attention-based metrics for these circuits, provide a theoretical accuracy analysis for SCC, show consistency with random-label robustness, explore how data properties alter phase structure, and extend the analysis to multi-head attention and GPT2-XL.

Significance. If the multi-phase claim is robust, the paper extends mechanistic interpretability of ICL beyond induction heads, connecting toy experiments to phenomena such as task vectors and random-label robustness. Strengths include a clean controlled setup, released code, a theoretical derivation for SCC accuracy that is empirically verified (Figure 5), and pruning experiments (Appendix C) that independently test the role of each identified circuit. These contributions are valuable regardless of the exact threshold used, but the central discrete-phase claim and the LLM generalization claim require stronger evidence.

major comments (4)
  1. [Section 4.1, Figure 3] The definition of phase boundaries uses the hand-set threshold ΔAccuracy > 0.025 with Δt = 100, and the paper never tests sensitivity to these choices or compares against an independent change-point detection method. Since the number and location of phases is the load-bearing premise for the claim that 'a unique circuit emerges in each phase,' the current evidence does not establish discreteness; it shows that three intervals can be carved from the accuracy curve with this particular threshold. I ask the authors to report phase boundaries across thresholds, window sizes, and random seeds, or to apply a model-free change-point criterion (e.g., Bayesian change-point analysis) to verify that three phases are not an artifact of the chosen threshold.
  2. [Section 4.2, Table 2 and Figure 4] The three attention metrics in Table 2 were defined after inspecting attention maps, and the phase boundaries in Figure 3 were chosen after observing accuracy jumps. The 'close correspondence' between metric transitions and phase boundaries in Figure 4 is therefore partly built into the analysis. This circularity is partially mitigated by the pruning experiments in Appendix C and the theoretical SCC analysis in Section 4.3, but the alignment claim should be validated by applying pre-specified metrics and threshold procedures to held-out seeds or newly trained models, rather than only to the runs used to define them.
  3. [Section 6, Figure 10] The GPT2-XL analysis uses raw attention probabilities averaged over all heads in each layer, from a single prompt format, with no error bars, no baseline comparison (e.g., random labels or permuted prompts), and no statistical test for the layer-wise differences. The claim that the observed circuits 'generalize to LLMs' is not supported by this evidence. Please add multiple prompts or seeds, confidence intervals, and a null model for attention, such as comparing to attention on shuffled token sequences.
  4. [Section 4.4, Figure 7] The claims that the model 'skips Phase 1' or 'skips Phase 2' for different values of K, ε, α, and β are made by visual inspection of the accuracy curves in Figure 7, without applying the operational ΔAccuracy criterion used elsewhere. Because these skipping claims are used to support the multi-phase framework, they should be evaluated with the same threshold-based or change-point procedure, and the results should be reported systematically.
minor comments (5)
  1. [Sections 3.2 and 4.2] The direction of attention is described inconsistently: Section 3.2 says p^{µ,h}_{ij} quantifies how strongly position i attends to position j, while Section 4.2 defines p^{µ,h}_{i,j} as the attention from token j to token i. Please unify the notation, especially because Table 2 depends on it.
  2. [Appendix F, Figure 6, Figure 7(a), Appendix N, References] There are several typos and formatting issues: 'Birstiness' in Appendix F header, 'Traning Accuracy' in Figure 6, 'nunber' in Figure 7(a), 'Experimemt' in Appendix N title, and 'V on Oswald' in the references (should be 'von Oswald').
  3. [Appendix J, Figure 19] The axis labels in Figure 19 contain the literal placeholder '/glyph1197', which appears to be a rendering error and should be fixed.
  4. [Figure 4 and Figure 2 captions] Figure 2 states that lighter-shaded curves indicate different random seeds, but Figure 4 does not clarify whether the metric curves are averaged over seeds or shown for a single run. Please state this explicitly for all plots.
  5. [Section 6, Figure 10] The metric definitions for the GPT2-XL analysis are too loose; for example, 'p(query,query)' needs clarification when there are multiple query-related positions in the prompt, and the averaging over heads and tokens should be specified precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central multi-phase claim is an empirical finding with independent theoretical, pruning, and external-model checks; the hand-set DeltaAccuracy threshold is a robustness concern, not a definitional reduction.

