acceptodds
Under review as a conference paper at ICLR 2027

When Data Can Teach: Prerequisite Circuits Gate Learning

Abstract

The same training data can teach a capability to one model and nothing to another. Curricula and data attribution depend on this effect, yet it remains unclear which parts of a model's state determine what data can teach. We test whether a previously learned circuit acts as a prerequisite: holding the training data fixed, we suppress an induction circuit only while a transformer trains on a task that builds on it, and we evaluate every model with the circuit intact. In a two-layer transformer, suppressing the circuit's previous-token head during training blocks learning of long-range copying, and supplying the head's output restores it; without the circuit, the same data teach nothing for 600,000 updates. In pretrained Pythia models from 160M to 6.9B parameters, suppression blocks or delays learning of a retrieval task that builds on induction, whereas suppressing control heads does not. With longer training, most pretrained models learn the task by rebuilding the retrieval circuit from dormant backup heads. Suppressing these heads after training removes the capability, and suppressing them from the start prevents learning at 410M. The delay, 2 to 18 times the unimpaired learning time, depends on how many backup heads remain. Previously learned circuits can thus determine whether, and how quickly, later data teach a new capability.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, learning dynamics, causal interventions, transformer circuits, curriculum learning, data attribution

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_VjKAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors studies whether circuits learned earlier affect a model’s ability to learn from later data. The authors suppress induction-related heads during target-task training, keep the data fixed, and restore the heads for evaluation. In a two-layer transformer, suppressing a previous-token head blocks long-range copying, while supplying its output restores learning. In early Pythia checkpoints, induction-head suppression prevents or delays learning of a mapped-copy task. Longer runs show recovery through other heads, investigated through ablations and adaptive suppression. These results demonstrate that the selected interventions affect learning, but the evidence that recovery requires rebuilding the same prerequisite computation is less conclusive: models can learn the target while ordinary induction remains suppressed. The restriction to two copy-based tasks and early checkpoints from one pretrained model family also limits the broader conclusions about when data can teach a model.

Critical strengths
  1. Training-only suppression addresses a question that evaluation-time ablation cannot answer. Keeping the target batches identical and restoring heads for evaluation is a useful design for studying how internal computations affect learning.

  2. The central result is supported by repeated runs and several controls. The sixteen Pythia-160M pretraining runs, loss-matched controls, and key-dependence tests make the observed learning difference difficult to dismiss as an isolated failure or a simple positional shortcut.

  3. The longer training runs qualify the initial blocking result. Showing that models can eventually recover, and examining the heads involved in recovery, adds useful evidence about the difference between preventing learning and delaying it.

Critical weaknesses
  1. The prerequisite-rebuilding interpretation is not sufficiently established. Target-task learning precedes recovery on the ordinary-induction probe at several model sizes. More importantly, adaptive suppression triggered by ordinary induction keeps that score near zero while the model still learns a key-dependent solution. The subsequent key-triggered intervention suppresses heads based on retrieval performance on the target itself. This shows that disrupting emerging target-relevant computation can prevent learning, but does not establish that the original induction prerequisite must be rebuilt. The paper needs a clearer account of what computation is necessary and how the replacement is shown to implement it.

  2. Whole-head suppression leaves uncertainty about the cause of the learning failure. Removing a head deletes all of its contributions, not just induction, and changes the representations used by downstream layers. The loss-matched controls are helpful, but matching damage to natural-text loss does not match disruption on the target examples. In Pythia, the suppressed heads also continue to change through other training examples, so enabling them at evaluation does not restore their original function. The evidence supports a causal effect of the selected intervention, but its specificity to induction remains less clear. Finer interventions would strengthen the central interpretation.

  3. The experimental scope is too narrow for the broader claims. There are many repetitions and intervention variants, but they all concern one prerequisite and two closely related copy-based tasks. The pretrained experiments use early checkpoints from a single model family. These are reasonable settings for a case study, but they provide limited evidence about other capabilities or learning from natural data. No experiment tests whether circuit information improves curriculum design or predicts data usefulness outside these tasks. This limits the significance of the paper’s broader framing.

Questions for authors

refer to the weakness, if you can solve my questions, I will raise my score

1 comment

Thank you for stating precisely what would establish the interpretation. We answer the three weaknesses in order.

