acceptodds
Withdrawn from ICLR 2027

Contextual Control of Transformers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Computational neuroscience models show that gain modulation can alter neural computation while connectivity remains fixed. Inspired by this principle, we ask whether context can similarly provide an explicit control signal over computation in a pretrained Transformer while its weights remain fixed. We introduce contextual control, a lightweight learned pathway that maps contextual states to bounded gains over feed-forward computation. Because the control signal is explicit, we can manipulate it while holding the input and prediction target fixed. Across GPT-2 and Qwen, contextual control improves prediction beyond static and token-only controls. We then ask whether the learned signal is merely context-dependent or whether its assignment to a particular context matters. Moving control between occurrences of the same token in different contexts changes prediction, and in pre-tanh control coordinates its absolute effect exceeds that of perturbations matched in layer-wise magnitude on both backbones. The assigned control also contributes positively to prediction on average, as neutralizing it increases loss. Together, these interventions show that contextual assignment has causal consequences within the learned control interface, providing an explicit framework for studying how context regulates computation in frozen Transformers.

Primary area:
interpretability and explainable AI
Keywords:
contextual control, context-dependent modulation, gain modulation, transformers, language models, frozen language models, mechanistic interpretability, intervention-based analysis.

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_yD17OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This work presents a method called contextual control, which learns a set of control signals that modulate the magnitude of the hidden activations at feed-forward blocks of a frozen transformer. They show that these control signals improve perplexity once learned on particular domains and reflects contextual information (superior to token-only modulations and noise baselines).

Critical strengths
  • The experiments are comprehensive, providing valid evidence for getting the control signal modulation to work and testing on different models and benchmarks.
  • The ablation experiments are also quite comprehensive, covering different baselines (token-only controls and multiple reassignment baselines) and scenarios (e.g. neutralization). Section 7's investigation of the small benefit on qwen on PG-19 is also a nice addition.
  • The motivation/connection to computational neuroscience is interesting.
Critical weaknesses
  • My biggest concern, and question after reading, is what insights we learn from these control signals. Beside a loose motivation from neuroscience and overall evidence that we can get these signals, I find the paper significantly lacking justification of why this method or these results are useful, especially beyond other interpretability methods that establish the contextual influence of token representations. If the control signals are more intended for interpretability, actual interpretable insights are missing from the write-up. The paper states "contextual control introduces an explicit, bounded signal whose neutral value exactly recovers the frozen model", however, it's unclear what we get from an explicit but obscure control signal.
  • The writing is quite hard to follow, and I think a large part is due to references of concepts like "cohort", "donor", "gain", "endpoint" that are not well-defined nor well-established.
Questions for authors
  • Could you clarify the practical value of this method?
  • Can you share the motivation behind the contiguous grouping assumption?
Reviewer_VZ9uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adds a small learned controller to a frozen language model and treats its output as an experimental variable for studying how context regulates computation. At every layer and position the controller reads the state entering the feed-forward block and sets bounded gains on eight groups of feed-forward channels. Zero control gives back the frozen model exactly, so the control at a single position can be replayed, reassigned to the control produced for the same token in another context, perturbed, or removed while the text and the target stay fixed. On GPT-2 and Qwen2.5-0.5B with WikiText-2 and PG-19, the contextual controller lowers perplexity more than static controls and token-only controls of similar size. Moving control between occurrences of the same token changes the loss, and in pre-tanh coordinates the absolute change is larger than for perturbations matched in per-layer magnitude. Removing the assigned control raises loss on average. The authors also look at where in the context the benefit sits, and report for Qwen that it is concentrated at the start of blocks on PG-19, and near section headers on WikiText-2, with a smaller effect persisting far from them.

Critical strengths

Having zero control exactly recover the frozen model is a good design choice, and together with exact replay of recorded controls it keeps the interventions clean. A single-position reassignment or neutralization measures the effect of that control alone, and differences of a few hundredths of a nat can be read without worrying that something else moved. I also found the comparator ladder in Table 2 useful beyond this paper. On Qwen the absolute contrast is negative against a random perturbation matched only in total norm, and it turns positive once the perturbation is matched layer by layer. Someone reading only the first row would draw the opposite conclusion, which shows how much an intervention result can depend on the choice of comparator.

