Research Digest — 2026-10-04

Scan window: 2026-09-28 → 2026-10-04 Sources: arXiv (cs.AI / cs.LG / cs.MA / cs.CR / cs.CV / cs.CY), Papers With Code Verification scope: For every paper below I fetched the arXiv abs page and confirmed (a) the ID prefix encodes a September 2026 submission (2609 = 2026-09) and (b) the "Submitted on" line falls inside the scan window. Titles and headline numbers are quoted from the abstracts fetched today. Nothing else is claimed as verified.

Selected (5): 2609.35928, 2609.36642, 2609.34890, 2609.34427, 2609.35706 — all submitted 28–29 Sep 2026, ID dates confirmed.

Discarded:

• ●2402.03578 — LLM Multi-Agent Systems: Challenges and Open Problems — first submitted Feb 2024 (revised Jan 2026); outside the 7-day window.
• ●2503.16416, 2508.04652, 2503.10009 — 2025 IDs, outside window.
• ●Papers With Code scan returned only stale benchmark pages with no verifiable new state-of-the-art claim dated within the window; nothing selected from PwC this week.

1. Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems ⚠ Breakthrough candidate

• ●2609.35928 — Submitted 28 Sep 2026 (verified)
• ●Authors: Xavier Del Giudice, Alessio Palma, Matteo Migliarini, Fabio Galasso, Indro Spinelli
• ●Method: Controlled experiments on multi-agent LLM teams (9–25 agents, up to 5 open-weight model families), two cooperative games + one reasoning benchmark. Manipulated variable: whether each agent's underlying model identity is visible to peers.
• ●Findings: Exposing model identity causes "factionalism" — agents preferentially interact with same-label peers although the task neither rewards nor requests it. Labeled groups spend on average 30% more rounds and 55% more tokens to reach a decision; success drops from 96% to 81%. Shuffling or replacing labels with arbitrary ones moves the factions accordingly; removing labels removes the behavior. Withholding identity information is the paper's own simple and effective mitigation.
• ●Applicability to LocalKin: High, immediately actionable. Config-level change: stop forwarding model-identity metadata between agents. Cost: low. Measure rounds/tokens/success before-after.
• ●Detailed analysis: output/data_scientist/breakthrough/prompted_identity_factionalism_2026-10-04.md

2. PR-OPD: Privileged Representation On-policy Self-Distillation for Agentic Reinforcement Learning

• ●2609.36642 — Submitted 29 Sep 2026 (verified)
• ●Method: Observation: a "skill" in context lifts WebShop success 42.2% → 56.2% while changing probabilities of fewer than a quarter of sampled tokens, but changes hidden states of over 80% of response tokens (linear probes can trace the skill). Fix: after a GRPO warm start, the policy writes a hindsight skill per trajectory, re-reads its own response with that skill as a stop-gradient teacher, and aligns hidden states layer-by-layer alongside the reward objective.
• ●Findings: Best overall on ALFWorld and WebShop across two backbones: up to +4.7 pts ALFWorld success, +14.0 pts WebShop accuracy over GRPO — no external skill library, no separate teacher, no inference-time overhead.
• ●Applicability: Medium — benefits us only if we fine-tune backbones; the hindsight self-teaching pattern can be approximated at inference time cheaply. Cost: full = high (RL infra); proxy = low.

3. KITA AI: A Multi-Agent LLM System for Pluralistic Policy Deliberation

• ●2609.34890 — Submitted 28 Sep 2026 (verified)
• ●Authors: United Nations University Institute in Macau; AI Singapore; University of Pretoria (first-listed: Arnau Mayoral-Macau, 13 authors total)
• ●Method: Multiple LLM agents, each grounded in distinct demographic stakeholder personas and conceptual frameworks, deliberate over policy scenarios; the system surfaces who is affected, with rationales and quantitative indicators per positional cluster.
• ●Key insight: Treats non-convergence as a first-class explainable output, not a failure mode — decision-makers get the trade-off map, not forced consensus.
• ●Applicability to LocalKin: High. Add a "non-convergence report" output mode to swarm debates (position clusters + reasons + quantitative deltas). Cost: low–medium (prompt/schema changes only).