full rationale

The paper's central claim is that an ICML-trained transformer passes through three phases with distinct attention circuits. The phase boundaries are operationalized by a hand-chosen DeltaAccuracy threshold (Section 4.1), and the three attention metrics in Table 2 are defined from the visualized attention patterns. These choices are post hoc and could inflate the apparent alignment in Figure 4, but they do not constitute circularity: the circuit metrics are not fitted to the accuracy curve, the phases are not derived from the metrics, and the central claim would fail if the attention patterns did not actually change at those boundaries. The SCC theoretical accuracy (Section 4.3 and Appendix D) is a parameter-free combinatorial derivation tested against a pruned model, and the pruning experiments (Appendix C) and GPT2-XL layer analysis (Section 6) provide independent support. The only self-citations (Furuta et al. 2024; Minegishi et al. 2025) appear as background in Section K and are not load-bearing. Threshold sensitivity and post-hoc metric selection are methodological limitations, not circular reasoning.

Assumptions & free parameters 1 free parameters · 4 assumptions · 1 invented entities

The central empirical claim rests on a hand-chosen phase threshold and on the assumption that attention patterns reflect causal circuits. No fitted parameters are needed for the theoretical SCC accuracy, which is derived from explicit task conditions. The named circuits are descriptive constructs, not independent postulates.

free parameters (1)
  • Phase transition threshold Delta-Accuracy = 0.025
    Chosen by hand to mark phase boundaries in Section 4.1; the number and location of the three phases depend on this value.
assumptions (4)
  • domain assumption Two-layer attention-only transformer with a token-wise MLP classifier is a sufficient model class to study in-context meta-learning circuit emergence.
    All experiments use this architecture; Section 3.2. The paper does not prove that results extend to deeper or standard transformers, though Sections L and O provide some evidence.
  • domain assumption The synthetic task distribution (T=3, K=64, L=32, N=4, epsilon=0.1, pB=0) captures the essential characteristics of practical in-context meta-learning.
    Defined in Section 3.1; the claim that this setting requires task inference and is a practical ICL scenario depends on this choice.
  • domain assumption Attention probabilities are the right level of analysis for identifying circuits.
    The paper measures circuits solely via attention maps and metrics (Table 2), following mechanistic interpretability conventions; value and MLP computations are not directly analyzed.
  • ad hoc to paper The phase segmentation threshold Delta-Accuracy = 0.025 is a valid way to define learning phases.
    Introduced in Section 4.1; not derived or validated against other thresholds.
invented entities (1)
  • Named circuits NCC, SCC, FCC (Non-Context, Semi-Context, Full-Context Circuit) independent evidence
    purpose: Descriptive labels for the attention patterns observed in each learning phase: bigram-only, label attention, and chunk-example plus label attention.
    The circuits are operationalized with the attention metrics in Table 2 and the paper finds similar attention patterns in a pretrained GPT2-XL (Section 6), providing an external handle. They are descriptive abstractions rather than new physical entities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence." pith.science (2026). https://pith.science/paper/ND7ML7JQ

@misc{pith2026250516694,
  author       = {Pith},
  title        = {Pith review of: Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ND7ML7JQ}},
  note         = {Machine review of arXiv:2505.16694}
}
read the original abstract

Transformer-based language models exhibit In-Context Learning (ICL), where predictions are made adaptively based on context. While prior work links induction heads to ICL through a sudden jump in accuracy, this can only account for ICL when the answer is included within the context. However, an important property of practical ICL in large language models is the ability to meta-learn how to solve tasks from context, rather than just copying answers from context; how such an ability is obtained during training is largely unexplored. In this paper, we experimentally clarify how such meta-learning ability is acquired by analyzing the dynamics of the model's circuit during training. Specifically, we extend the copy task from previous research into an In-Context Meta Learning setting, where models must infer a task from examples to answer queries. Interestingly, in this setting, we find that there are multiple phases in the process of acquiring such abilities, and that a unique circuit emerges in each phase, contrasting with the single-phases change in induction heads. The emergence of such circuits can be related to several phenomena known in large language models, and our analysis lead to a deeper understanding of the source of the transformer's ICL ability.

Figures

Figures reproduced from arXiv: 2505.16694 by the authors.

