acceptodds
Under review as a conference paper at ICLR 2027

Borrowed Futures: Recipe-Dependent Attribution in Two Matched Revision Gates

Abstract

Attention gates such as CASTLE and Selective Attention let later tokens change how a language model reads earlier positions, a lookahead credited with their lower loss. The credit treats the share of the gain that needs the sequence's own future as fixed by the architecture. We build a pooled-lookahead gate and an accumulated mask from the published equations, each matched to a causal baseline in small models. Retraining a gate with its later-token input taken from an unrelated sequence, at unchanged parameters and FLOPs, isolates the future-free part of its gain. In the pooled gate this part is 91% of the gain at the shortest training length and indistinguishable from zero by a third of the Chinchilla ratio. In the accumulated mask it persists at the full Chinchilla ratio, and a causal model with only a data-dependent decay gains about as much as this part. Limiting the trained pooled gate's lookahead to a short window costs more than its whole gain, while retraining with that window costs nothing measurable. In both gates, the part that survives losing the future is not specific to revision. Attributing these gates' gains therefore needs a retrained substitution at the training length in question.

Primary area:
foundation or frontier models, including LLMs
Keywords:
language models, transformer architectures, attention mechanisms, ablation methodology, counterfactual retraining, training recipe, attribution, lookahead attention

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rx6KOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper asks how much of the gain from lookahead attention gates actually comes from reading a sequence’s own future tokens. The authors retrain each gate from scratch with its lookahead inputs borrowed from a different sequence, which splits the gain into a future-dependent part and a future-free part. In the experiments, the paper finds that in CASTLE, the future-free part dominates early in training but vanishes by 1B tokens. For Selective Attention, most of the benefit also does not need future tokens.

Critical strengths

The presentation of a look ahead gate is novel, as it provides a method to quantify if a model utilizes future lookahead on subsequent tokens.

Critical weaknesses
  1. Only one model size is mainly explored in the experiments, at 0.16B parameters. This does not even cover the 7B parameter models that are typically used in other training baselines.
  2. Only perplexity is ever evaluated. Downstream tasks like long-context retrieval or multi-hop reasoning is not explored. Tasks that require lookahead would be especially relevant to evaluate.
  3. The section 6 ablation is only conducted on one answer point, limiting its generalizability.
  4. The paper does not provide analysis as to why the pooled gate's future-free part disappears between 0.7B and 1B tokens.
Questions for authors

How would your implementation fare under a long context retrieval task, or multi-document summarization?

How do these future free gates affect larger models like 7B or 13B parameter models?

Reviewer_offaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies "lookahead"/"revision" attention methods that change how much a model attends to a past token using the tokens between it and the current prediction position. More specifically, the authors try to investigate whether the gain in lower pretraining language-modeling loss is from access to such "future tokens" or is a property of the architecture. During pretraining, the authors replace these inputs with representations from an unrelated sequence without changing the parameter count or attention computation. They compare prediction losses to measure the gains lost and retained across training budgets, including the Chinchilla reference of about 20 tokens per parameter. For the lookahead models that pool gated later token representations into a summary (e.g., CASTLE), the retained gain declines with longer training. For the attention design that accumulates penalties on past tokens (e.g., Selective Attention), more gain remains and is comparable to a simple forgetting/decaying model. The authors also present a preliminary ablation study on the lookahead window size for inference-time and re-training.

Critical strengths
  1. The paper asks an interesting question about why a lookahead/revision attention design helps -- is it mostly driven by data (the input sequences' own "future tokens") or driven by specific architecture design. Replacing later-token input while keeping parameters and nominal compute matched gives a useful controlled comparison, and repeating it across training budgets avoids relying on one early checkpoint.

  2. The simple forgetting model provides a relevant baseline for the more complex attention design. The window experiment also shows how strongly the answer can change between restricting an already trained model and training with that restriction.

Critical weaknesses
  1. Unclear Role of the Chinchilla Reference: The original Chinchilla laws use a normal autoregressive causal language model and establish a compute-optimal ratio through extensive study on different model sizes [1]. The paper uses about 20 tokens per parameter as a reference, and it does not establish that this is compute-optimal for the modified attention designs. It is not very clear why this reference should anchor the comparison between model variants, or how the findings extend to even larger models and longer training.

