Rewired or Gated? How Instruction Tuning Shapes Knowledge-Conflict Circuits in LLMs
Shubham Santosh Pandere, Gautam Ranka, Ritika Varshney, Navya Deshmukh, Roushni Sareen, Roshan Kumar Singh
cs.LG · cs.CL
Quick verdict
Claims instruction tuning gates rather than rewires conflict circuits in 3 LLM families; rigorous multi-method case, but only 3-4B models, no chat template.
Summary
The paper asks whether instruction tuning rewires or merely reweights ('gates') the attention-head circuits that arbitrate between contextual and parametric knowledge in LLMs. Comparing base and instruct variants of Llama-3.2-3B, Qwen-2.5-3B, and Gemma-3-4B on ParaConflict-derived single-token conflict prompts (764-870 per family), the authors triangulate five methods: logit-derivative node saliency (LDS), EAP-IG edge attribution, a superposition role index, causal ablation, and path patching. All point to gating: top-head Jaccard between base and instruct is 0.60-0.82, edge overlap is lower (0.39-0.61) but never below the pre-registered 0.3 rewiring threshold; roles are stable (2-3 flips per family) and late-layer localization is preserved. Behaviorally, tuning cuts context-following (CRR) under terse substitution conflicts (e.g., Llama 0.675 to 0.263; all p<10^-15), yet under coherent evidential passages both variants follow the false context near ceiling (0.88-1.00), so the added skepticism is framing-specific. Causal ablations match predicted signs in 36/36 cases with random-head controls near zero. Limitations: only 3-4B models, one dataset dominated by Athlete-Sport prompts, no chat templates applied to instruct models, edge analysis restricted to the substitution form, and path-patching cells individually non-significant. Code is released.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Instruction tuning gates rather than rewires the knowledge-conflict circuit: base and instruct variants share top heads (Jaccard 0.60-0.82) with stable superposition roles (2-3 flips per family) and preserved late-layer localization (median layer moves at most 2).
- Tuning shifts conflict behavior parametric-ward: instruct models reject terse counterfactual context far more than base (substitution CRR falls 0.41-0.50 across families, all p < 10^-15).
- The added skepticism is framing-specific: under coherent evidential passages both base and instruct follow the false context near ceiling (0.88-1.00), so tuning is not a defense against well-written misinformation.
- Because the circuit is preserved, interpretability and control tools such as PH3 head pruning calibrated on base models should transfer to their instruct siblings.
- Edge wiring drifts more than nodes (edge J@50 0.39-0.61) but never reaches the rewiring threshold, yielding strict gating for Llama and an ambiguous edge sub-type for Qwen and Gemma.
Strengths
Pre-registered decision rule guards against threshold shopping
Table 1 fixes bands before analysis (overlap >0.6 high, <0.3 low) and explicitly subordinates the edge axis to the node axis, so Qwen/Gemma's edge drift cannot flip the gating verdict — a structural defense against motivated cutoff choice.
Causal validation with a specificity control
36/36 ablation sign matches (binomial p ~ 2^-36, Figure 5), plus Table 10's random-head floor: ten random 5-head sets per variant move CRR by only -0.03 to +0.02 versus -0.31 to -0.83 for LDS-selected context heads, ruling out a generic late-layer-perturbation artifact.
Behavioral dissociation bounds the headline claim
Statistical hygiene unusual for interpretability work
Weaknesses
Instruct models evaluated without chat templates
Dataset narrowness undermines the definite article in 'the conflict circuit'
'Five independent methods' overstates independence
Edge axis incomplete and verdicts left 'ambiguous'
Abstention confound inside CRR
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper poses a binary mechanistic question: does instruction tuning **rewire** the attention-head circuit that arbitrates knowledge conflicts, or merely **gate** (reweight) it? Three claims are decomposed: C1 gating (node overlap 0.60-0.82), C2 a memory-ward behavioral shift whose added skepticism is framing-specific (C2a), and C3 preserved late-layer localization. Evidence comes from three base/instruct pairs (Llama-3.2-3B, Qwen-2.5-3B, Gemma-3-4B) on ParaConflict-derived single-token prompts — 764 for Llama/Qwen, 870 for Gemma — across four relation domains and two conflict forms (terse substitution, coherent passage). The stated payoff is practical: tools like PH3 pruning calibrated on base checkpoints should transfer to deployed instruct siblings (Section 5).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does 'gating, not rewiring' mean in this paper?
The authors find that after instruction tuning, the same attention heads still carry the knowledge-conflict decision: their identities, memory-vs-context roles, and late-layer positions are preserved, and only their importance (gain) is reweighted. Top-head overlap between base and instruct models is 0.60-0.82 Jaccard, far above the pre-registered 0.3 threshold that would indicate rewiring.
Do instruction-tuned models follow the user's prompt less?
Only for terse, unsupported false statements. Instruct models reject a one-sentence counterfactual far more than base models (context-following drops from 0.65-0.87 to 0.22-0.37), but when the same false claim is written as a coherent evidential passage, both follow it near ceiling (0.88-1.00). The skepticism is specific to terse framing, not a general drop in context-following.
Which models were studied and how?
Base and instruction-tuned pairs of Llama-3.2-3B, Qwen-2.5-3B, and Gemma-3-4B, analyzed without any training using TransformerLens on ParaConflict-derived single-token conflict prompts. Llama is classified as 'strict gating'; Qwen and Gemma as 'gating with an ambiguous edge sub-type.' No chat template was applied to the instruct models.
Is the code and data available for this paper?
Yes. The paper links a public repository (github.com/IvLabs/Rewired-or-Gated), and all experiments use publicly released model weights plus the public ParaConflict dataset, so results are in principle reproducible with GPU access.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are accessible only via captions and in-text numbers; per-head layer assignments and scatter details (Figures 3, 5, 6) could not be inspected directly.
- arXiv v1 preprint, apparently not yet peer-reviewed; the 'first mechanistic base-vs-instruct comparison' priority claim was not independently verified.
- Appendix content is included in the extracted text, but any supplementary material beyond it (e.g., per-domain breakdowns of ablation effects) may be missing.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.