Refusal on harmful prompts is a key safety behaviour in instruction-tuned large language models (LLMs), yet the internal causes of this behaviour remain poorly understood. We study two public instruction tuned models-Gemma-22B-IT and LLaMA-3.1-8B-IT using sparse autoencoders (SAEs) trained on residual-stream activations. Given a harmful prompt, we search the SAE latent space for feature sets whose ablation flips the model from refusal to compliance, demonstrating causal influence and creating a jailbreak. Our search proceeds in three stages: 1. Refusal Direction - Finding a refusal mediating direction and collecting SAE features close to that direction, followed by 2. Greedy Filtering - to prune this set to obtain a minimal set and finally 3. Interaction Discovery - a factorization-machine (FM) model that captures non-linear interactions among the remaining active features and the minimal set. This pipeline yields a broad set of jailbreak-critical features, offering insight into the mechanistic basis of refusal. Moreover, we also find evidence of redundant features which remain dormant unless earlier features are suppressed. Our findings highlight the potential for fine-grained auditing and targeted intervention in safety behaviours by manipulating the interpretable latent space. Code- https://github.com/Social- AI-Studio/LLM refusal interp Extended version- https://arxiv.org/abs/2509.09708
Large language models (LLMs) have shown strong potential in automating the design of agentic workflows. However, existing methods still rely heavily on manually predefined operators, limiting generalization and scalability. To address this issue, we propose A(2)Flow, a fully automated framework for agentic workflow generation based on self-adaptive abstraction operators. A(2)Flow employs a three-stage operator extraction process: 1) Case-based Initial Operator Generation: leveraging expert demonstrations and LLM reasoning to generate case-specific operators; 2) Operator Clustering and Preliminary Abstraction: grouping similar operators across tasks to form preliminary abstractions; and 3) Deep Extraction for Abstract Execution Operators: applying long chain-of-thought prompting and multi-path reasoning to derive compact and generalizable execution operators. These operators serve as reusable building blocks for workflow construction without manual predefinition. Furthermore, we enhance node-level workflow search with an operator memory mechanism, which retains historical outputs to enrich context and improve decision-making. Experiments on general and embodied benchmarks show that A(2)Flow achieves a 2.4% and 19.3% average performance improvement and reduces resource usage by 37% over state-of-the-art baselines.
Recent years have witnessed growing scholarly interest in binary post-training quantization (PTQ) techniques for large language models (LLMs). While state-of-the-art (SOTA) binary quantization methods significantly reduce memory footprint and computational demands, they introduce additional memory overhead beyond binary weight tensors to mitigate performance degradation. Moreover, binary LLMs still suffer from substantial accuracy loss. To address these limitations, we propose MemeBQ, a novel binary PTQ framework for LLMs that reduces the memory overhead of auxiliary flag bitmaps in existing binary quantization methods. Specifically, we first design a greedy row clustering method, which leverages the similarity between the row vectors of weights to partition the weight rows into different groups. By sharing the common flag bitmap within each row group, we significantly mitigate the memory overhead associated with flag bitmaps. Besides, to improve the performance of binary LLMs, we propose a novel weight splitting method for each row group of weights, which determines the flag bitmap's values in a fine-grained way. Extensive experiments on OPT, Llama-2, and Llama-3 models demonstrate that MemeBQ reduces 50% extra memory demand while achieving comparable accuracy compared with current SOTA methods. Alternatively, MemeBQ outperforms SOTA binary quantization methods up to 7% with the same extra bits on reasoning benchmarks.
Exploration in sparse-reward tasks remains a fundamental challenge in multi-agent reinforcement learning (MARL) due to complex inter-agent interactions and the expansive exploration space. To address this issue, we propose Targeted Multi-Agent Exploration (TMAE), a novel framework that uncovers the causal relationships between the state space and the reward function, thereby reducing the exploration space and enabling more targeted exploration. Specifically, we construct a structural causal model (SCM) to model the causality between sub-state variables and sparse rewards, providing a robust analytical foundation for subsequent causal inference. Through counterfactual causal intervention, TMAE identifies the most critical subspaces for discovering rare but pivotal events while filtering out confounders. By incorporating these causal insights into the exploration process, TMAE prioritizes subspaces with stronger causal effects on sparse rewards, significantly enhancing exploration efficiency. We evaluate TMAE on a range of MARL benchmarks featuring sparse rewards, consistently demonstrating superior exploration efficiency compared to state-of-the-art methods. Furthermore, visualized causal insights derived from TMAE reveal its ability to effectively capture intricate dependencies and priorities in targeted exploration, showcasing strong alignment with prior domain knowledge.
