Skip to content

Adaptive Knowledge Routing for Complex Reasoning in LLMs

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/hzt12345/normative-circuits
Area: LLM Reasoning
Keywords: normative reasoning, mechanistic interpretability, path patching, sparse circuits, attention head knockout and amplification

TL;DR

Addressing the confound between substantive normative evaluation and generic multiple-choice answer-formatting machinery, this paper proposes a controlled locate–disentangle–intervene pipeline with non-normative MMLU-STEM controls that isolates 5–10 late-layer normative-specific (NS) attention heads (<1% of total) across Qwen3-8B, Llama3.1-8B, and Mistral-7B-v0.1, establishing bidirectional causal effects via knockout and amplification.

Background & Motivation

When a large language model determines that discharging industrial toxic waste into a public waterway is illegal or that finding a lost wallet requires returning it intact, where in the underlying neural network is that evaluative judgment computed? Existing legal and ethical benchmarks—such as LegalBench, LEXam, SafetyBench, and MoralChoice—evaluate model competence almost exclusively through macroscopic task accuracy. However, behavioral accuracy cannot reveal whether the underlying moral or legal reasoning is implemented by a localized, sparse circuit or diffusely distributed across hundreds of billions of model parameters. If normative computation is localized, practitioners can execute precision inference-time interventions to enforce ethical boundaries without degrading broader language capabilities; conversely, if it is completely diffuse, targeted model alignment and steerability remain fundamentally difficult.

Mechanistic interpretability investigations into normative judgment have faced severe methodological confounds. Standard normative benchmarks are predominantly formatted as multiple-choice question answering (MCQA). Consequently, model representations must simultaneously execute substantive normative evaluations and drive generic task infrastructure—such as recognizing answer options, aggregating contextual cues, and predicting terminal answer letter tokens (A/B/C/D). Naive path patching on normative benchmarks inevitably entangles these format-processing heads with genuine normative reasoning, leading prior studies to misclassify universal MCQA machinery as domain-specific reasoning components. Furthermore, existing mechanistic work on moral judgment (such as studies reporting the attenuation of opposed moral categories across intermediate layers, or observing zero behavioral flips across tens of thousands of single-component activation swaps) has not pinpointed cross-task, causal normative subcircuits across model families.

To resolve the tension between format-processing infrastructure and genuine normative evaluation, this paper develops a controlled circuit identification framework. By analyzing four comprehensive legal and moral benchmarks across both Chinese and English (totaling 5,919 samples per model) and pairing them with a controlled set of 500 non-normative MMLU-STEM questions, the authors employ set-difference filtering to subtract generic answer-formatting machinery before running causal interventions. Core idea: by pairing normative benchmarks with same-format non-normative MCQA controls and applying set-difference filtering, this work isolates 5–10 late-layer normative-specific (NS) attention heads (<1% of total) in 7–8B LLMs, proving via bidirectional knockout and amplification that normative judgment is driven by sparse attention routing of candidate options rather than progressive MLP-based computation.

Method

Overall Architecture

The paper introduces a three-stage "Locate–Disentangle–Intervene" pipeline designed to isolate and validate causal normative circuits while systematically stripping away generic multiple-choice formatting artifacts. The pipeline takes as input bilingual normative MCQA datasets (Chinese and English legal and moral questions) along with a non-normative STEM MCQA control corpus, and outputs a validated, sparse set of Normative-Specific (NS) attention heads along with their causal bidirectional characterization.