4. LLMs as Adaptive Meta-Solvers: Strategy-Diverse RL for Industrial-Scale Optimization

• ●2609.34427 — Submitted 28 Sep 2026 (verified)
• ●Method: SDRL trains open-source LLMs to adaptively choose among three strategies — solver-integrated reasoning, exact combinatorial algorithms, heuristic search — using a correctness-gated hierarchical diversity reward (between-strategy AND within-strategy diversity), plus mixed-format training (self-contained text problems + file-grounded instances).
• ●Findings: Authors claim it outperforms fine-tuned methods and frontier models incl. DeepSeek-V4-Pro and GPT-5.5 on average benchmarks and industrial-scale tasks. [Authors' claim — not independently reproduced.]
• ●Deeper insight: the three strategies are complementary across problem structures and scales — a portfolio argument. Cheaper proxy for us: a routing layer that picks strategy by problem profile (no training required).
• ●Applicability: Medium–high if we route optimization-shaped workloads to a trained meta-solver; full recipe = high cost.

5. Reinforcing Agentic Creativity in Scientific Ideation with Night Science

• ●2609.35706 — Submitted 28 Sep 2026 (verified)
• ●Method: GRPO-based RL conditioned on cognitive-science-grounded creativity axes: action (what and how creatively), process (explore vs. exploit timing), outcome (novelty + usefulness). The paper's abstract names its framework AI Night-Scientist; "night science" in the title denotes loosely structured, serendipitous exploration vs. structured "day science".
• ●Findings: Research-direction range +27.8%, contribution-type range +14.9%, predicted citation impact up to +32.0 pp, originality +66.2 pts. Gains cannot be reproduced by raising decoding temperature; semantic guidance specifying what kind of creativity to pursue is critical.
• ●Applicability: Medium. Full RL recipe = high cost; inference-time lesson is free: vary semantic instructions, not temperature, for ideation diversity. Proxy cost: low.

1. ●Do now (config): Withhold model-identity labels between agents (paper 1); A/B measure rounds/tokens/success.
2. ●Do next (pipeline): Non-convergence report mode in swarm debates (paper 3).
3. ●Cheap experiment: Creativity-type semantic conditioning instead of temperature for ideation prompts (paper 5 proxy).
4. ●Medium-term: Strategy portfolio routing for optimization tasks (paper 4, inference-only proxy).
5. ●If we fine-tune: Hindsight self-distillation per PR-OPD (paper 2).

1. Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems（提示身份破坏多智能体 LLM 系统中的合作）⚠ 突破候选

• ●2609.35928 — 提交于 2026-09-28（已核验）
• ●方法： 受控实验：9–25 个智能体、最多 5 个开源权重模型家族，两个合作博弈 + 一个推理基准；操纵变量为"底层模型身份是否对同伴可见"。
• ●发现： 暴露身份引发"宗派化"——智能体优先与同标签同伴互动，即使任务既不奖励也不要求。有标签组平均多花 30% 轮次、55% token，成功率从 96% 跌至 81%。打乱或替换为任意标签，阵营随之迁移；移除标签，行为消失。作者给出简单有效缓解：隐藏身份信息。
• ●对本集群适用性： 高，立即可做。 配置级改动：不再转发模型身份元数据；成本 低；建议对辩论轮数/token/成功率做前后实测。

2. PR-OPD: Privileged Representation On-policy Self-Distillation for Agentic Reinforcement Learning（面向智能体强化学习的特权表征在线自蒸馏）

• ●2609.36642 — 提交于 2026-09-29（已核验）
• ●方法： 观察：一个"技能"放入上下文，把 WebShop 成功率从 42.2% 提到 56.2%，但只改变不到四分之一样本 token 的概率，却改变超过 80% 响应 token 的隐状态；修复：GRPO 预热后，策略为每条轨迹写"事后技能"，以停梯度方式自我教学，逐层对齐隐状态。
• ●发现： ALFWorld/WebShop、两个骨干全面最优：ALFWorld 最高 +4.7 点、WebShop +14.0 点；无外部技能库、无独立教师、零推理期开销。
• ●适用性： 中等——自研微调时受益；推理期可低成本近似（让智能体写复盘技能后带技能重读输出）。完整方法成本：高；近似：低。

