Large language model agents in social deduction games must reason and speak under hidden roles, shifting public beliefs, and social pressure. Existing methods mainly target final win rate through strategic prompting, belief modeling, or reinforcement learning, but they leave the utterance-level trade-off between belief correction and social risk largely implicit. In false-consensus intervention, useful private information can move the group toward the truth, while poorly timed or overly direct speech can trigger suspicion or rejection. We propose PACR-Pro, a risk-aware rhetorical decision framework for false-consensus intervention. PACR-Pro maps each speaking step to consensus-conflict encoding, structured candidate-action generation, Pareto-based selection, and risk-adaptive utterance realization. A dual-branch Epistemic-Social Critic scores candidate actions, and a state-conditioned 𝜆 policy adjusts the trade-off between epistemic value and social safety. We train these decision modules offline from historical games and evaluate PACR-Pro on One Night Ultimate Werewolf (ONUW) as the test platform. Experiments on standard ONUW games and the Wolf-Pack Induction Benchmark show that PACR-Pro achieves the best mixed-agent win rate, the highest attack win rate, the lowest attack vote rate, and the smallest attack-control win-rate gap among all baselines.

Multi-agent systems enhance complex reasoning but face a critical cost and effectiveness dilemma: API costs grow exponentially while performance degrades due to context degradation and error propagation. This paper proposes an evaluator-guided clause-level pruning and early halting method for lower-cost multi-agent systems named HALT, which implements a macro-to-micro intervention. At the macro-level, a difficulty aware router classifies query complexity, bypassing redundant multi-agent systems overhead for simple tasks. At the micro-level, a lightweight hindsight evaluator monitors generation in real time for complex reasoning, driving a dual-axis intervention policy: spatial pruning removes erroneous clauses, while temporal halting terminates multi-agent interaction early upon reaching a purified consensus. The extensive experiments across three different datasets demonstrate the outperformance of HALT over the baselines, achieving the 97.98%, 90.82%, 87.14% in GSM8K, MMLU, MMLU-Pro in accuracy while drastically mitigating computational overhead, with a maximum reduction of 94.3% in prompt tokens and an unprecedented reduction of 98.9% in completion tokens, which validate the effectiveness of HALT. Code is available at URL.

Reasoning based on chains of thought (CoTs) enables large language models (LLMs) to solve problems by thinking step by step and becomes the mainstream solution for Question-Answering (QA) tasks. Knowledge graph (KG)-enhanced CoT technology helps correct factual errors or predict reasoning direction. Existing KG-enhanced methods find relevant information in KGs “within” each reasoning step of CoTs. However, in some cases, logical connections “between” reasoning steps may be missing or wrong, leading to broken reasoning chains and wrong reasoning direction. To solve the above problem, we argue that the errors between reasoning steps require collaborative verification and mining of multiple triplets and multiple paths in KG. So we propose the DCMKC (Dual Consistency Matching for KG and CoT) method, aiming to maintain semantic and structural consistency between KG and CoT. The main idea is to convert CoTs and KGs into two granularity-aligned graphs, transforming multi-hop reasoning and KG matching into iterative matching and modification of two graphs. In each iteration, DCMKC matches the KG reasoning chains with CoTs based on semantic similarity and judges the structural consistency between them. Then it modifies CoTs using the matched chains. After iterations, the CoTs and KG reasoning chains reach high semantic and structural consistency, which is theoretically and experimentally demonstrated by kernel and spectral methods. The two kinds of chains are then used to generate the final answers. Experimental results show that our method outperforms baselines on multiple datasets, especially on multi-answer questions, with up to 5.1% improvement over the baseline.

Instruction tuning (IT) improves large language models (LLMs) by aligning their outputs with human instructions, but its success depends critically on training data quality, and datasets such as Alpaca often contain noisy or suboptimal examples that undermine fine‐tuning. Prior selection strategies score samples using general‐purpose LLMs (e.g., GPT), leveraging their strong language understanding yet introducing inherent biases that misalign with the target model’s behavior and yield unstable downstream performance. Influence‐based methods address this by estimating each example’s marginal contribution to overall performance, but they typically assume additive contributions and therefore overlook higher‐order interactions among samples. To overcome these limitations, we propose JI²S, a novel framework that jointly models both marginal and combinatorial influences within sample groups. Applying JI²S to select the top 1,000 most influential examples from Alpaca, we fine‐tune LLaMA2‐7B, Mistral‐7B, and LLaMA2‐13B and evaluate them on Open LLM Benchmarks, MT‐Bench, and GPT‐4–judged pairwise comparisons. Our experiments show that JI²S consistently outperforms full‐dataset training and strong baselines, highlighting the value of capturing joint influence for high‐quality instruction fine‐tuning. We provide our code in this GitHub repository.