In the Locate stage, Attribution Patching (AtP) is applied across all attention heads using counterfactual prompts with masked domain terms, ranking heads by their first-order gradient-approximated causal effect on the correct answer logit. In the Disentangle stage, the top-20 attention heads from normative tasks are compared against the top-20 heads from a 500-sample non-normative MMLU-STEM control task using set-difference filtering, discarding general-purpose MCQA infrastructure heads and retaining only those heads that recur across multiple normative domains. In the Intervene stage, bidirectional causal tests—mean-ablation knockout to test necessity and activation amplification to test sufficiency—are executed on the isolated NS head sets under both logit-difference and categorical accuracy metrics.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Bilingual Normative & Counterfactual Prompts<br/>(JEC-QA / LEXam / SafetyBench / MoralChoice)"] --> B["Locate Candidate Heads via Attribution Patching<br/>(AtP Causal Effect Approximation)"]
    B --> C["Disentangle Normative Heads via Set-Difference<br/>(Subtractive Filter vs. MMLU-STEM Controls)"]
    C --> D["Bidirectional Causal Intervention & Attention Analysis<br/>(Knockout for Necessity & Amplification for Sufficiency)"]
    D --> E["Cross-Model Architectural Convergence Analysis<br/>(Evaluation across Qwen3, Llama3.1, and Mistral)"]

Key Designs

1. Locate Candidate Heads via Attribution Patching: Overcoming the computational intractability of full combinatorial path patching

Standard activation path patching requires replacing activations for each individual component (amounting to \(36 \times 32 = 1,152\) attention heads in Qwen3-8B) followed by a full forward pass, incurring prohibitive GPU-hour costs when scaling across thousands of multi-step prompts. To render network-wide localization feasible, the authors implement Attribution Patching (AtP), a first-order Taylor series approximation that replaces exhaustive forward passes with two forward passes and one backward pass. Domain-specific counterfactuals are constructed by identifying key domain terms (via curated lists of 74 Chinese legal, 50 English legal, 36 Chinese moral, and 28 English moral terms) and masking 30%–50% of them with bold single-letter placeholders (such as V or E), destroying semantic normative meaning while preserving syntactic token length. At the final token position (END), the causal attribution score \(\hat{E}_{l,h}\) for head \((l,h)\) is computed with respect to the correct answer log-probability \(L = \log P(y \mid X_r)\):

\[\hat{E}_{l,h} = \left( A_c(l, h) - A_r(l, h) \right) \cdot \left. \nabla_{A(l,h)} L \right|_{A=A_r}\]

where \(A_r\) and \(A_c\) denote activations from the reference and counterfactual runs respectively. Filtering for heads with \(|\hat{E}_{l,h}| > 0.20\) reveals remarkable sparsity (<5% of heads) and strong spatial concentration in late layers (layers 22–35 in Qwen3-8B).

2. Disentangle Normative Heads via Set-Difference: Eliminating false-positive attribution caused by generic MCQA task machinery

A critical conceptual vulnerability in naive circuit analysis is that attention heads highlighted by attribution patching frequently govern generic multiple-choice infrastructure (e.g., matching letter labels A–D to option positions) rather than substantive ethical or legal judgment. To untangle this confound, the authors introduce a controlled set-difference filter using 500 samples from MMLU-STEM as a non-normative, same-format control. Empirical measurement reveals that the top-20 heads identified on normative tasks exhibit an average Jaccard similarity of 0.44 with the top-20 heads of the general MCQA control, confirming that nearly half of the candidate circuit consists of format-handling machinery.

The authors formally define the Normative-Specific (NS) head set as those attention heads appearing in the top-20 of at least two normative datasets while being strictly absent from the general-MCQA top-20. This subtractive filter contracts the candidate substrate to just 5–10 heads per model: 7 heads in Qwen3-8B, 5 in Llama3.1-8B, and 8 in Mistral-7B-v0.1. This represents a sparse subnetwork occupying merely 0.5%–0.8% of the total attention capacity, proving that format disentanglement is an essential prerequisite for interpreting task-specific reasoning.

