Key Path Identification for Resolving Knowledge Conflicts via SAE-based Steering
Wenbo Zhang, Zhongxiang Sun, Zhiguang Han, Jun Xu
cs.AI
Quick verdict
KPI picks few key SAE feature paths for conflict steering; consistent gains over mass steering across 3 models, but stats thin and causal framing overreaches.
Summary
The paper addresses knowledge conflicts in RAG, where LLMs ignore retrieved context in favor of parametric knowledge. Existing SAE-based steering methods (STA, SPARE) steer large batches of features selected by correlation (mutual information); the authors show empirically that most such features are redundant, and a few steered individually can rival or beat mass steering (Section 3, Figure 3). Their Key Path Identification (KPI) steers each candidate feature singly on a 10-example development set, records which features' activations rise most in downstream layers, builds a directed interaction graph, ranks features by indegree, locates a key layer via the best single-steering feature, and steers only top-ranked features in that layer and preceding layers. Across Gemma-2-9B, Llama-3.1-8B and Llama-3-8B with 131K-wide residual SAEs, KPI wins all six model-dataset settings on NQ-Swap and Macnoise (e.g., 75.83 vs 43.00 for SPARE on Gemma-2-9B/Macnoise), leads on FaithEval-counterfactual, and causes smaller side effects on faithful examples, MMLU and GSM8K. Limitations: 200-example test sets with unexplained std and no significance tests; perturbation-based edges make the causal framing heuristic; per-model hyperparameter tuning; no usable SAEs above 70B. A code URL is provided.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 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
- SAE-based mass steering contains many redundant features; steering a small subset achieves comparable or better conflict-resolution performance (Section 3, Figure 3).
- KPI identifies key features via indegree in a feature-interaction graph and steers key paths (key layer plus preceding layers), winning all six model-dataset settings in Table 1 with an 18% average improvement over the best mass-steering baseline.
- KPI's compact feature set reduces side effects on faithful examples and preserves MMLU/GSM8K better than SPARE (Tables 2, 6-8).
- Steering works by enhancing attention on golden answer tokens, and knowledge-selection behavior forms gradually across a range of layers (Figures 6-10).
Strengths
Clean motivating diagnosis of feature redundancy
Section 3 / Figure 3: on layers 30-31 of Gemma-2-9B (200-example dev set), single-feature steering of a few features beats mass steering over all positive features, and mutual-information rank poorly predicts steering effect — a concrete, actionable failure mode of correlation-based selection (SPARE's Eq. 2).
Consistent directional wins with multi-angle validation
Table 1: KPI best in all six model-dataset cells (e.g., Gemma-2-9B/Macnoise 75.83±5.62 vs SPARE 43.00±6.54); Table 2 side effects on faithful examples (KPI 95.00-99.17 vs SPARE 76.83-95.33); Table 5 OOD FaithEval; Tables 6-8 MMLU/GSM8K preservation, including the dramatic all-token MMLU setting where SPARE collapses to 22.95-30.09 while KPI holds 60.13-67.12.
Cheap and stable pipeline
Mechanistic corroboration beyond scores
Candid limitations section
Weaknesses
Statistical rigor is thin
The 18% headline needs unpacking
Causal framing outruns the evidence
Partially circular collaboration ablation
Citation error and limited baseline breadth
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets knowledge conflicts in RAG — models preferring parametric knowledge over retrieved context — and proposes Key Path Identification (KPI), a feature-selection layer on top of SAE-based steering. The claims chain: (1) correlation-based mass steering (STA, SPARE) selects many redundant features; (2) only a small causally-connected subset matters, so steering a compact key path through a key layer and its predecessors beats mass steering; (3) KPI improves accuracy by 18% on average over the best mass-steering baseline while reducing side effects. The 18% figure needs unpacking: across Table 1, absolute KPI−SPARE gains average roughly 10 points (4.0, 1.8, 9.8, 8.2, 4.2, and an outlier +32.8 on Gemma-2-9B/Macnoise: 75.83 vs 43.00); 18% holds only as a relative improvement and is inflated by that single cell. All six model-dataset cells do favor KPI.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 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 is Key Path Identification (KPI) in LLM steering?
KPI is a method for resolving knowledge conflicts in retrieval-augmented generation, where models ignore retrieved context in favor of their own parametric knowledge. Instead of steering the many sparse-autoencoder (SAE) features picked by correlation, KPI steers each candidate feature individually, builds a graph of which features boost which downstream features, keeps the most-connected ones (highest indegree) around a key layer, and steers only that small key path — a handful of features instead of dozens.
How much better is KPI than previous SAE steering methods?
On NQ-Swap and Macnoise with Gemma-2-9B, Llama-3.1-8B and Llama-3-8B, KPI scores best in all six model-dataset settings; for example 75.83% versus 43.00% for the SPARE baseline on Gemma-2-9B/Macnoise. The abstract's 18% is an average over these settings relative to the best mass-steering baseline; absolute gains average roughly 10 points and are dominated by that single largest gap.
Does SAE feature steering damage the model's normal abilities?
The paper checks side effects: on faithful (non-conflicting) examples KPI keeps 95.00-99.17% accuracy versus SPARE's 76.83-95.33%, and on MMLU and GSM8K KPI stays close to the unsteered model while SPARE drops more — the authors attribute this to filtering out noisy redundant features. The comparison holds in an extreme setting where steering is applied to all prompt tokens.
Which models, SAEs and datasets does the paper use?
Three open models — Gemma-2-9B (Gemma Scope SAEs), Llama-3.1-8B (Llama Scope) and Llama-3-8B (EleutherAI SAEs) — all with 131K-wide residual-stream SAEs. Conflict datasets are NQ-Swap and Macnoise, with FaithEval-counterfactual for out-of-distribution testing; Qwen3-14B judges answer correctness and test sets contain 200 examples each.
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 available only as OCR'd axis labels and captions in the extraction; visual details (bar heights in Figures 3-10) could not be independently inspected, though every number cited in this review comes from tables or prose present in the text.
- Table headers and cells are flattened in the extraction; column assignments (e.g., Table 1's NQ-Swap vs Macnoise halves) were cross-checked against in-prose statements such as SPARE's low 43.00 on Gemma-2-9B/Macnoise before being cited.
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.