When interpretability researchers search a language model for the "circuit" behind a behavior, meaning the small set of attention heads and MLP blocks that does the work, they report one answer. Language models, however, often carry backup machinery. If two heads can each move the same information, deleting either one leaves behavior intact, so a search that prunes components one at a time keeps one head, drops the other, and reports a circuit that is sufficient but incomplete. Tabular machine learning has studied the same phenomenon under another name. The Rashomon effect is the existence of many different models, or feature subsets, that fit the data almost equally well. That field now has algorithms that list all near-optimal decision trees at a cost that is polynomial per tree listed, rather than exponential overall.
We propose to bring that enumeration machinery into circuit discovery. For a given behavior, a threshold $\epsilon$, and a fixed budget of model evaluations, the method outputs the whole family of minimal circuits that reproduce the behavior within $\epsilon$, not one circuit. That family is then summarized into two parts: a core of components that appears in every member, and interchangeable groups that act as logical ORs. The payoff is a description of redundancy that can be tested. It should predict which combinations of knockouts break the model. A single circuit, or a union of circuits, cannot make that prediction.
Core Claim
Fix a model $M$ whose nodes are attention heads and MLP blocks, a task metric $f$ (for example, the logit difference on Indirect Object Identification), and a tolerance $\epsilon$. Let $f(C)$ be the metric when every node outside $C$ is mean-ablated. The circuit Rashomon set is
$$R_\epsilon = \{\, C \subseteq V : f(C) \ge (1-\epsilon)\, f(V) \text{ and no proper subset of } C \text{ satisfies this} \,\},$$
where $V$ is the full node set. The claim has two parts. First, $R_\epsilon$ can be enumerated for models the size of GPT-2 Small on standard tasks, within a model-evaluation budget comparable to running ACDC a few dozen times. Second, the structure of $R_\epsilon$ predicts held-out multi-node ablations better than any single circuit or any union of circuits found by existing methods.
The enumeration rests on a duality. Formal Mechanistic Interpretability: Automated Circuit Discovery with Provable Guarantees uses minimum hitting set duality to compute one certified minimal circuit on small vision networks. A blocking set is a minimal group of nodes whose joint ablation breaks the behavior. Every faithful circuit must contain at least one node from every blocking set, so, when faithfulness is monotone in $C$, minimal circuits are exactly the minimal hitting sets of the blocking sets. We reuse this duality to enumerate circuits rather than to find a single one. We run it on transformers with sampled ablations rather than formal bounds. The interval-bound verification in that paper is intractable for transformers, so we give up the certificate in exchange for scale.
Relation to Prior Work
Three papers motivate the gap. Towards Automated Circuit Discovery for Mechanistic Interpretability (ACDC) prunes edges greedily, one at a time, against a KL-divergence threshold. Attribution Patching Outperforms Automated Circuit Discovery (EAP) scores all edges at once using a gradient approximation. Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates shows that both miss backup (OR) pathways. It fixes this with a combined noising-plus-denoising objective that returns one more complete circuit.
That paper is our closest competitor. It recovers OR-gated components, but it folds them into a single union and does not say which components substitute for which. Our output keeps the alternatives separate and quantifies them.
Two papers constrain the design. The Computational Complexity of Circuit Discovery for Inner Interpretability proves that cardinally-minimal circuit discovery is NP-hard and inapproximable, while quasi-minimal circuits can be found in polynomial time. We therefore target subset-minimal members and an output-sensitive budget, not an optimality guarantee. The formal paper above also proves monotonicity only under its robust patching. With ordinary mean ablation, monotonicity can fail, and we measure how often it does.
The method itself comes from tabular Rashomon enumeration. From Rashomon Theory to PRAXIS: Efficient Decision Tree Rashomon Sets contributes three ideas. It uses a cheap proxy optimizer to bound the search. It moves leftover error budget between sibling subproblems. It hashes subproblems so repeated work is cached. ArborEnum: Decision Tree Rashomon Sets over Continuous Features adds an anytime variant whose partial output stays useful if the search stops early. Neither paper touches neural networks. Our contribution is to translate these ideas into the circuit setting: EAP scores serve as the proxy, the error budget is the $\epsilon$ slack in the metric, and caching is keyed on ablation masks.
Research Program
The algorithm alternates between two steps.
The first step is a blocking-set oracle. Given the blocking sets found so far, it computes a candidate minimal hitting set and tests that circuit. If the circuit is unfaithful, the oracle grows a new blocking set from the complement of the circuit, using binary search over complement nodes ordered by EAP score. This is the $O(\log |G|)$ quasi-minimal routine from the formal-interpretability work, followed by a linear pass that makes the result subset-minimal. If the circuit is faithful, it is added to $R_\epsilon$ and excluded from future candidates.
The second step is budget control borrowed from PRAXIS. EAP scores give optimistic bounds on how much any candidate's completion could recover, and branches whose bound falls below $(1-\epsilon) f(V)$ are cut. Evaluations are cached on 64-bit hashes of the ablation mask. The anytime variant returns circuits ordered by size, so a truncated run still yields the core and the largest OR groups.
We work at node granularity: 144 heads and 12 MLPs in GPT-2 Small. Edge granularity is reserved for a follow-up, because the 32,000-edge Greater-Than graph reported by ACDC is far larger.
From $R_\epsilon$ we extract the core, which is the intersection of all members. We also extract OR groups, defined as maximal sets of nodes that swap one-for-one across members while the rest of the circuit is held fixed. Together these yield a predicted Boolean formula for "behavior survives ablation set $A$."
