acceptodds
Withdrawn from ICLR 2027

BICD: Bidirectional Circuit Discovery by Dynamic Forward Selection and Backward Pruning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Mechanistic interpretability (MI) has developed many mature methods for discovering subcircuits in large language models. However, methods that abstract an LLM as a computational graph often face an enormous number of edges and consequently inefficient pruning. Methods based on activation patching, on the other hand, depend heavily on manually specified patching interventions, and the subcircuits they discover often have limited ability to perform the task independently. Drawing on the complementary ideas of forward search and backward pruning, we propose Bidirectional Circuit Discovery (BiCD), a simple and efficient bidirectional search algorithm. BiCD seeks to obtain a subcircuit that is as small as possible while preserving as much of the model's functionality as possible.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretability, Automated Circuit Discovery, Transformer Circuits

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_NuaeOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Bidirectional Circuit Discovery (BiCD), an automated circuit discovery method for mechanistic interpretability. BiCD operates at the component node level under zero ablation and a correct-answer cross-entropy (CE) constraint, to find a small subcircuit capable of performing fine-tuned arithmetic tasks on GPT-2 small and medium.

Critical strengths

Combining dynamic gradient rescoring with a forward-then-backward search trajectory helps account for context-dependent component interactions while reducing the number of exact evaluation trials compared to static rankings.

The writing is clear.

Critical weaknesses

Most importantly, the proposed combination of zero ablation and node-level pruning reverts to an early formulation whose limitations are already well-documented in the mechanistic interpretability literature—precisely the limitations that motivated the community's shift toward interchange (corrupted, you mean) patching and edge-level pruning. First, zero ablation severely ruins the activation space and pushes representations out-of-distribution; forcing a component's output to zero does not correspond to rendering a neuron or component functionally "inactive" or neutral. Second, prior foundational works (such as Path Patching and IOI circuit analysis) have long established that complex, fine-grained edge paths are required to characterize true algorithmic mechanisms, whereas coarse node pruning (especially while retaining all 12 MLP layers unconditionally) obscures the underlying causal routing.

Building on this paradigm shift, the authors discard standard evaluation frameworks in favor of a standalone zero-ablation accuracy metric, claiming that subcircuits found by activation patching methods "often have limited ability to perform the task independently" (Section 1 and Appendix F.2). In reality, under standard interchange/corrupt ablation evaluations (e.g., EAP and EAP-IG on the canonical IOI task), existing methods reliably recover over 90% of the model's task performance and faithfulness. Evaluating methods or objectives designed for distribution-preserving patching under harsh zero-ablation conditions—and using a fine-tuned checkpoint that suffers catastrophic forgetting on standard benchmarks—does not substantiate the claim that existing methods fail.

Questions for authors
  1. How do the authors justify that setting a component's output to zero represents the component being "inactive"? If a component is genuinely inactive on a given task, would its natural output ever be an all-zero matrix (particularly in the presence of biases and non-zero baseline activations), and what strict conditions would its input have to satisfy for that to occur?

  2. Consider two attention heads A1,A2A_1, A_2 and an MLP M1M_1. Suppose Mechanism 1 involves M1M_1 reading directly and independently from both A1A_1 and A2A_2 (A1→M1A_1 \to M_1 and A2→M1A_2 \to M_1), whereas Mechanism 2 involves A1A_1 passing information to A2A_2, which then passes it to M1M_1 (A1→A2→M1A_1 \to A_2 \to M_1, with no direct functional path A1→M1A_1 \to M_1). Since A1,A2,A_1, A_2, and M1M_1 must all be retained as nodes in both cases, how can the proposed node-level circuit representation distinguish between these two fundamentally different mechanisms?

Reviewer_WC4qOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a new method for circuit discovery. The method consists of two steps: in the forward step, they start with a circuit that contains only MLPs and, at each iteration, calculate gradients to decide which component to add next. In the backward step, they again calculate gradients and remove the component with the lowest gradient, as long as the loss remains within a certain threshold.

They compare their method with ACDC, EAP, and EAP-IG on GPT-2 Small and an addition task. Their results show that they can find smaller yet faithful circuits compare to other methods.

Critical strengths
  • On the addition task their method was able to find the smallest yet faithful circuit.
  • Their pruning step is very useful. it's able to prune a significant amount of nodes from the circuit. Im not aware to other methods that prune edges after defining the circuit, and it seems to be useful.
Critical weaknesses
  • The main weakness of the paper is the evaluation: they run their experiments only on GPT-2 Small and on the addition task. GPT-2 Small is a weak and old model (they even need to fine-tune it on the addition task so that it can solve it). For this reason, I don't think I can really evaluate their method.
  • Their method always includes all the MLPs by default. They don't really give a good explanation for why they are doing it, and I also think it can lead to an unfair comparison with other methods that don't do that.
  • This method is really slow because it requires many gradient steps . For example, edge-level BiCD runs for 27 hours while EAP runs for 1.28 seconds!
Questions for authors

I would be happy if you could refer to the Weakness section.

Reviewer_JR6TOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Mechanistic interpretability uses circuit discovery to identify the components responsible for a certain task, but the number of components and edges exponentially increase with model size. Thus, the authors propose Bidirectional Circuit Discovery (BiCD), which uses a forward search algorithm to identify the components that preserve the original model's behavior and then applies a backward pruning process to eliminate any selected components that are not necessary. The authors test their method on the addition task using GPT-2 small, comparing with various established circuit discovery method.

Critical strengths
  1. Interesting idea: I believe that most circuit discovery methods only complete the first stage, identifying the important components, without verifying that all the selected components are essential after identification. This approach does address a valid concern.
Critical weaknesses
  1. IOI: The experiments were performed on the addition task rather than the popular IOI task. The benefit of IOI is that it has an established circuit that was human evaluated, which then can be used to evaluate the outputs of BiCD.
  2. Limited evaluation models: The experiments were also only tested on GPT-2.
Questions for authors
  1. It appears that the cross entropy loss of the original model is fairly small. Does that affect the gradient computation in the first stage? What if the cross entropy loss was only computed between possible answers rather than the entire distribution?
  2. Zero ablation is known to greatly disrupt the model's computation. How does the method perform if you use counterfactual ablation or mean ablation?
  3. Looking at Table 7, the best performing BiCD is component-level, but the number of induced edges is significantly greater than the baselines. At a comparable number of edges, BiCD performs worse than the baselines (i.e. edge-level BiCD vs ACDC). How does BiCD perform if you keep the number of edges fixed, so EAP or EAP-IG at 2,440 edges?

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

Related papers

Discussion (0)

Sign in to comment.