Research map / Large language models / Agents, code and math
Mitigating Content Effects on Reasoning in Language Models Through Fine-Grained Activation Steering
Marco Valentino, Geonhee Kim, Dhairya Dalal, Zhixue Zhao, André Freitas
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add static contrastive activation steering (CAA) | Llama 3.2 1b · Syllogistic validity (zero-shot) | accuracy (%): 58.17 -> 73.56 | Improved |
| add static contrastive activation steering (CAA) | Qwen 2.5 7b · Syllogistic validity (zero-shot) | Acc/CE: 16.48 -> 93.65 | Improved |
| add static CAA steering | Llama 3.2 3b and Qwen 2.5 3b · Syllogistic validity (zero-shot) | accuracy (%): 77.79 -> 77.79 (Llama 3b) | No clear effect |
| add static CAA steering | Gemma 2 9b · Syllogistic validity (zero-shot) | accuracy vs content effect: accuracy 85.00 -> 83.27; CE 8.46 -> 1.92 | Mixed |
| add 4-shot in-context examples | Llama 3.1 8b · Syllogistic validity | accuracy (%): 78.27 -> 40.58 | Hurt |
| use K-CAST conditional steering instead of unsteered model | Llama 3.2 3b · Syllogistic validity (zero-shot) | accuracy (%): 77.79 -> 92.60 | Improved |
| use kNN-based condition (K-CAST) instead of aggregated condition vectors (CAST) (ablation) | Llama 3.2 3b · Syllogistic validity (zero-shot) | accuracy (%): 81.04 -> 92.60 | Improved |
| use K-CAST instead of CAST (ablation) | Qwen 2.5 3b · Syllogistic validity (zero-shot) | accuracy (%): 85.86 -> 85.28 | No clear effect |
| apply content-effect steering | Llama 3.2 1b · C4 English perplexity | perplexity: 24.29 -> 24.77 | No clear effect |
| apply content-effect steering | Gemma 2 9b · C4 English perplexity | perplexity: 18.95 -> 20.58 | Hurt |
| apply steering vectors from syllogisms | Llama 3.2 1b · ProntoQA (OOD) | accuracy: 49.6 -> 53.6 | Improved |
| apply steering vectors from syllogisms | Gemma 2 9b · ProntoQA (OOD) | accuracy: 62.2 -> 52.2 | Hurt |
Closest papers by meaning
- Calibrating Reasoning in Language Models with Internal Consistency (NeurIPS 2024 · similarity 0.90)
- ActivationReasoning: Logical Reasoning in Latent Activation Spaces (ICLR 2026 · similarity 0.89)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.89)
- AdaMCoT: Rethinking Cross-Lingual Factual Reasoning Through Adaptive Multilingual Chain-of-Thought (AAAI 2026 · similarity 0.89)
- The Tell-Tale Norm: $\ell_2$ Magnitude as a Signal for Reasoning Dynamics in Large Language Models (ICML 2026 · similarity 0.89)
- Longer Context, Deeper Thinking: Uncovering the Role of Long-Context Ability in Reasoning (NeurIPS 2025 · similarity 0.89)
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.