The study has two settings. The first is Tracr (Tracr: Compiled Transformers as a Laboratory for Interpretability), which compiles human-written programs into exact transformer weights. We compile token counting, sorting, and Dyck programs in which we deliberately duplicate RASP operations, so the true Rashomon set is known by construction. The second setting is GPT-2 Small on IOI, Greater-Than, and Syntactic Agreement. IOI is the natural stress test, because its reverse-engineered circuit already includes backup name-mover heads.
Baselines
The naive reference is a single ACDC run at its default threshold. Its single circuit implies a trivially simple redundancy prediction: behavior breaks if any circuit node is ablated. On IOI, ACDC recovered a 26-head circuit in 8 minutes on an A100, and the manual circuit from Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 Small has 28 heads with 94% faithfulness. Both serve as reference circuits for checking that our core and OR groups contain them.
The strongest existing method for completeness is noising-plus-denoising from the AND/OR/ADDER paper. It explicitly targets OR gates on the same three GPT-2 Small tasks. It reports that ablating its circuits causes larger performance drops than ablating ACDC or EAP circuits, but no reported result found on Lacuna for its numeric ablation drop. We will therefore rerun it ourselves.
The strongest method for compactness is Multi-Granular Node Pruning for Circuit Discovery, which learns Hard Concrete masks, a differentiable relaxation of on/off node gates. It reports "only 21 attention heads, compared to 116 for EAP and 41 for EP" on IOI. EAP itself is a further baseline, with top-$k$ nodes swept over $k$.
The most dangerous baseline is a restart ensemble. We run ACDC, EAP-with-threshold, and Hard Concrete pruning from many random seeds and node orderings. For each method we take the set of distinct circuits found, plus their union and per-node frequencies, as an empirical Rashomon set.
For fairness, every method gets the same mean-ablation scheme, the same prompt sets and held-out split, the same $\epsilon$, and the same total number of forward passes. Each baseline gets an equal hyperparameter sweep, tuned on the training prompts only. Beating them means two things. On Tracr, we need higher recall of ground-truth alternative circuits. On GPT-2 Small, we need higher accuracy at predicting held-out multi-node ablation outcomes, with both differences measured across five seeds of prompt sampling.
One ablation isolates the key ingredient by removing the hitting-set duality and keeping everything else, so that new candidates come from random restarts guided by EAP. A second ablation removes the PRAXIS-style bounds and caching, to measure how much of the budget they save.
Why not just…?
Why not just restart ACDC or Hard Concrete pruning many times and take the distinct circuits? This is the real threat. If redundancy in GPT-2 Small is shallow, with a handful of two-way OR groups, restarts may find most members. Restarts sample circuits in proportion to how easy the optimizer makes them to reach, not how many exist. The duality guarantees that each new candidate differs from all previous ones on some blocking set. The equal-budget comparison on Tracr, where the true set is known, settles this directly. If restarts match our recall on GPT-2 tasks, we will report that circuit multiplicity in small models is low and that cheap restarts suffice. That would itself be a useful negative result.
Why not just use noising-plus-denoising, which already recovers OR-gated components? It returns their union. A union cannot distinguish "heads A and B are both needed" from "either A or B suffices." Those two cases imply opposite outcomes for the double knockout of A and B. The held-out multi-ablation test below targets exactly this distinction.
Why not just read redundancy off EAP scores or single-node ablations? Under an OR gate, each backup head has near-zero individual effect. By construction, single-node and first-order scores are blind to it, and EAP's linear approximation is known to break near nonlinear saturation. We keep EAP only as a pruning proxy and never as the final judge.
Proposed Experiments and Decisive Checks
On Tracr, we compile programs with $r \in \{1,2,3\}$ duplicated pathways. The true Rashomon set then has known size, and we measure set recall in the style of PRAXIS, which reports average recall "never dipping below 0.98" for trees. ACDC is reported to reach AUC 1.0 on unmodified Tracr models. Our question is whether it, or any baseline, still recovers all alternatives once duplication is introduced.
On GPT-2 Small, we report three things for IOI, Greater-Than, and Syntactic Agreement. The first is the size of $R_\epsilon$ for $\epsilon \in \{0.05, 0.1, 0.2\}$. The second is the size of the core and the OR groups, and whether known backup name movers appear as OR partners of the primary name movers. The third is the evaluation budget used.
The decisive check is held-out combinatorial ablation. We sample a few thousand ablation sets of size two to six, biased toward pairs and triples within predicted OR groups. We then compare each method's predicted survive/break label with the true outcome, scored by balanced accuracy. Our method succeeds if it clearly beats the restart ensemble and noising-plus-denoising on joint knockouts, while matching them on single knockouts.
Finally, we report how often mean ablation violates monotonicity, meaning a superset of a faithful circuit turns out unfaithful. A high violation rate would limit how much weight the duality can bear, and we would state that limitation plainly.
References
- Formal Mechanistic Interpretability: Automated Circuit Discovery with Provable Guarantees
- Towards Automated Circuit Discovery for Mechanistic Interpretability
- Attribution Patching Outperforms Automated Circuit Discovery
- Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates
- The Computational Complexity of Circuit Discovery for Inner Interpretability
- From Rashomon Theory to PRAXIS: Efficient Decision Tree Rashomon Sets
- ArborEnum: Decision Tree Rashomon Sets over Continuous Features
- Tracr: Compiled Transformers as a Laboratory for Interpretability
- Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 Small
- Multi-Granular Node Pruning for Circuit Discovery