Critical weaknesses

My main concern is that much of the predictive gain may come from position, or from a part that does not depend on context at all, rather than from what the context says, and the baselines cannot tell these apart. On GPT-2 with PG-19 a static control with under a hundred parameters already gives about three quarters of the contextual gain (Table 1), although the static run there has a single seed and the share is much smaller in the other settings. On Qwen with PG-19 the contextual gain is tiny to begin with (Fig. 4), and most of what the control contributes sits in the first 64 tokens of each block (Fig. 3a). A pre-specified splice test there found no recovery of that early benefit after artificial boundaries, which points to position or short context rather than to boundaries as such. On WikiText-2 the benefit at later positions is concentrated just after section headers, although a smaller effect persists far from them. The token-only baselines have neither context nor position, and, as the authors acknowledge, there is no position-only baseline. Their answer is that reassignment holds token and position fixed and is therefore unaffected, which puts most of the weight on the reassignment result.

The second issue is what the reassignment result shows. Reassignment alone raises the loss, but so do the comparators, and the paper says this establishes an effect, not donor specificity. Against random changes matched layer by layer in pre-tanh coordinates, the donor's change has a larger absolute effect, which shows that the change depends on which displacement is applied, beyond its size and layer allocation. The paper is careful to say that this does not establish that the original assignment is optimal or that the difference reflects semantic structure, so the result says less about context than the framing in the abstract suggests. On the Qwen follow-up, where the signed contrast CC was the pre-registered primary endpoint, CC is negative against every comparator, and against two of them its interval lies below zero (Table 6). The paper labels this unresolved under its pre-registered rule, but taken at face value, reassignment there is if anything less harmful than a random change. The abstract instead reports that, in pre-tanh coordinates, the absolute effect exceeds layer-matched perturbations on both backbones. On Qwen this absolute contrast AA is the secondary endpoint and only a few thousandths of a nat, and the paper itself notes that it is not established under gain-space matching. The only norm-matched comparator built from real controller outputs, comparator 4, differs in token as well as context (Section 9), so a direct test of context specificity is still missing (Question 1). The effects are uneven as well. On Qwen the median neutralization effect is close to zero and the control helps at barely more than half of the positions, and on GPT-2 the per-pair absolute contrast is only weakly correlated between seeds. The Qwen follow-up rests on two dozen articles and the neutralization numbers on 11. The initial GPT-2 cohort also gives an absolute contrast about half as large in every seed, for reasons the paper does not identify, and Section 9 notes that an earlier implementation gave smaller GPT-2 effects with a configuration that is no longer available. The paper is also hard to follow in places, since which endpoint is primary, the cohort rules and the "unresolved" labels are spread across Sections 3 and 5 and Appendices C–E, and there is no reproducibility statement or mention of releasing code, controllers or pair lists.

Questions for authors
  1. Suggestion: For a fixed token, does the donor effect depend on how similar the donor context is to the recipient context? Comparing similar and dissimilar donors would test context specificity directly, which comparator 4 cannot do because it differs in token as well as context.
  2. Clarification: How does a position-only controller, or one that reads the state from a shuffled position in the same block, compare on prediction? This would separate the contextual signal from position effects.
  3. Suggestion: State in the abstract that the signed contrast, the pre-registered primary endpoint in the Qwen follow-up, is unresolved, and that the absolute contrast reported there is the secondary endpoint.
  4. Clarification: Will the code, trained controllers and reassignment pair lists be released? The paper has no reproducibility statement.
  5. Suggestion: The biological motivation rests on gain modulation of neural computation, and Section 9 sets it aside because the control here is generated from the Transformer's own state. Astrocytes may be a closer analogue. Murphy-Royal et al. [R1] describe them as contextual gates that shape circuit function in an adaptive, state-dependent way, a framework they call contextual guidance, and neuron–astrocyte networks have been related to Transformer computation [R2, R3]. A short discussion of this line of work would help position the interface.

References:

[R1] Ciaran Murphy-Royal, ShiNung Ching, and Thomas Papouin. A conceptual framework for astrocyte function. Nature Neuroscience, 26(11):1848–1856, 2023. https://doi.org/10.1038/s41593-023-01448-8

[R2] Leo Kozachkov, Ksenia V. Kastanenka, and Dmitry Krotov. Building transformers from neurons and astrocytes. Proceedings of the National Academy of Sciences, 120(34):e2219150120, 2023. https://doi.org/10.1073/pnas.2219150120

[R3] Md Zesun Ahmed Mia, Malyaban Bal, and Abhronil Sengupta. Delving Deeper Into Astromorphic Transformers. IEEE Transactions on Cognitive and Developmental Systems, 17(6):1436–1446, 2025. https://doi.org/10.1109/TCDS.2025.3564285