1. Must recovery rebuild the original prerequisite? No. We agree and corrected the claim in the abstract, Section 5 and the title of Figure 3. Recovery can use a specialized route, and the acquisition result is unchanged: removing selected computations changes what identical subsequent data teach within the tested budget.

The shared causal contribution we identify is transmission of source-position information through the selected heads. We test it with the same operation in the original and the replacement heads. In the original heads, replacing only the source-position value prevents learning, and keeping only that value preserves it (point 2). In the replacement heads of a recovering Pythia-160M model, replacing that value lowers target accuracy from 84.4% to 1.2%; removing the four heads lowers it to 3.9%; removing four matched heads leaves 85.5% (Appendix U).

How the replacement implements it. With the original heads suppressed, the model outputs the mapped partner on 80.5% of marker-source probes and the marker itself on 9.8%, while ordinary induction is zero. Key relocation gives 73.6% paired accuracy and deleting the key lowers accuracy to 2.7%, so the solution uses the key. The model uses the replacement heads for source-dependent, task-specific recovery; the mapping may also involve other components. In the frozen-lower-network 1B experiment, the upper layers reach 98% target accuracy and pass the key tests while ordinary induction stays at or below 2/256, all with suppression in place (Appendix N). The key-triggered adaptive intervention disrupts the emerging target route; it does not show that this route is the original induction algorithm.

2. Does whole-head damage explain the effect? We edited one vector per selected head: the value it reads from the source position. The operation preserves queries, keys, attention and other value entries at that layer. Earlier edits can change later-layer inputs; other heads retain source access (Appendix R). Replacing only this vector with a distractor's value prevents learning (at most 6.6% over 500 updates). Keeping only this vector, with the head's other values set to zero, leaves learning unchanged (crossing at update 200, as without intervention; 99.3% of key pairs). The source-only arm retains its source-value gradient. Comparing detached correct and distractor values then matches the local gradient cut while varying content. In three pretraining runs × two streams, correct content permits learning in all six cells (88.3–94.9%, all passing the key tests), while the distractor's content stays below 50% (Appendix S). Natural-text-loss controls are complementary; we do not claim they match damage on the target examples.

Does enabling a head restore its original function? In the controlled model, yes: its weights and inputs are fixed. In pretrained models, enabling heads only removes the intervention. Holding the selected heads' weights at their initial values still prevents learning (at most 2.7% over 500 updates) while matched frozen controls cross at update 200 (Appendix R). New experiments report evaluation with the intervention removed and retained. Within 500 updates no suppressed model in these experiments reaches 50% in either mode; in longer runs two GPT-2 models learn under the mask about 1.5-fold later than their controls (General response 2).

3. Scope. Five independently pretrained GPT-2 models, three final GPT-2 checkpoints and final Pythia-160M reproduce the effect (Appendix T). The tasks remain copying-based, and the paper does not demonstrate a readiness predictor or a curriculum benefit (Section 8).

Summary. (1) The rebuilding claim is corrected; source-value interventions identify a causal contribution in both the original and replacement heads. (2) Edits confined to one source-position vector per head reproduce the effect, and the acquisition contrast persists with matched local gradient cuts and with head weights fixed. (3) Coverage now includes a second model family and final checkpoints; task generality remains a stated limitation.

Reviewer_A1BEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether previously learned circuits can act as prerequisites for learning new tasks. Specifically, it asks whether induction circuits/previous-token heads effectively gate the ability of models to learn mode demanding copying tasks.

First, the authors study a two-layer transformer trained on short-range copying before training it for long-range copying. They find that suppressing the previous-token head during long-copy training prevents the model from learning the task, even when the head is restored for evaluation. They also find that long-copy examples teach the model only little when presented before the induction circuit forms, but become effective when replayed after short-copy training has established it.

The authors then take a look at early pretrained Pythia checkpoints on a mapped-copy task (retrieve the marker associated with a repeated key and predict its partner under a fixed mapping.). They, again, study the effect of training order and find that if they first induce learning an induction circuit and then introduce the task the model is able to learn the task. This does not appear to be the case when the order of the data is reversed (first train on the task and then on data that incentivises learning an induction circuit).