Figure 1
Figure 1. (a) Task Structure: Previous studies focused on a copying-task setup, where the query’s answer remains unchanged by context, allowing the model to either memorize pairs or match and copy from context. In contrast, this work explores a more practical scenario where (x, ℓ) pairs vary by task, requiring the model to infer the task from examples and predict the query’s answer. (b) Network Structure: we mainly use two la… view at source ↗
Figure 2
Figure 2. (left) Changes in accuracy and loss across three distinct phases during training, with lighter-shaded curves indicating different random seeds. Each phase is highlighted with a different background color: Phase 1 (yellow), Phase 2 (orange), and Phase 3 (red). (right) Visualization of the attention maps (circuits) corresponding to each phase, with characteristic attention patterns indicated by red arrows and their ci… view at source ↗
Figure 3
Figure 3. Accuracy (blue) and ∆Accuracy (green) as functions of the training step. Here, ∆Accuracy = Acc(t + ∆t) − Acc(t) with ∆t = 100. Vertical dashed lines indicate where ∆Accuracy exceeds 0.025, marking the transition points between the three observed phases (Phase 1, Phase 2, Phase 3). The classifier is a two-layer MLP with ReLU activations, followed by a softmax layer producing probabilities over L labels. We train this… view at source ↗
Figures from the paper (18 more)
Figure 4
Figure 4. Figure 4: Evolution of the three attention metrics (Bigram, Label Attention, and Chunk Example) across optimization steps for the first (green) and second (red) layers. The shaded regions represent the three learning phases: Phase 1 (yellow), Phase 2 (orange), and Phase 3 (red),…
Figure 6
Figure 6. Figure 6: Comparison of training accuracy and random-label ac￾curacy (RLA). The plot demonstrates the rise in both metrics, with RLA following a trend similar to emergence Phase 2. This indicates that SCC acquired in Phase 2 contributes to improved accuracy even with shuffled la…
Figure 7
Figure 7. Figure 7: The relationship between learning phase dynamics and data distribution properties is explored by varying key parameters: the number of tasks (T), the number of classes (K), the noise magnitude (ϵ), the sampling bias for classes (k −α ), and the sampling bias for tasks …
Figure 8
Figure 8. Figure 8: Comparison of accuracy dynamics between single-head (blue) and multi-head (orange) attention models (left). The multi￾head model exhibits smoother accuracy improvements, without the distinct learning phases observed in the single-head model. On the right, the attention…
Figure 10
Figure 10. Figure 10: Layer-wise analysis of Bigram, Label Attention, and Chunk Example metrics in a pretrained LLM (GPT2-XL). We observe that chunk example scores peak in earlier layers while label attention scores are higher in middle or later layers, consistent with the final circuit (F…
Figure 11
Figure 11. Figure 11: illustrates an induction circuit consisting of a previous token head in Layer 1 and an induction head in Layer 2. After Layer 1, the side-by-side x and ℓ tokens are chunked into a single token. In Layer 2, two operations occur: matching of x via queries and keys (in p…
Figure 12
Figure 12. Figure 12: (left) The emergence of induction heads is observed as single-learning phase. (right) The attention maps on the right illustrate the circuit mechanism, where Layer 1 groups (x, ℓ) pairs into single-token representations, and Layer 2 then copies this label. 13 [PITH_F…
Figure 13
Figure 13. Figure 13: Controlled pruning experiments to validate the relationship between identified circuits and model performance. Networks trained with only the circuits from a specific phase plateaued at accuracies corresponding to that phase (yellow: Phase 1, orange: Phase 2, pink: Ph…
Figure 14
Figure 14. Figure 14: Rank-frequency distributions for different values of the power-law exponent α, following the Zipfian distribution f(k) = k −α . As α increases, the distribution becomes more skewed, with a few high-frequency items dominating while the majority appear infrequently. In …
Figure 15
Figure 15. Figure 15: (Left) Accuracy curves for different values of B at a fixed pB = 0.25. Increasing B tends to skip Phase 1, where NCC memorizes pairs through weights. (Right) Accuracy curves for different values of pB with B = 1. As pB increases, the learning process becomes smoother,…
Figure 16
Figure 16. Figure 16: (Left) shows accuracy curves over training steps for different numbers of attention heads (1, 2, 4, 8, and 16). Models with multiple heads exhibit a smooth increase in accuracy, whereas the single-head configuration undergoes multi-learning phases, where accuracy impr…
Figure 17
Figure 17. Figure 17: presents circuit metrics for each attention head, analyzed by layer in a two-head attention model. Head 1 consistently maintains high bigram values across both Layer 1 and Layer 2. This indicates that it primarily performs token-level copying operations, forming an NC…
Figure 18
Figure 18. Figure 18: shows accuracy (left) and loss (right) for models trained with increasing numbers of tasks (T = 3, 6, 9, 12, 15, 18). Even with higher T, models exhibit sudden accuracy jumps. As T increases, initial accuracy decreases, and it takes longer for models to achieve sharp …
Figure 19
Figure 19. Figure 19: (a) presents accuracy curves when increasing the number of few-shot examples (N), resulting in a longer total context. Multiple learning phases are visible for contexts up to N = 8. For N ≥ 16, the model quickly achieves perfect accuracy, indicating easier learning wi…
Figure 20
Figure 20. Figure 20: shows accuracy and attention maps for a standard transformer (with attention and MLP layers) trained on the same task as the simpler 2-layer attention-only model from the main text. At 50k steps, the model shows clear label-attention and bigram patterns similar to the…
Figure 21
Figure 21. Figure 21: shows accuracy curves for two conditions: predicting only the final label token (Last Label Only) and predicting every label token (All Labels). Both conditions show clear learning phases, but the All Labels setting does not reach perfect accuracy due to the need for …
Figure 22
Figure 22. Figure 22: shows accuracy curves of attention-only Transformers with 2 to 5 layers. Clear multiple learning phases are observed in the 2- and 3-layer models, whereas the 4- and 5-layer models exhibit smoother transitions without distinct phases [PITH_FULL_IMAGE:figures/full_fig…
Figure 23
Figure 23. Figure 23: Attention maps of 3-layer (top), 4-layer (middle), and 5-layer (bottom) models at 100% accuracy. In each model, the first layer displays a chunk example pattern, and the second layer exhibits label attention. This suggests that the core circuit (FCC) for achieving 100…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 14 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Understanding in-context learning in transformers and llms by learning to learn discrete functions

