ROAST: Rollout-based On-distribution Activation Steering Technique

Activation steering provides parameter-efficient control over large language models (LLMs) at inference time, but many methods rely on off-distribution supervision and discrete masking, leading to brittle interventions. We propose ROAST (Rollout-based On-distribution Activation Steering Technique), which estimates steering directions from the model's own on-distribution rollouts via ROC and avoids hard sparsification via Continuous Soft Scaling (CSS) and Grouped Mean Normalization. Our empirical analysis reveals that while activation magnitude correlates moderately with directional consistency, the variance in magnitude is significant and often disproportionate to semantic quality. This suggests that high-magnitude activations risk dominating the global steering direction if not properly normalized. To address this, ROAST employs grouped normalization to balance contributions across samples, ensuring a more robust estimation of the consensus steering direction. Across models (0.6B to 32B), ROAST consistently improves performance on diverse tasks (e.g., +9.7% on GSM8K for Qwen3-0.6B and +12.1% on TruthfulQA for GLM4-32B), and analyses show that CSS better preserves activation energy.

Mistake Notebook Learning: Selective Batch-Wise Context Optimization for In-Context Learning

Large language models (LLMs) adapt to tasks via gradient fine-tuning (heavy computation, catastrophic forgetting) or In-Context Learning (ICL: low robustness, poor mistake learning). To fix this, we introduce Mistake Notebook Learning (MNL), a training-free framework with a persistent knowledge base of abstracted error patterns. Unlike prior instance/single-trajectory memory methods, MNL uses batch-wise error abstraction: it extracts generalizable guidance from multiple failures, stores insights in a dynamic notebook, and retains only baseline-outperforming guidance via hold-out validation (ensuring monotonic improvement). We show MNL nearly matches Supervised Fine-Tuning (93.9% vs 94.3% on GSM8K) and outperforms training-free alternatives on GSM8K, Spider, AIME, and KaggleDBQA. On KaggleDBQA (Qwen3-8B), MNL hits 28% accuracy (47% relative gain), outperforming Memento (15.1%) and Training-Free GRPO (22.1) - proving it's a strong training-free alternative for complex reasoning.