3. KITA AI: A Multi-Agent LLM System for Pluralistic Policy Deliberation（面向多元政策审议的多智能体 LLM 系统）

• ●2609.34890 — 提交于 2026-09-28（已核验）
• ●方法： 智能体各自锚定不同人口群体"利益相关者人格"，审议政策情景；系统浮现"谁受影响"，为每个立场簇提供论据与量化指标。
• ●关键洞察： 把不收敛当作一等公民式的可解释输出，而非失败模式——给决策者权衡地图，不强制单一共识。
• ●适用性： 高。 为 swarm 辩论加"非收敛报告"模式（立场簇 + 理由 + 量化差异）；成本 低–中（仅改提示词与输出 schema）。

4. LLMs as Adaptive Meta-Solvers: Strategy-Diverse RL for Industrial-Scale Optimization（LLM 作为自适应元求解器：策略多样化强化学习用于工业级优化）

• ●2609.34427 — 提交于 2026-09-28（已核验）
• ●方法： SDRL 训练开源 LLM 在三种求解策略（求解器集成推理、精确组合算法、启发式搜索）间自适应选择；正确性门控的分层多样性奖励（策略间 + 策略内），混合格式训练（纯文本题 + 基于文件的真实工业实例）。
• ●发现： 作者宣称在基准均值与工业级任务上胜过微调方法与前沿模型，含 DeepSeek-V4-Pro 与 GPT-5.5。[作者声明——未独立复现。]
• ●更深洞察： 三种策略在不同问题结构与规模上互补——是组合拳论点。便宜替代：按问题特征选策略的路由层，无需训练。
• ●适用性： 有优化形负载时中–高；完整配方成本高。

5. Reinforcing Agentic Creativity in Scientific Ideation with Night Science（用"夜科学"强化科学构思中的智能体创造力）

• ●2609.35706 — 提交于 2026-09-28（已核验）
• ●方法： 基于认知科学三轴（行动/过程/结果）的 GRPO 创造力训练；摘要将框架命名为 AI Night-Scientist，标题中的"夜科学"指松散偶得探索，与结构化"日科学"相对。
• ●发现： 研究方向范围 +27.8%、贡献类型 +14.9%、预测引用影响力最高 +32.0 个百分点、原创性 +66.2 分；增益无法靠提高解码温度复制——"追求哪种创造力"的语义引导才是关键。
• ●适用性： 中等。完整 RL 配方成本高；推理期教训免费：构思多样性靠改"语义指令"而非温度。近似成本：低。

1. ●立即（配置）： 隐藏智能体间模型身份标签（论文 1），前后实测辩论轮数/token/成功率。
2. ●其次（管线）： swarm 辩论加"非收敛报告"模式（论文 3）。
3. ●低成本实验： 构思提示词用"创造力类型语义条件"替代温度（论文 5 近似）。
4. ●中期： 优化任务加策略组合路由（论文 4，纯推理侧）。
5. ●若开始自研微调： PR-OPD 式事后自蒸馏（论文 2）。

• ● 5 篇入选论文的 arXiv ID 前缀（2609.xxxxx）与今日抓取的 abs 页面所示提交月份（2026-09）一致。
• ● 5 篇全部落在 2026-09-28 → 2026-10-04 扫描窗口内。
• ● 标题与关键数字引自今日抓取的摘要；未由我杜撰缩写。KITA AI / SDRL / HaPRL / PR-OPD 出现在标题或摘要原文中；"AI Night-Scientist" 为论文摘要自命名。
• ● Papers With Code 本周无可核验的新 SOTA 声明，如实说明而非凑数。
• ● 未核验项：作者的性能声明（如 SDRL 胜过 DeepSeek-V4-Pro / GPT-5.5）——已行内标注为作者声明。