The authors then study the effect of suppressing heads relevant for induction behaviour during training on the mapped-copy task. They find that this substantially delays learning, but the models eventually recover. Finally, they experiment with ablating additional heads during evaluation and training. Adaptively suppressing more heads when the model passes the key-swap test prevents a reliable key-dependent solution within the tested period, whereas adding heads when ordinary induction returns does not prevent learning.

Critical strengths
  • The paper studies an interesting question about dependencies between learned computations and the implicit curricula that may arise during training. It presents several complementary experiments, including training-order comparisons, replay, and training-only suppression, to investigate why the same examples become more useful after a model acquires another capability. The controls that match the damage to language modelling caused by head suppression is an interesting idea.
Critical weaknesses
  • The authors rank heads by prefix matching, which they call “induction attention”, and select a set whose ablation reduces the model’s copying accuracy. This differs from the copying score in Olsson et al. (2022; In-context learning and induction heads), which measures whether an individual head’s output directly raises the prediction score for the token it attends to. As a result, ablations show that the selected heads collectively support copying, but do not verify this copying behaviour in each head. The authors should report head-level copying scores or equivalent evidence to support identifying these heads as induction heads. Otherwise, the claims should refer more narrowly to heads with high prefix-matching scores that support copying.
  • I found the presentation difficult to follow because the paper moves between different models, tasks, and interventions without separating them clearly enough. Details such as which parameters remain trainable, whether heads are restored during evaluation, and how resetting differs from suppression are important for interpreting the findings.
Questions for authors
  • How well does the short-copy-trained model perform on long copying before long-copy training?
  • In the two-layer model’s output-restoration experiment (Figure 1C), you suppress the previous-token head and add the output of an identical frozen copy. Since the original head and its inputs are frozen in all conditions, is this equivalent to leaving the original head active? If so, what additional evidence does this experiment provide?
1 comment

Thank you for pressing on head identification and on the clarity of the intervention details.

Do the selected heads individually copy? (Section 2; Appendix Q) Yes. For all 64 selected heads (four in each of the 16 Pythia-160M runs):

  • OV eigenvalue diagnostic used by Olsson et al. On our 4,932-token probe vocabulary, the score is positive for all 64 (0.20–0.46, median 0.30); layer-matched controls score −0.15 to 0.13 (median −0.02). Among the 132 heads above layer 0, the selected heads rank at median 3 within their run.
  • Direct effect on the attended token. On 128 fresh random-token sequences, the part of each head's output that comes from the source position raises the source token's logit relative to the average token (all 64 positive, 0.081–1.470; controls −0.0005 to 0.0016). This provides direct output-based evidence without the normalization in Olsson et al.'s activation score.
  • Single-head ablation. Removing each head alone lowers the copied token's log-probability by 0.057–5.595 nats; for matched heads the drop is at most 0.011 nats.

Selection still uses prefix-matching attention (renamed from "induction attention"), because layer-0 heads also have high copying scores (mean 0.30).

Which parameters train, and what is restored? (Section 2; Table 1) Table 1 lists each intervention, what changes during training and the evaluation mode.

  • Controlled two-layer model: the suppressed head and the parameters producing its inputs are frozen in every condition, so enabling the head at evaluation restores its original computation exactly.
  • Pythia and GPT-2: all parameters train; only target-training rows are masked. Enabling the heads removes the intervention but keeps training-induced changes, so new experiments report evaluation with the intervention removed and retained.
  • A reset reinitializes parameters permanently; suppression removes a head's output temporarily.

Each result paragraph states its endpoint (50% crossing or the stricter key-dependent criterion) and horizon.

Q1: Long-copy performance before long-copy training (Section 3). The three short-copy-trained models start with copy gains of 0.26, 0.27 and 0.34 nats, below the 1-nat criterion, and distance-64 accuracies of 19.1%, 0.4% and 27.7%. Short-copy training gives partial transfer; Section 3 reports these values before the learning results.

Q2: Is output supply equivalent to leaving the head active? (Section 3; Appendix Q) Yes. With the head and its inputs frozen, supplying its identical output gives the same forward computation and the same gradients for trainable parameters. A float64 check over three optimizer steps finds maximum differences of 8.5×10−148.5\times10^{-14} in logits, 1.3×10−151.3\times10^{-15} in gradients and 6×10−146\times10^{-14} in parameters. Section 3 now presents supply as a check that the intervention removes and restores exactly the head's contribution; supplying another head's output does not restore learning. The content-specific test in pretrained models is the new source-value experiment: replacing only the value each selected head reads from the source position prevents learning, while keeping only that value preserves it (General response 1).

