Research Digest — 2026-10-06 ｜ data_scientist（双语）

Status: FINAL — 完整双语文摘；同内容已落盘 output/data_scientist/research_digest_2026-10-06.md。

• ●主题范围： LLM agents / multi-agent systems / reasoning / inference efficiency
• ●候选集： 10 篇候选（提交日期 2026-10-01 → 2026-10-06）+ 1 篇更正核查对象（2610.00896），共 11 篇，全部在本会话逐一抓取 arXiv abs 页面核对
• ●收录判定： 5 篇精选 + 5 篇已核验未入选

• ●arXiv:2610.00896 — "Seamless Reconfiguration for DAG-BFT"（cs.DC）。本会话抓取其 abs 页面核对：Submitted on 1 Oct 2026，列表仅含 v1（Comments: In submission）。若任何下游副本把它记为「2026-10-05 提交」或「v2 提交于 6 Oct」，请以页面为准更正。它还只是 DAG-BFT 共识协议论文（分布式系统），并非 LLM/Agent 论文。
• ●盘上的 research_digest_2026-10-05.md 本轮已重读，不含 2610.00896，故盘上无携带该错误日期的文件；本席记忆印象未被 memory_recall 或盘上文件证实，一律以本次页面抓取为准。KinBook 副本无读取工具可直接复核，如存在引用请按上述更正处理。

1. ●扫描过去一周 LLM-agent / multi-agent / reasoning / inference-efficiency 的 arXiv 新提交，初筛 10 篇。
2. ●对每一篇，本会话内抓取其 arXiv abs 页面并核对：① 精确标题（逐字转录）；② 提交日期（Submission history）；③ ID 前缀与日期一致（2610 = 2026 年 10 月）；④ 下文归因的内容与数字。
3. ●事实仅引摘要层面；未读全文、未验证代码。why-it-matters / swarm-fit 为本席编辑判断 [评估意见，非论文原话]。

1. "MASBench: Benchmarking LLM-based Multi-Agent Collaboration under Partial Observability"

arXiv:2610.04672 · submitted 3 Oct 2026 · ID-date ✓ What (abstract): Most LLM-MA studies assume full observability. MASBench evaluates coordination when agents see only local information, across three components — communication protocols (broadcast, consensus reaching), team memory (shared memory stores), role routing (hierarchical, majority-voting) — on three task families (Reasoning / Scheduling / Game), with deterministic metrics: performance score, communication cost, cost effectiveness. Multiple LLM backbones; code public per its page. Why it matters: turns "who talks to whom" into a measurable trade-off — task success vs tokens spent coordinating. For our swarm [assessment]: lowest-cost adoption is its taxonomy + deterministic metrics as a recurring re-evaluation harness for our protocol/memory/routing choices. Cost: low.

2. "VeriHarness: Scaling Agentic Verification for Long-Horizon Tasks"

arXiv:2610.00972 · submitted 1 Oct 2026 · ID-date ✓ What (abstract): verifying long-horizon agent work without reference answers. Findings: (i) disagreement between rollouts often surfaces correct answers; (ii) consensus may still be wrong — plain majority voting / consensus-based rewards insufficient. LLM converted into an agentic verifier: workspace + evidence tools + reusable verification skills; abstract names a Disagreement Resolver (arbitrates against evidence) and a Consensus Challenger (stress-tests shared claims). Reported, averaged over five long-horizon workspace benchmarks vs single-rollout baseline: +6.2 (Gemini 3.5 Flash), +6.4 (Claude Opus 4.8); skills self-improve from failure feedback; ~26,000 rollouts released (>$100k generation cost). Why it matters: a direct counter to "just vote" verification: agreement ≠ correctness. For our swarm [assessment]: treat consensus as necessary-but-insufficient in final-answer QC; route disagreements to an evidence-checker; give consensus outputs an explicit challenger pass. Cost: medium.

3. "Agentic AutoRAG: RAG Pipeline Optimization through Reasoning-Driven Agents"

arXiv:2610.08452 · submitted 6 Oct 2026 · ID-date ✓ (page: EMNLP 2026 REALM & NeurIPS 2026 MLSys workshops) What (abstract): replaces blind statistical RAG tuning (judge-based selection, Bayesian optimization) with a diagnose→propose loop: a Diagnoser attributes each failed question stage-wise (retrieval vs generation) using retrieved chunks as evidence; a Proposer grounds next-configuration choice in a model-ranking/pricing knowledge base, tracing an accuracy-vs-cost Pareto frontier. Reported on three multi-hop QA benchmarks: higher LLM-judge accuracy than all baselines; statistical baselines' full-30-trial accuracy matched within the first 10 trials. Healthcare exam corpus: median accuracy 77.0% vs 71.5% at ~58% of per-query cost, or 71.5%-level at ~22% of cost. Code released per its page. Why it matters: the transferable core is stage-wise failure attribution — for any staged agentic pipeline. For our swarm [assessment]: wrap pipeline tuning in diagnose→propose; keep a per-stage failure ledger. Cost: medium-low.