    Bhattamishra, S., Patel, A., Blunsom, P., and Kanade, V. Understanding in-context learning in transformers and llms by learning to learn discrete functions. arXiv preprint arXiv:2310.03016, 2023

    arXiv 2023

  3. [3]

    Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, ...

    arXiv 2005

  4. [4]

    Chan, S. C. Y., Santoro, A., Lampinen, A. K., Wang, J. X., Singh, A., Richemond, P. H., McClelland, J., and Hill, F. Data distributional properties drive emergent in-context learning in transformers. arXiv preprint arXiv:2205.05055, 2022

    arXiv 2022

  5. [5]

    Revisiting in-context learning inference circuit in large language models

    Cho, H., Kato, M., Sakai, Y., and Inoue, N. Revisiting in-context learning inference circuit in large language models. arXiv preprint arXiv:2410.04468, 2024

    arXiv 2024

  6. [7]

    N., Lynch, A., Heimersheim, S., and Garriga-Alonso, A

    Conmy, A., Mavor-Parker, A. N., Lynch, A., Heimersheim, S., and Garriga-Alonso, A. Towards automated circuit discovery for mechanistic interpretability, 2023 b . URL https://arxiv.org/abs/2304.14997

    arXiv 2023

  7. [8]

    and Shutova, E

    Crosbie, J. and Shutova, E. Induction heads as an essential mechanism for pattern matching in in-context learning. arXiv preprint arXiv:2407.07011, 2024

    arXiv 2024

  8. [9]

    Why can GPT learn in-context? language models secretly perform gradient descent as meta-optimizers

    Dai, D., Sun, Y., Dong, L., Hao, Y., Ma, S., Sui, Z., and Wei, F. Why can GPT learn in-context? language models secretly perform gradient descent as meta-optimizers. In Findings of the Association for Computational Linguistics: ACL 2023, pp.\ 4005--4019, 2023

    work page 2023

Show all 42 references
  1. [10]

    Selective induction heads: How transformers select causal structures in context

    D'Angelo, F., Croce, F., and Flammarion, N. Selective induction heads: How transformers select causal structures in context. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=bnJgzAQjWf

    2025

  2. [11]

    A survey on in-context learning

    Dong, Q., Li, L., Dai, D., Zheng, C., Ma, J., Li, R., Xia, H., Xu, J., Wu, Z., Chang, B., Sun, X., Li, L., and Sui, Z. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2024

    2024 arXiv

  3. [12]

    L., Malach, E., and Goel, S

    Edelman, E., Tsilivis, N., Edelman, B. L., Malach, E., and Goel, S. The evolution of statistical induction heads: In-context learning markov chains. In Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., and Zhang, C. (eds.), Advances in Neural Informati...

