Tom 文献雷达 · Agent RAG LongContext · 2026-08-30 08:40
候选摘要(8 条)
| # | 标题 | 来源 | 日期 | 信号 |
|---|---|---|---|---|
| 1 | Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models | HF Daily | 2026-08-25 | ⭐⭐⭐ 133票 |
| 2 | PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents | HF Daily | 2026-08-26 | ⭐⭐ 27票 |
| 3 | Procedura: Agentic 3D Modeling with Procedural Control | HF Daily | 2026-08-25 | ⭐ 10票 |
| 4 | CritICL: Inference-Time Weak-to-Strong Generalization from Small LM Failure Modes | HF Daily | 2026-08-26 | ⭐ 8票 |
| 5 | Luce: Relightable Gaussians for 3D Asset Generation | HF Daily | 2026-08-24 | 7票 |
| 6 | TacForcing: Streaming Action Generation with Execution-Time Tactile Feedback | HF Daily | 2026-08-25 | 5票 |
| 7 | EditaLive! Unified Character Video Editing for Live Streaming | HF Daily | 2026-08-26 | 3票 |
| 8 | What Does an Evaluation License? A Commit-Bound Census of Claim-Relative Inference | HF Daily | 2026-08-24 | 2票 |
高价值条目(4 条)
⭐⭐⭐ Agentic Game Dev → RL Data Engine(133票,2026-08-25)
链接: https://arxiv.org/abs/2608.25518
Scaling world models 的瓶颈不是数据量,而是 reward signal 质量。代码 agent 成功是因为编译器/runtimes 提供可执行 reward;3D/spatial 生成仍依赖 CLIP 等模糊 proxy,难以支持 RL post-training。本文提出用游戏开发构建可验证的 RL data engine——游戏天然有 grounded reward,是继 code agent 之后最有潜力的 RL 训练数据源。Agent + RL + World Model 三领域交叉,值得重点跟进。
⭐⭐ PILOT in the Loop:实时自我改进(27票,2026-08-26)
链接: https://arxiv.org/abs/2608.26530
现有 self-improvement 方法在执行结束后才处理 experience,导致无法重定向当前 run。本文主张 live self-improvement:在执行过程中同时 redirect 当前 run 并更新 persistent harness。架构上需解决单 agent 自纠正(任务执行+轨迹评估同 context)与 subagent 委托(执行+评估分离)的矛盾。对长期 agent 系统的 memory/反思机制有直接启发。
⭐⭐ CritICL:推理时弱到强泛化(8票,2026-08-26)
链接: https://arxiv.org/abs/2608.27455
核心洞察:同系列模型中,弱模型的 failure modes 在强模型上也有结构化规律。CritICL 利用弱模型 failure modes 作为 guidance source 做推理时增强,兼顾效率与性能。直接关联 RAG 评测场景——检索失败模式的结构化分析可用于设计更强的 retrieval-augmented reasoning 系统。
⭐ Eval Benchmark 深水区(2票,2026-08-24)
链接: https://arxiv.org/abs/2608.19269
标题: What Does an Evaluation License? A Commit-Bound Census of Claim-Relative Inference in Inspect Evals
124 个 Inspect Evals 单元全量审查发现:110 个在确定性推理之前就停止了,因为所需的历史证据或语义 grounding 不可用。论文形式化了"claim-replay layer",指出 eval artifact 报出的 metric 并不必然 license 其 claim——这对评测驱动的 RAG/长上下文献系统有重要警示:不要迷信 benchmark 数字,要追踪 claim replay 的完整链路。
Substack 线索(1 条)
The AI Engineer · The AI Agents Stack (2026 Edition) https://theaiengineer.substack.com/p/the-ai-agents-stack-2026-edition
三个值得关注的 convergence 点: 1. Agent guardrails 从 I/O filter 演变为 tool call 授权 + rate limit + 行为验证 2. Eval as infrastructure 三层架构:PR fast check → nightly LLM judge → production drift monitoring 3. 新 benchmark:Context-Bench(memory)、Recovery-Bench(错误恢复)、Terminal-Bench(coding agents)
数据来源:HF Daily;候选 JSON:/shared/research-kb/inbox/tom/_candidates/2026-08-30-agent-rag-longcontext-candidates.json