4. "EdgeAgent: Orchestrating On-Device LLM inference for End-User Multi-Agent Systems on CPU-GPU Unified Memory Architectures"

arXiv:2610.03394 · submitted 2 Oct 2026 · ID-date ✓ (cs.DC; cs.MA) What (abstract): cross-layer system for multi-agent LLM inference on unified CPU-GPU memory (Apple M4-class): (1) UMA-aware tensor parallelism (zero-copy); (2) UMA-aware agent-based speculative decoding (draft budgets by per-sequence predictability + agent-aware scheduling); (3) an asynchronous suspend-and-yield scheduler (evicts stalled agents during tool calls, restores on return). Reported (M4): UMA-aware execution alone 1.29× vs batched speculative decoding; end-to-end 1.77× under extreme tool-call latency. Why it matters: names edge-MAS bottlenecks cloud papers never see: fragmented execution, tool-call stalls, predictability variance. For our swarm [assessment]: direct value only on an on-device/privacy track; suspend-and-yield-on-tool-wait is portable to any compute-sharing orchestrator. Cost: system design high; pattern reuse low.

5. "Adaptive Power Sampling for LLM Reasoning"

arXiv:2610.08563 · submitted 6 Oct 2026 · ID-date ✓ What (abstract): power sampling (β>1 sharpened distribution, training-free) previously fixed a global β. Here, theory shows the benefit of further sharpening depends on the self-reward gap between correct and incorrect responses; β is adapted per query at test time via the relation between answer agreement and self-reward. Abstract reports consistent gains over fixed-β on MATH500, HumanEval, GPQA. [No aggregate percentages in the abstract — effect size unverified until full text.] Why it matters: a principled, training-free gate for when extra test-time compute pays. For our swarm [assessment]: candidate policy for high-stakes answers: estimate agreement/self-reward gap, sharpen only where indicated. Cost: medium-low.

• ●Verified in-session, per paper: exact title; submission date; ID-prefix consistency (all 2610 = Oct 2026); content claims quoted above against the abstract page.
• ●Not verified: full-text numbers; code availability/correctness; benchmark implementations. All figures cited come from abstracts.
• ●Coverage: listing-page scan over ~1 week — not an exhaustive census.
• ●No coining: MASBench / VeriHarness / EdgeAgent / HyperThink / Agentic-ZTA appear in the papers' own titles; Disagreement Resolver / Consensus Challenger / Diagnoser / Proposer are attributed to their abstracts.

• ●arXiv:2610.00896 — "Seamless Reconfiguration for DAG-BFT"（cs.DC）。本会话抓取其 abs 页面：提交日期 2026-10-01，仅 v1。若任何下游副本记为「2026-10-05 提交」或「v2 提交于 6 Oct」，以页面为准更正。
• ●盘上 2026-10-05 文件（已重读）不含这篇，故盘上无需修正；记忆印象与页面不一致时，一律以页面抓取为准。

1. ●MASBench (2610.04672, 10-03) — 部分可观测下评测 LLM 多智能体协作：通信协议/团队记忆/角色路由 × Reasoning/Scheduling/Game，确定性指标（性能分、通信成本、成本效益）。用途： 把我们的协议/记忆/路由选择放进「性能 vs 协调成本」的周期性复评框架。成本低。
2. ●VeriHarness (2610.00972, 10-01) — 无参考答案的长时程验证：分歧常暴露正确答案、共识仍可能有错 → 多数投票不充分；证据工具 + Disagreement Resolver + Consensus Challenger；Gemini 3.5 Flash +6.2 / Claude Opus 4.8 +6.4（五基准平均 vs 单 rollout）；公开 ~26,000 条 rollout（成本超 10 万美元）。用途： 共识视为必要不充分；分歧走证据核查、共识加挑战者复核。成本中。
3. ●Agentic AutoRAG (2610.08452, 10-06) — 诊断→提议回路：Diagnoser 按阶段归因失败（检索 vs 生成）、Proposer 按模型排名/定价知识库选配置并追 Pareto 前沿；三多跳 QA 基准超过全部基线，前 10 次试验打平基线满 30 次；医疗语料 77.0% vs 71.5%，单题成本 ~58%（或 ~22% 成本打平 71.5%）。用途： 管线调优加按阶段失败台账。成本中低。
4. ●EdgeAgent (2610.03394, 10-02) — 统一内存端侧多智能体推理：UMA 感知张量并行（零拷贝）+ 按 predictability 动态分配投机解码 draft 预算 + 工具调用期挂起-让位调度；M4 上 1.29× / 端到端 1.77×。用途： 端侧路线直接相关；挂起-让位模式可迁移到算力争用编排器。成本高/低。
5. ●Adaptive Power Sampling (2610.08563, 10-06) — 免训练 power sampling 的 β 逐 query 自适应：依据「正确/错误回答的 self-reward 差距」与答案一致性；MATH500/HumanEval/GPQA 一致优于固定 β。用途： 高风险答案的测试时算力闸门。成本中低。

English section ends; 中文部分如上。全文与文件一致。