Research map / Large language models / Language, safety and interpretability
The Logical Implication Steering Method for Conditional Interventions on Transformer Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LIMS, which inserts a sensing-gated steering circuit into a transformer's attention output projection so that a concept P triggers behavior Q, built from contrastive mean-difference concept vectors using about 100 examples.
Problem
Models often fail to translate internal recognition of a concept (e.g., hallucination, insufficient information, toxicity, math) into the desired conditional generation behavior, and fine-tuning or prompting is opaque and data-hungry for such conditional logic.
Approach
Extract a sensing concept vector p and a steering vector q from contrastive activation means, then add a gated circuit q·sigmoid(p^T h(x) - b) to the attention output projection, so behavior Q is steered only when P is sensed; the mergeable variant folds this into W as W + q p^T.
Key results
- LIMS reaches 83.0% on HaluEval and 85.0% on AdvBench versus 53.3% and 61.7% for the base model; m-LIMS reaches 94.8% on AdvBench.
- On SQuAD 2, LIMS with 500 examples (81.4%) outperforms DPO trained on 20,000 examples (81.3%).
- LIMS keeps or improves MT-Bench (7.30 base vs 7.30 to 7.43 LIMS variants), while DPO drops to 6.72 on HaluEval-tuned models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add conditional LIMS steering | Mistral 7B Instruct v0.2 · HaluEval | accuracy (%): 53.3 -> 83.0 | Helped |
| Add conditional LIMS steering | Mistral 7B Instruct v0.2 · SQuAD 2 | accuracy (%): 61.8 -> 79.6 | Helped |
| Add conditional LIMS steering | Mistral 7B Instruct v0.2 · AdvBench | accuracy (%): 61.7 -> 85.0 | Helped |
| Add conditional LIMS steering | Mistral 7B Instruct v0.2 · GSM8K test | accuracy (% of COT-prompt performance): 55.5 -> 72.2 | Helped |
| Merge LIMS into model parameters as m-LIMS | LIMS · AdvBench | accuracy (%): 85.0 -> 94.8 | Helped |
| Merge LIMS into model parameters as m-LIMS | LIMS · SQuAD 2 | accuracy (%): 79.6 -> 79.9 | No clear effect |
| Merge LIMS into model parameters as m-LIMS | LIMS · GSM8K test | accuracy (% of COT-prompt performance): 72.2 -> 76.8 | Helped |
| Use 10-shot prompting instead of LIMS | Mistral 7B Instruct v0.2 · HaluEval | accuracy (%): 53.3 -> 50.5 | Hurt |
| Use DPO fine-tuning instead of LIMS | Mistral 7B Instruct v0.2 · SQuAD 2 | accuracy (%): 61.8 -> 61.9 | No clear effect |
| Combine task circuits in one multitask LIMS model (ablation) | single-task LIMS · HaluEval | accuracy (%): 83.0 -> 81.3 | Hurt |
| Combine task circuits in one multitask LIMS model (ablation) | single-task LIMS · SQuAD 2 | accuracy (%): 79.6 -> 78.5 | Hurt |
| Combine task circuits in one multitask LIMS model (ablation) | single-task LIMS · AdvBench | accuracy (%): 85.0 -> 85.0 | No clear effect |
Surprising findings
- DPO-tuned models with best task accuracy regress on open-ended MT-Bench generation, while LIMS does not.
- Steering vectors extracted from HaluEval sensing reach 85% accuracy from answer length alone, suggesting shortcut features.
Limitations
- Circuit inserted only at a single layer (17 of 32) and last token position; layer choice not optimized.
- Hallucination sensing is harder, and SQuAD 2 sensing spills over into earlier tokens, requiring careful alpha tuning.
- Only direct implications and two-sided implications are tested; more complex logic is left to future work.
Experimental setup
Datasets: HaluEval SQuAD 2 AdvBench Alpaca GSM8K
Benchmarks: HaluEval SQuAD 2 AdvBench GSM8K MT-Bench GSM-Symbolic
Metrics: accuracy MT-Bench score average tokens
Baselines: 10-shot prompting DPO Base model COT prompt GPT-4o
Builds on: linear representation hypothesis activation steering DPO RepE
Closest papers by meaning
- When can transformers reason with abstract symbols? (ICLR 2024 · similarity 0.86)
- Are Transformers Able to Reason by Connecting Separated Knowledge in Training Data? (ICLR 2025 · similarity 0.86)
- Grokking of Implicit Reasoning in Transformers: A Mechanistic Journey to the Edge of Generalization (NeurIPS 2024 · similarity 0.85)
- Transformers Provably Solve Parity Efficiently with Chain of Thought (ICLR 2025 · similarity 0.85)
- The Expressive Power of Transformers with Chain of Thought (ICLR 2024 · similarity 0.84)
- Beyond Components: Singular Vector-Based Interpretability of Transformer Circuits (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.