    2024

  4. [13]

    A mathematical framework for transformer circuits

    Elhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., Askell, A., Bai, Y., Chen, A., Conerly, T., DasSarma, N., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D., Jones, A., Kernion, J., Lovitt, L., Ndousse, K., Amodei, D., Brown, T., Clark, J., Kaplan...

    2021

  5. [14]

    Towards empirical interpretation of internal circuits and properties in grokked transformers on modular polynomials

    Furuta, H., Minegishi, G., Iwasawa, Y., and Matsuo, Y. Towards empirical interpretation of internal circuits and properties in grokked transformers on modular polynomials. Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URL https://openreview.net/forum?id=MzSf70uXJO

    2024

  6. [15]

    D., Tillman, H., Goh, G., Troll, R., Radford, A., Sutskever, I., Leike, J., and Wu, J

    Gao, L., la Tour, T. D., Tillman, H., Goh, G., Troll, R., Radford, A., Sutskever, I., Leike, J., and Wu, J. Scaling and evaluating sparse autoencoders, 2024. URL https://arxiv.org/abs/2406.04093

    2024 arXiv

  7. [16]

    S., and Valiant, G

    Garg, S., Tsipras, D., Liang, P. S., and Valiant, G. What can transformers learn in-context? a case study of simple function classes. In Advances in Neural Information Processing Systems, volume 35, pp.\ 30583--30598, 2022

    2022

  8. [17]

    Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic tasks

    He, T., Doshi, D., Das, A., and Gromov, A. Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic tasks. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=aVh9KRZdRk

    2024

  9. [18]

    In-context learning creates task vectors

    Hendel, R., Geva, M., and Globerson, A. In-context learning creates task vectors. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. URL https://openreview.net/forum?id=QYvFUlF19n

    2023

  10. [19]

    Loss landscape degeneracy drives stagewise development in transformers, 2025

    Hoogland, J., Wang, G., Farrugia-Roberts, M., Carroll, L., Wei, S., and Murfet, D. Loss landscape degeneracy drives stagewise development in transformers, 2025. URL https://arxiv.org/abs/2402.02364

    2025 arXiv

  11. [20]

    Meta-learning in neural networks: A survey

    Hospedales, T., Antoniou, A., Micaelli, P., and Storkey, A. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44 0 (9): 0 5149--5169, 2021

    2021

  12. [21]

    and Suzuki, T

    Kim, J. and Suzuki, T. Transformers learn nonlinear features in context: Nonconvex mean-field dynamics on the attention landscape. arXiv preprint arXiv:2402.01258, 2024

    2024 arXiv

  13. [22]

    Circuit component reuse across tasks in transformer language models

    Merullo, J., Eickhoff, C., and Pavlick, E. Circuit component reuse across tasks in transformer language models. arXiv preprint arXiv:2310.08744, 2024

    2024 arXiv

  14. [23]

    M eta ICL : Learning to learn in context

    Min, S., Lewis, M., Zettlemoyer, L., and Hajishirzi, H. M eta ICL : Learning to learn in context. In Carpuat, M., de Marneffe, M.-C., and Meza Ruiz, I. V. (eds.), Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics:...

    2022 doi

  15. [24]

    Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837, 2022 b

    Min, S., Lyu, X., Holtzman, A., Artetxe, M., Lewis, M., Hajishirzi, H., and Zettlemoyer, L. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837, 2022 b

    2022 arXiv

  16. [25]

    Bridging lottery ticket and grokking: Understanding grokking from inner structure of networks

    Minegishi, G., Iwasawa, Y., and Matsuo, Y. Bridging lottery ticket and grokking: Understanding grokking from inner structure of networks. Transactions on Machine Learning Research, 2025. ISSN 2835-8856. URL https://openreview.net/forum?id=eQeYyup1tm

    2025

  17. [26]

    Progress measures for grokking via mechanistic interpretability

    Nanda, N., Chan, L., Lieberum, T., Smith, J., and Steinhardt, J. Progress measures for grokking via mechanistic interpretability. arXiv preprint arXiv:2301.05217, 2023

    2023 arXiv

  18. [27]

    Zoom in: An introduction to circuits

    Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M., and Carter, S. Zoom in: An introduction to circuits. Distill, 2020. URL https://distill.pub/2020/circuits/zoom-in

    2020

  19. [28]

    In-context learning and induction heads

    Olsson, C., Elhage, N., Nanda, N., Joseph, N., DasSarma, N., Henighan, T., Mann, B., Askell, A., Bai, Y., Chen, A., Conerly, T., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D., Johnston, S., Jones, A., Kernion, J., Lovitt, L., Ndousse, K., Amodei, D., Brown, T., Cla...