3. Bidirectional Causal Intervention & Attention Analysis: Proving necessity and sufficiency while isolating the routing mechanism

To ensure that the identified NS heads are causally operative rather than correlated byproducts, the framework applies bidirectional interventions. For necessity testing, target heads are mean-ablated using position-wide dataset-mean activations drawn from a 100-prompt counterfactual pool, benchmarked against size-matched random-rank head sets across 5 random trials per cell. Effects are assessed via the continuous logit difference metric (\(\text{logit}_y - \max_{y' \neq y} \text{logit}_{y'}\)) to detect subtle margin compressions that discrete accuracy obscures. For sufficiency testing, incorrect baseline predictions (\(N=500\)) are subjected to activation amplification by multiplying NS head outputs by scaling factor \(\alpha = 2.0\):

\[A'_{l,h} = \alpha \cdot A_{l,h}\]

Under amplification, NS heads flip incorrect predictions to correct answers at an advantage ratio of 1.8×–5.4× over random head sets. Furthermore, attention analysis demonstrates that key NS heads (e.g., L35-H25 in Qwen3) focus 6.64% of their attention weight onto candidate answer-option tokens—approximately 5 times the uniform baseline rate. In sharp contrast, multi-position MLP path-patching yields causal effects bounded below \(10^{-3}\) (two to three orders of magnitude smaller than attention heads). This proves that normative reasoning in MCQA settings operates primarily via attention heads routing candidate option representations to the decision position, rather than progressive nonlinear transformation across MLP layers.

4. Cross-Model Architectural Convergence Analysis: Testing cross-family universality versus functional re-identification

To determine whether normative circuits reflect immutable cross-architecture coordinates or emergent structural properties, the entire experimental pipeline is independently replicated across Qwen3-8B, Llama3.1-8B, and Mistral-7B-v0.1. Results show profound structural convergence: across all three families, \(\ge 50\%\) of top-20 normative heads concentrate in the final third of the network layers (62.5% in Qwen3, 53.8% in Llama, 50.0% in Mistral), and within-model Legal-versus-Moral grouped Jaccard overlaps remain stable between 0.31 and 0.44. However, at the microscopic coordinate level, the (layer, head) indices of the NS sets across the three families are completely disjoint. This establishes that normative reasoning circuits do not transfer by fixed numerical indices across architectures; rather, mechanistic transferability requires functional re-identification using standardized disentanglement protocols.

Benchmark & Evaluation Protocol

The evaluation spans four diverse normative corpora representing two major legal systems and distinct ethical dimensions: 1. Chinese Legal (JEC-QA): 1,999 multiple-choice questions from the National Judicial Examination of China covering civil, criminal, and administrative law (4-choice format). 2. English Legal (LEXam): 619 Swiss law examination questions translated and curated across multiple legal domains (4-choice format). 3. Chinese Moral (SafetyBench): 1,934 ethics and morality questions from the SafetyBench evaluation suite (2-to-4 choice format). 4. English Moral (MoralChoice): 1,367 moral dilemma scenarios. Due to substantial original answer-label bias (>95% choice A), the authors construct an unbiased stratified 50/50 A/B subset of \(N=500\) samples for intervention tests. The non-normative control comprises 500 questions from MMLU-STEM. All models evaluate single-token output logits at the final prompt token.

Key Experimental Results

Main Results: Selective NS-Head Knockout (Necessity Testing)

The table below presents mean-ablation knockout results across 15 cells (\(N=500\) per cell, 5 random trials). The primary significance criterion requires \(|\Delta\text{logit-diff}| \ge 0.05\) in the predicted direction with trial-level statistic \(|s_{ld}| \ge 2.0\) (denoted by S; L denotes Bonferroni-corrected logit-diff significance, a denotes Bonferroni-corrected accuracy significance):