Relation to prior work. Prior work clamps subcircuits while induction forms (Singh et al., 2024) or suppresses syntactic attention during pretraining (Chen et al., 2024). Here the circuit is already learned, the subsequent data stream is fixed exactly, and the intervention is limited to training on the target. In Pythia, those batches yield no target accuracy after preparation alone, but produce acquisition within 100 updates when replayed. Separate suppression runs recover through backup heads that already attend to the source.

Reviewer_AiwyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether a circuit that a model has already learned can decide whether it learns a new capability from later data. Keeping the training data and its order fixed, the authors zero the output of a prerequisite induction circuit only on the new task's training examples and evaluate every model with the circuit intact (training-only suppression). In a two-layer transformer, suppressing the previous-token head prevents learning of long-range copying, even over 600,000 updates, while supplying the head's output restores learning. In Pythia-160M, suppressing induction heads prevents learning of a retrieval task in all 16 public pretraining runs. From 160M to 6.9B, suppressed models eventually learn after 2–18 times as many updates as controls, by rebuilding the circuit from backup heads that attend to the right tokens before training but do not copy them; the number of remaining backup heads governs the delay.

Critical strengths
  1. Rigorous causal design. The training data and its order are fixed across conditions, and every model is evaluated with the circuit intact. On top of this, the paper adds output-supply experiments, wrong-head supply controls, loss-matched controls that harm language modeling as much as the suppressed heads, and a sign test over 16 public pretraining runs. Together these carefully support the causal claim that the circuit is needed for learning, not only for performing, the task. The paper is also transparent: it states which choices were fixed before which outcomes and reports failed and interrupted runs.

  2. The backup-head rebuilding finding. The paper shows that pretrained models rebuild a suppressed circuit from backup heads that already attend to the correct source before training but do not copy it. This is supported by three lines of evidence: the delay shortens as more backup heads remain, suppressing the rebuilt heads removes the capability, and suppressing the circuit as it forms prevents learning. This is a genuinely new insight into learning dynamics.

Critical weaknesses
  1. The central claim is intuitive, and no alternative hypothesis is articulated. The core hypothesis, that a capability built on an existing computation is hard to learn when that computation is unavailable, is intuitive and close to what existing theory (e.g., leap complexity) and the known staged formation of induction circuits would predict. The paper does not state what outcome would have been plausible if the hypothesis were false—for instance, that a large model could learn the target task directly from the data through another route, as indeed happens when induction heads are reset in the two-layer model. As a result, it is hard to see what is surprising about the headline result. In my reading, the paper's main value lies in its methodology (separating "needed to perform" from "needed to learn") and in the backup-head rebuilding finding, but these are not emphasized as the central contributions.

  2. Generality is limited, and the experiments are distant from the stated motivation. The study examines a single prerequisite (induction) and two synthetic copy-based target tasks. In the two-layer model in particular, the target task (long-range copying) is close to an extension of induction itself, which blurs the distinction between the prerequisite and the new capability. The Pythia experiments start from very early pretraining checkpoints, and comparisons across sizes are confounded by architecture, starting checkpoint, number of suppressed heads, and training length. Finally, while the introduction and abstract motivate the work through curricula and data attribution, the paper does not show how either would change in practice.

Questions for authors
  1. How do the authors explain the rebuilding mechanism theoretically? In particular, how do they relate to frameworks such as leap complexity (a) the asymmetry in the two-layer model, where removing the previous-token head (the "writer") blocks learning while resetting the induction heads (the "readers") allows rapid recovery, and (b) the role of backup heads as starting points for rebuilding?

  2. Is the rebuilt circuit initially specialized to the target task (e.g., biased toward outputting the 32 marker tokens)? Would this explain why target accuracy recovers before general induction (masked induction) does?

1 comment

Thank you for the careful reading and for identifying the training-only method and the backup-head recovery as the paper's main value. We agree; the revised introduction lists them as the contributions (Section 1).