  2. Confounding Effects from Optimization: The same-round comparisons, early-stopping test, and higher peak-learning-rate check support the decline trend found for the pooled gate models. Robustness across other optimization choices is still unclear. Several comparisons for the accumulated-penalty design also span code versions and become uncertain under the paper's more conservative control for those changes. (Section 4; Appendices D-E; Table 4.)

  3. The paper too often reports a comparison without making its scientific meaning clear.: The substitution procedure and purpose of several controls are hard to follow from the main text. The conclusion section tries to provide design lessons for model architecture, but they are not consistently grounded in the experimental results. In particular, similar performance from the simple forgetting baseline does not establish that it explains the more complex design's retained gain.

References:

[1] Hoffmann, Jordan, et al. "An empirical analysis of compute-optimal large language model training." Advances in neural information processing systems 35 (2022): 30016-30030.

Questions for authors
  1. Could you explain why you anchor your analysis to the Chinchilla law? How is it relevant to ablation for future-dependent and future-free decomposition, considering you are working with a different model architecture and a different training setup from the Chinchilla paper?

  2. Could you clarify how sensitive the findings are to training choices beyond the controls already reported?

  3. Could you provide clear pseudocode and more explanation for the substitution process, the main experiment control for this paper? Ideally, it should be explained with some qualitative examples to show how information passes between sequences across layers. Please also give an explanation of how the experiment controls would go beyond 12-layer models.

  4. Could you give a clearer roadmap for the main insights for each experiment, instead of enumerating the experiment results?

Reviewer_D9oUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how much of the improvement from revision gates depends on the sequence’s own later tokens. The authors retrain two equation-based implementations—a pooled-lookahead gate and an accumulated mask—with their lookahead inputs borrowed from unrelated sequences, preserving the gate’s parameters and computation. This intervention splits the measured gain into “future-dependent” and “future-free” components. Experiments show that the pooled gate’s future-free benefit declines substantially with longer training recipes, whereas the accumulated mask retains a residual benefit comparable to that of a decay-only model under the main analysis. The paper also demonstrates a substantial discrepancy between inference-time window ablation and retraining under the same restriction.

Critical strengths
  1. Borrowing inputs from unrelated sequences preserves the gate’s computational structure while disrupting its access to sequence-specific lookahead information. Retraining allows adaptation to the intervention. Multiple seeds, shared evaluation windows, causality tests, and donor-cycle checks strengthen the experimental design.
  2. The pooled gate’s future-free benefit declines robustly across recipes. The window experiment is particularly informative: restricting lookahead after training increases loss by 0.102 nats, whereas retraining with the same restriction incurs no measurable cost. These findings clearly distinguish a trained model’s dependence on a component from the architecture’s ability to perform without it.
Critical weaknesses
  1. Appendix D reports architecture-dependent loss shifts, while the interval-widening procedure assumes a common round effect. Under the larger observed shift bound of 0.036 nats, the accumulated mask’s future-free gain and the decay-only gains have intervals that include zero (Table 4). Consequently, the persistent mask/decay findings are less secure than the pooled gate’s declining future-free benefit. The paper acknowledges this issue, but its prominence in the conclusions should better reflect the sensitivity.
  2. Similar gains from the substituted mask and the decay-only model are consistent with a shared explanation, but do not establish that distance decay causes the residual benefit. The mean-accumulation control in Appendix G changes both distance dependence and penalty magnitude, and its comparison with the substituted mask is statistically inconclusive. The decomposition therefore provides a useful intervention-specific measurement, while the mechanism behind the residual gain remains only partially resolved.
Questions for authors
  1. What specific code changes produced the training-round shifts? Can the key accumulated-mask, substituted-mask, causal, and decay-only comparisons at 3.21B tokens be validated using a single fixed implementation and a common training round?
  2. Can you isolate the effect of distance dependence from overall suppression strength—for example, by comparing penalties with matched overall magnitude but different distance profiles? This would more directly test the interpretation of the mask’s residual gain.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.