Model Dataset NS Count \(K\) Baseline Acc NS Acc Rand Acc \(\Delta\text{acc}\) (pp) \(\Delta\text{logit-diff}\) \(s_{ld}\) Significance
Qwen3-8B Chinese Legal (JEC-QA) 8 0.53 0.494 0.508 -1.4 -0.090 -5.5 S
Qwen3-8B English Legal (LEXam) 6 0.32 0.296 0.316 -2.0 -0.155 -8.3 S
Qwen3-8B Chinese Moral (SafetyBench) 5 0.47 0.472 0.476 -0.4 +0.049 +2.1 —
Qwen3-8B English Moral (MoralChoice) 10 0.64 0.570 0.650 -8.0 -0.337 -16.2 S, L
Qwen3-8B General MCQA (MMLU-STEM) 22 0.48 0.378 0.439 -6.1 -0.320 -6.2 S (Control)
Llama3.1-8B Chinese Legal (JEC-QA) 7 0.52 0.492 0.528 -3.6 +0.006 +0.4 —
Llama3.1-8B English Legal (LEXam) 7 0.26 0.244 0.246 -0.2 +0.132 +10.5 ◦ (Floor bound)
Llama3.1-8B Chinese Moral (SafetyBench) 6 0.28 0.274 0.298 -2.4 -0.083 -2.5 S (Sub-chance)
Llama3.1-8B English Moral (MoralChoice) 10 0.44 0.524 0.439 +8.5 +0.133 +5.6 ◦ (Outlier)
Llama3.1-8B General MCQA (MMLU-STEM) 24 0.26 0.250 0.251 -0.1 +0.011 +0.8 — (Control)
Mistral-7B Chinese Legal (JEC-QA) 9 0.49 0.500 0.496 +0.4 -0.043 -5.6 — (<0.05 mag)
Mistral-7B English Legal (LEXam) 8 0.22 0.218 0.226 -0.8 -0.060 -9.1 S
Mistral-7B Chinese Moral (SafetyBench) 9 0.26 0.308 0.262 +4.6 +0.127 +6.2 ◦ (Floor bound)
Mistral-7B English Moral (MoralChoice) 9 0.40 0.334 0.407 -7.3 -0.460 -14.4 S, a, L
Mistral-7B General MCQA (MMLU-STEM) 25 0.25 0.226 0.249 -2.3 -0.156 -6.5 S (Control)

Ablation & Sufficiency: NS Activation Amplification (\(\alpha = 2.0\))

The table below reports answer recovery rates when amplifying NS heads versus random head sets on \(N=500\) originally-incorrect samples per cell. Bonferroni significance (\(p_{\text{Bonf}} < 0.05\) across 12 cells) is marked with ✓; point-estimate positive cells failing correction are marked with ◦; null cell is marked with ⃝:

Model Metric Chinese Legal (JEC-QA) Chinese Moral (SafetyBench) English Legal (LEXam) English Moral (MoralChoice)
Qwen3-8B NS / Random Recovery Count 90 / 43 (18.0% vs 8.6%) 85 / 38 (17.0% vs 7.6%) 10 / 5 (2.0% vs 1.0%) 65 / 12 (13.0% vs 2.4%)
Qwen3-8B Advantage Ratio (NS / Rand) 2.1× 2.2× 2.0× 5.4×
Qwen3-8B Bonferroni Significance ✓ (\(p=1.4 \times 10^{-4}\)) ✓ (\(p=7.2 \times 10^{-5}\)) ◦ (\(p_{\text{Bonf}}=1.0\)) ✓ (\(p=3.9 \times 10^{-9}\))
Llama3.1-8B NS / Random Recovery Count 45 / 17 (9.0% vs 3.4%) 15 / 5 (3.0% vs 1.0%) 30 / 11 (6.0% vs 2.2%) 0 / 0 (0.0% vs 0.0%)
Llama3.1-8B Advantage Ratio (NS / Rand) 2.6× 3.0× 2.7× 0.0 (Null cell)
Llama3.1-8B Bonferroni Significance ✓ (\(p=0.0029\)) ◦ (\(p_{\text{Bonf}}=0.287\)) ✓ (\(p=0.029\)) N/A
Mistral-7B NS / Random Recovery Count 25 / 14 (5.0% vs 2.8%) 30 / 15 (6.0% vs 3.0%) 15 / 12 (3.0% vs 2.4%) 70 / 21 (14.0% vs 4.2%)
Mistral-7B Advantage Ratio (NS / Rand) 1.8× 2.0× 1.25× 3.3×
Mistral-7B Bonferroni Significance ◦ (\(p_{\text{Bonf}}=0.868\)) ◦ (\(p_{\text{Bonf}}=0.266\)) ◦ (\(p_{\text{Bonf}}=1.0\)) ✓ (\(p=8.6 \times 10^{-7}\))