What would the data have taught without a gate? Three alternatives were plausible, and the experiments test each:

  1. Direct learning. The source stays in the input and all other parameters train, so a model could learn the task by another route. Downstream-reset two-layer models do recover within 500 updates. Under suppression, Pythia-160M learns in none of 32 run–stream pairs (16 pretraining runs) within 500 updates, while every matched control learns.
  2. Expression of early exposure after preparation. Target batches presented early could yield target performance once induction forms. Instead, target-first accuracy is still zero after preparation; replaying the same 500 batches then produces acquisition (Figure 2A–B). This measures expressed performance; latent changes may still have occurred.
  3. Generic damage. Removing any heads that harm language modeling as much might slow learning equally. Sixteen loss-matched heads preserve rapid learning (updates 350/400 versus 350); the four induction heads delay it to 6,300–6,400.

Q1: writer/reader asymmetry, backup heads and leap complexity. A retained writer exposes a feature that a reader can learn to use. Resetting the writer changes that feature-learning problem; surviving source-routing heads provide alternative starting points. Leap complexity offers a compatible framework, conditional on a specified representation, target expansion and learning rule. We do not equate leap complexity with the number of processing steps or assign a leap index to these transformers (Section 6; Appendix V). Two measurements bear on it:

  • Optimization sets the length of the block. Without weight decay, the three upstream-reset models recover after 39,000, 100,000 and 121,000 updates; downstream-reset models recover within 500 updates even with decay. These comparisons establish optimizer dependence of the upstream block; the downstream reference uses the original optimizer (Section 3).
  • Backup heads shorten the delay. At 410M and 1B they attend to the source before training (prefix-matching attention 0.11–0.67, against a median of 0.01). Within a model, the delay grows as fewer remain: at 160M, suppressing two, three or four heads gives learning at 1,100–1,300 updates, none within 2,000, and 6,300–6,400; at 410M, suppressing all 13 heads with attention of at least 0.2 prevents learning for 2,000 updates (Section 5).

Q2: Are the replacement outputs specialized to the target? Yes, in the recovering Pythia-160M model we analyzed (Section 5; Appendix U). With the original heads suppressed, the model outputs the mapped partner on 80.5% of marker-source probes and the marker itself on 9.8%, while ordinary induction accuracy is zero. Removing the four replacement heads lowers target accuracy from 84.4% to 3.9%; replacing their source values lowers it to 1.2%; removing four matched heads leaves 85.5%. The model uses the replacement heads for source-dependent, task-specific recovery, supporting this explanation of the earlier target recovery. Other components may implement part of the mapping. In the frozen-lower-network 1B experiment, the upper layers reach 98% target accuracy and pass the key tests under suppression while ordinary induction stays at or below 2/256 (Appendix N).

Generality and motivation. Five GPT-2 pretraining runs and final GPT-2 and Pythia checkpoints reproduce the effect (Section 4; Appendix T). Cross-size comparisons remain separate case studies, since architecture, starting checkpoint and the number of suppressed heads differ; the 2.8B control is descriptive only. Short-to-long copying extends the same computation; mapped copy adds a learned mapping and key tests that separate the target from the prerequisite. A practical curriculum or attribution method is not demonstrated (Section 8).

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of revisions and the source-content experiment

We thank the reviewers for their detailed reviews. The revised PDF marks added and replaced wording in blue. The table maps each concern to the new evidence.

ConcernNew evidenceLocation
Selected heads may not copy individually (A1BE)Individual copying evidence for all 64 selected heads: OV score, direct contribution, ablationSec. 2; App. Q
Whole-head suppression is not specific (VjKA)Editing only the value each head reads from the source position reproduces the effect; keeping only that value preserves learningSec. 4; App. R, S
Head drift and evaluation mode (VjKA)Head weights held fixed; new runs evaluated with the intervention removed and retainedSec. 4; App. R, T
Recovery need not rebuild induction (VjKA; Aiwy Q2)Claim corrected; replacement heads carry source information used for task-specific recoveryAbstract; Sec. 5; App. U
Early checkpoints of one family (all)Five GPT-2 pretraining runs; final GPT-2 and Pythia checkpointsSec. 4; App. T
Alternatives and theory (Aiwy)Alternatives stated; zero-decay continuations establish optimizer dependenceSec. 1, 3, 6; App. V
Presentation (A1BE)Intervention table; initial long-copy scores; supply equivalenceSec. 2, 3; Table 1; App. Q

Source-position content carries the effect (Section 4; Appendices R and S). Whole-head suppression removes everything a head contributes. We therefore edit one vector per selected head, on target-training rows only: the value the head reads from the source position (the marker after the first key). At each edited layer, queries, keys, attention and other value entries are unchanged by the operation. Earlier edits can affect later-layer inputs; the source token stays available to other heads.