    2022 arXiv

  20. [29]

    F., Lubana, E

    Park, C. F., Lubana, E. S., and Tanaka, H. Competition dynamics shape algorithmic phases of in-context learning. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=XgH1wfHSX8

    2025

  21. [30]

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

    Raventos, A., Paul, M., Chen, F., and Ganguli, S. Pretraining task diversity and the emergence of non-bayesian in-context learning for regression. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=BtAz4a5xDg

    2023

  22. [31]

    The mechanistic basis of data dependence and abrupt learning in an in-context classification task

    Reddy, G. The mechanistic basis of data dependence and abrupt learning in an in-context classification task. arXiv preprint arXiv:2312.03002, 2023

    2023 arXiv

  23. [32]

    Identifying semantic induction heads to understand in-context learning

    Ren, J., Guo, Q., Yan, H., Liu, D., Zhang, Q., Qiu, X., and Lin, D. Identifying semantic induction heads to understand in-context learning. arXiv preprint arXiv:2402.13055, 2024

    2024 arXiv

  24. [33]

    K., Chan, S

    Singh, A. K., Chan, S. C. Y., Moskovitz, T., Grant, E., Saxe, A. M., and Hill, F. The transient nature of emergent in-context learning in transformers. arXiv preprint arXiv:2311.08360, 2023

    2023 arXiv

  25. [34]

    K., Moskovitz, T., Hill, F., Chan, S

    Singh, A. K., Moskovitz, T., Hill, F., Chan, S. C., and Saxe, A. M. What needs to go right for an induction head? a mechanistic study of in-context learning circuits and their formation. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview....

    2024

  26. [35]

    L., McDougall, C., MacDiarmid, M., Freeman, C

    Templeton, A., Conerly, T., Marcus, J., Lindsey, J., Bricken, T., Chen, B., Pearce, A., Citro, C., Ameisen, E., Jones, A., Cunningham, H., Turner, N. L., McDougall, C., MacDiarmid, M., Freeman, C. D., Sumers, T. R., Rees, E., Batson, J., Jermyn, A., Carter, S., Olah, C., and H...

    2024

  27. [36]

    S., Mueller, A., Wallace, B

    Todd, E., Li, M., Sharma, A. S., Mueller, A., Wallace, B. C., and Bau, D. Function vectors in large language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=AwyxtyMwaG

    2024

  28. [37]

    N., Kaiser, L

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L. u., and Polosukhin, I. Attention is all you need. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R. (eds.), Advances in Neural Informatio...

    2017

  29. [38]

    a., Mordvintsev, A., Zhmoginov, A., and Vladymyrov, M

    Von Oswald, J., Niklasson, E., Randazzo, E., Sacramento, J. a., Mordvintsev, A., Zhmoginov, A., and Vladymyrov, M. Transformers learn in-context by gradient descent. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. JMLR.org, 2023

    2023

  30. [39]

    A., Vladymyrov, M., Pascanu, R., and Sacramento, J

    von Oswald, J., Niklasson, E., Schlegel, M., Kobayashi, S., Zucchet, N., Scherrer, N., Miller, N., Sandler, M., y Arcas, B. A., Vladymyrov, M., Pascanu, R., and Sacramento, J. Uncovering mesa-optimization algorithms in transformers. arXiv preprint arXiv:2309.05858, 2023

    2023 arXiv

  31. [40]

    Interpretability in the wild: a circuit for indirect object identification in gpt-2 small

    Wang, K., Variengien, A., Conmy, A., Shlegeris, B., and Steinhardt, J. Interpretability in the wild: a circuit for indirect object identification in gpt-2 small. arXiv preprint arXiv:2211.00593, 2022

    2022 arXiv

  32. [41]

    T., and Ni, L

    Wang, Y., Yao, Q., Kwok, J. T., and Ni, L. M. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53 0 (3): 0 1--34, 2020

    2020

  33. [42]

    M., Raghunathan, A., Liang, P., and Ma, T

    Xie, S. M., Raghunathan, A., Liang, P., and Ma, T. An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080, 2021

    2021 arXiv

  34. [43]

    Zipf, G. K. Human Behavior and the Principle of Least Effort. Addison-Wesley, Cambridge, MA, 1949

    1949

Pith tools

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