Key Findings

  • Extreme Sparsity of Normative Components: Set-difference filtering reveals that normative evaluation is mediated by an ultra-sparse subnetwork of only 5–10 attention heads per model (7 in Qwen3, 5 in Llama, 8 in Mistral), constituting less than 1% of total network attention heads.
  • Robust Cross-Model Necessity on English Moral: The necessity of NS heads is most pronounced on English Moral reasoning, exhibiting severe drops of 7.3pp to 8.0pp in accuracy and -0.337 to -0.460 in logit-difference for both Qwen3 and Mistral.
  • Directional Sufficiency Consistency: Amplification of NS heads produces directionally superior recovery of incorrect answers in 11 of 12 cells (binomial sign-test \(p \approx 0.003\)), with advantage ratios reaching 1.8×–5.4× over random head sets.
  • Attention Routing Over MLP Storage: While attention head ablations produce pronounced causal shifts (\(10^{-2}\) to \(10^{-1}\)), multi-position MLP path patching yields negligible effects (\(< 10^{-3}\)), demonstrating that normative judgment in MCQA is executed via dynamic attention routing of candidate options rather than localized MLP-stored factual associations.

Highlights & Insights

  • Standardized Set-Difference Paradigm: Introducing non-normative MCQA controls to subtract shared task infrastructure represents a critical methodological advance for circuit analysis, preventing the misattribution of generic multiple-choice machinery to substantive domain reasoning.
  • Structural Invariance Without Coordinate Sharing: The findings demonstrate that while different model families do not share identical (layer, head) coordinates for normative evaluation, they converge on the same macro-architectural motif: extreme sparsity (<1%) and strong late-layer concentration (>50% in the final third).
  • Practical Alignment Implication: Because normative evaluation is concentrated in a tiny cluster of late-layer attention heads, alignment techniques can target these specific heads for steering or auditing without fine-tuning the entire model or disrupting broad linguistic capabilities.

Limitations & Future Work

  • Format and Generative Limitations: All findings are established within multiple-choice formats (MCQA) with single-token output labels; whether sparse routing circuits govern open-ended generation, legal document drafting, or multi-turn conversational reasoning remains untested.
  • Model Scale and Control Language: Experiments are conducted on 7B–8B parameter base models; verification on larger models (e.g., 70B+) and instruction-tuned variants is needed. Additionally, the non-normative control was restricted to English STEM, requiring future expansion to Chinese non-normative controls.
  • Floor-Bounded Baseline Confounds: Datasets with baselines near chance (e.g., 22%–26% on 4-choice tasks) constrain the observable range of ablation effects and can induce spurious accuracy reversals.
  • vs. Zhang et al. (2024a) Arithmetic Circuit Analysis: Zhang et al. showed that arithmetic reasoning relies heavily on intermediate-layer MLPs for nonlinear numerical computation; this work reveals the opposite mechanism for normative judgment, where MLP effects are two orders of magnitude smaller and attention routing dominates.
  • vs. Wiegreffe et al. (2025) MCQA Infrastructure Analysis: While Wiegreffe et al. characterized how attention heads process multiple-choice formats generally, this work introduces subtractive filtering to cleanly decouple substantive normative evaluation from that format-processing infrastructure.
  • vs. Crosse (2026) Moral Fine-Tuning Interpretability: Crosse observed zero behavioral flips under single-component swaps, concluding moral representations are diffuse; by executing joint interventions on coordinated sparse head sets (5–10 heads), this paper proves that normative reasoning is causally concentrated and steerable.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ [Pioneering set-difference methodology decoupling normative judgment from task-format machinery]
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Comprehensive 3-family, 4-dataset, bilingual replication with AtP, knockout, amplification, and K-sweeps]
  • Writing Quality: ⭐⭐⭐⭐⭐ [Exceptional clarity, rigorous statistical framing, and candid discussion of floor confounds]
  • Value: ⭐⭐⭐⭐⭐ [Crucial foundational insights for mechanistic interpretability, legal AI audits, and targeted model alignment]