  • In the original Pythia-160M run, replacing this vector with the value of an earlier distractor prevents learning: target accuracy stays at or below 6.6% for 500 updates on both streams.
  • Keeping only this vector, with the head's other values set to zero, leaves learning unchanged: the model crosses 50% at update 200, as without intervention, and passes 99.3% of key-relocation pairs.

Distractor substitution detaches the donor value; the source-only arm retains the source-value gradient. To isolate content, we compare detached correct and distractor values, which share the local gradient cut. This comparison used three pretraining runs with two data streams each (24 runs; models, heads, streams, horizon and endpoint fixed before outcomes). All six correct-content cells learn (88.3–94.9% target accuracy, crossing at 300–350 updates versus 200 for intact training) and pass the key tests: 80.3–95.2% of key-relocation pairs, and deleting the key lowers accuracy to 3.1–5.9%. With the distractor's content, no cell reaches 50% within 500 updates. The selected heads therefore affect learning through the content they carry from the source position.

Some wrong-content runs partially recover (endpoints up to 42.6%), and changed content also alters downstream gradients. Wrong content is misleading rather than missing information; whole-head suppression, the missing-information case, also prevents learning.

AuthorsOpenReview ↗

Model coverage and evaluation modes

A second model family and final checkpoints (Section 4; Appendix T). Five independently pretrained GPT-2 small models (Stanford CRFM, checkpoint 2000) reproduce the 500-update contrast on two data streams each: every matched control crosses 50% target accuracy and no suppressed model does (one-sided sign test over the five pretraining runs, p=1/32p = 1/32). On the first stream, controls cross at the same update as unintervened training (400–500). At the final checkpoints of three CRFM runs, controls cross at 240–320 updates and suppressed models stay below 50% through 500. At the final Pythia-160M checkpoint, the control crosses at 550 and the suppressed model stays below 50% through 1,000. The task is still synthetic mapped retrieval, and the 50% crossing is weaker than the key-dependent criterion; Appendix T reports the key tests.

Both evaluation modes (Section 4; Appendix T). Within 500 updates, every suppressed model in the GPT-2, final-checkpoint and fixed-weight experiments stays below 50% both with all heads enabled and with the training intervention retained (at most 27% in the latter mode). Longer runs show that two GPT-2 models adapt to the intervention:

  • Final OpenAI GPT-2 crosses 50% under its training mask at update 800 (control 550) and reaches 99.6%, while accuracy with all heads enabled stays at or below 31.2% through 1,500 updates.
  • CRFM run 1 crosses under its mask at 600/650 (controls 400/450).

In these models suppression delays acquisition by about 1.5-fold, consistent with recovery through backup heads (Section 5). Their key tests were recorded only with all heads enabled, so we report accuracy recovery under the mask.

Fixed head weights (Appendix R). Holding the selected heads' query/key/value parameters and output-projection columns at their initial values preserves the Pythia contrast: suppressed models stay at or below 2.7% over 500 updates, while matched frozen controls cross at update 200. Their inputs can still change, so enabling the heads restores their weights but not necessarily their computation. Exact restoration holds in the controlled two-layer model, where the head and the parameters producing its inputs are frozen.

Task specificity (Appendix R). Suppressing the same four induction heads leaves learning of a task that maps the immediately preceding marker to its partner unaffected (crossing at 200/150 updates versus 200/200 without suppression). Suppressing two previous-token heads prevents learning of mapped copy within 500 updates, while the preceding-marker task learns at 350/300. The suppression affects the task that requires retrieving a distant source and spares the mapping of an adjacent marker.

Scope. These results extend model family and checkpoint age within copying-based retrieval. Other capability families, acquisition from natural data and a practical curriculum method remain open (Section 8).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.