今日速览
| 序号 | 标题 | 来源 | 日期 | 主题 | 推荐等级 |
|---|---|---|---|---|---|
| 1 | Do “True” and “False” Add Up to One? Complement Consistency in LLM-Agent Opinion Simulations | arXiv | 2026-10-23 | RAG | 高 |
| 2 | J-Comm: Jacobian-Guided Latent Communication between LLM Agents | arXiv | 2026-10-23 | RAG | 高 |
重点论文与技术动态
1. Do “True” and “False” Add Up to One? Complement Consistency in LLM-Agent Opinion Simulations
- 来源：arXiv
- 日期：2026-10-23
- 作者/机构：Anouar Imel, Jin Qi
- 主题标签：
RAG,arXiv - 推荐等级：高
- 分类：cs.AI, cs.CL, cs.MA, cs.SI
一句话结论
LLM智能体在观点模拟中违反互补一致性（正反命题概率和大于1），说明其模拟概率不能直接等同于内在信念。
核心内容
- 三款模型对互补命题的概率和均超1（Qwen达1.18），偏差比同义改写高2至10倍且在多种条件下持续存在。
- 互补初始状态的智能体群体迭代5轮后均出现正向发散，无法维持镜像演化。
- 与补集答案求均值虽可机械恢复一致性并降低平均Brier损失，但终轮仅对1款模型优于单侧最佳表述。
方法与数据
- 评估Qwen3-4B、Llama-3.2-3B及Gemma-3-4B三款小模型，基于2,304个模拟网络状态、48条AVeriTeC命题与24个双结果基准世界。
价值判断
- 值得关注：揭示了LLM概率输出受命题表述影响且违背概率公理的缺陷，提示 simulated probabilities 需先行验证一致性。
- 可复用点：互补改写一致性检验流程及对偶均值化校正策略。
- 局限/待核查：仅评测了小参数开源模型；均值化修正大多未在终轮超越较优的原始表述。
摘要
In LLM-agent opinion simulations, each agent reports the probability that a claim is true after seeing its own previous probability and the probabilities reported by its neighbors, and that number becomes its state in the next round. If the number describes the agent's position rather than the sentence it was shown, then rewriting the claim as its complement and replacing every probability p with 1 - p should return the complementary answer, so the two reports should sum to one. We test this with three small open models: Qwen3-4B, Llama-3.2-3B and Gemma-3-4B. On 2,304 agent states taken from a running network simulation, Qwen's paired reports sum to 1.18 on average, and on 48 real fact-checking claims from AVeriTeC they again exceed one. A controlled benchmark of 24 worlds, each with exactly two possible outcomes, shows the same failure in all three models. The discrepancy is two to ten times larger than the change produced by a same-meaning paraphrase, and it persists under repeated sampling and across peer group sizes and report strengths. When the reports are fed back into two populations that start from complementary states, the populations do not remain mirror images: their divergence is positive after five rounds for every model. Averaging each answer with the complement of its paired answer restores consistency by construction and lowers Brier loss averaged over the two wordings, but at the final round it beats the better of the two original wordings for only one of the three models. A schedule-by-targeting experiment on the same model gives a small, stable interaction because such sweeps keep the claim wording fixed. Consistency under rewording therefore needs its own test before simulated probabilities are read as beliefs2. J-Comm: Jacobian-Guided Latent Communication between LLM Agents
- 来源：arXiv
- 日期：2026-10-23
- 作者/机构：Xiangwei Wang, Wei Wang, Sachith Seneviratne, Ken Chen, Yu Xia, Peng Wang
- 主题标签：
RAG,arXiv - 推荐等级：高
- 分类：cs.CL, cs.MA
一句话结论
J-Comm提出免训练几何共识与H+J隐式通信接口，实现了无需共享KV缓存或匹配模型宽度的跨LLM智能体高效协同推理。
核心内容
- 提出免训练的GEAR机制，通过重加权词元在嵌入空间单调优化几何共识以增强隐式推理。
- 构建H+J双通道接口：H通道提供反馈分布的连续输入，J通道以雅可比透镜坐标传递中间层计算引导。
- 接收端利用自身词典重构线索并注入预填充阶段的中间层，较TextMAS减少了19.2%-74.8%的生成token。
方法与数据
- 在Qwen3-8B向Qwen3-4B通信场景下评测，覆盖MATH500、GPQA-Diamond、AIME及ALFWorld等7个基准。
价值判断
- 值得关注：突破了异构模型宽度及KV缓存共享限制，均分表现优异，性能超TextMAS 1.29-8.59个百分点。
- 可复用点：免训练的GEAR候选词元重加权算法，以及基于雅可比坐标的中间状态特征注入设计。
- 局限/待核查：摘要未明确在非同源系列模型或更大参数差异场景下的跨架构通信泛化效果。