Reviewer_je8uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a learned controller that maps contextual states of a frozen Transformer to bounded multiplicative gains over groups of feed-forward activations. The controller has an exact neutral setting, making its signal accessible to replay, reassignment, perturbation, and neutralization. Experiments on GPT-2 small and Qwen2.5-0.5B, using WikiText-2 and bounded PG-19 slices, compare contextual control with static and approximately parameter-matched token-only controls. Same-token cross-context reassignment is compared with perturbations matched in several ways, and single-position neutralization measures the contribution of the assigned control. Contextual controllers improve perplexity, and donor reassignment produces larger absolute loss changes than layer-wise matched comparators in pre-tanh coordinates. The signed comparison does not generalize across backbones, and the Qwen absolute contrast is inconclusive under post-tanh gain matching. Additional analyses examine attenuation of the benefit on PG-19 and at later block positions.

Critical strengths
  1. A concrete and well-validated interface for intervention experiments. The bounded gains, exact neutral point, same-token reassignment, and replay controls make the experimental variable explicit. Appendix D verifies properties including neutral equivalence, causality, frozen backbone weights, and replay consistency (pp. 12–13). These checks make it substantially easier to attribute an observed effect to the intended intervention within the augmented model.

  2. The paper investigates important confounds and reports unfavorable outcomes. It goes beyond a single isotropic perturbation by matching layer-wise norms, restricting directions to the decoder column space, and considering real displacements and post-tanh gain matching (Table 2 and Appendix E). It also reports the unresolved Qwen primary endpoint, the stronger LoRA reference, and the limited PG-19 benefit. This provides useful evidence about how fragile apparently straightforward claims of assignment specificity can be.

Critical weaknesses
  1. The strongest cross-backbone result is a coordinate- and comparator-dependent absolute effect, which leaves the substantive assignment-specific finding narrow. On Qwen, the absolute donor-minus-null contrast changes from −0.0088 nats under total-norm matching to +0.0077 under decoder-subspace, per-layer matching, but becomes +0.0023 with a 95% interval of [−0.00004, +0.0046] under post-tanh matching (Table 2, p. 5). The signed contrast was the pre-registered Qwen primary endpoint; the pooled signed estimates are negative and all are unresolved under its decision rule (Appendix E, pp. 15–16). Since gains actually act after tanh, pre-tanh norm matching does not control effective intervention magnitude. The authors appropriately acknowledge this boundary, but acknowledging it does not establish a robust assignment-specific advantage in the operational coordinates. Neutralization and positive mean donor loss establish usefulness and causal influence within the learned interface, but do not resolve this comparator issue.

  2. The remaining methodological and scientific contribution is insufficiently distinguished from established conditional modulation. The architecture combines familiar input-conditioned gains with a small bottleneck and a bounded neutral parameterization. Its main demonstrated properties—improved prediction, sensitivity to swapping its input-dependent signal, and degradation upon removing a trained component—are also plausible for ordinary conditional adapters. The paper does not yet demonstrate an inference or explanatory capability that clearly distinguishes this interface from that broader class. Furthermore, the token-only controls omit position as well as preceding context (Section 9, p. 8), so the perplexity improvement does not isolate contextual content from position-dependent calibration, particularly given the pronounced position effects on PG-19. The stronger rank-1 LoRA result in Appendix K is not itself disqualifying, since competitive adaptation is not the stated goal; it makes the distinctive scientific payoff of the proposed interface especially important to establish.

Questions for authors
  1. Clarification and assessment-critical validation: Can you provide the Qwen post-tanh matching diagnostics that Appendix E states were unavailable, and evaluate the donor contrast with verified gain-space matching and a real-displacement comparator conditioned on the same token identity? Please report both absolute and signed contrasts. This would help distinguish contextual assignment from differences in saturation, effective magnitude, and comparator token identity.

  2. Assessment-critical comparison: How much of the predictive advantage remains against a parameter-comparable token-and-position controller? The existing same-position reassignment experiment is useful, but does not resolve the separate interpretation of the corpus-level predictive gains. A comparison that includes position would directly affect the claimed benefit of access to contextual content.

  3. Contribution clarification: What specific inference about context-dependent computation has this interface enabled that would not follow from applying the same replay, swapping, and ablation procedures to an ordinary input-conditioned gain adapter? A concrete supported finding would help assess the contribution at its explicitly stated scope, without requiring a claim about native backbone circuitry or biological mechanisms.

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

Related papers

Discussion (0)

Sign in to comment.