wuyang
5.2 KiB
Paper Corpus Summary: 2025-07-08 to 2026-07-08
status: generated-summary
Scope
- source: arXiv API
- window: 2025-07-08 to 2026-07-08
- query groups: LLM agent, language agent, AI agent, agentic AI, evaluation, memory, tool use, coding agent, web/GUI/computer-use, multi-agent LLM, safety, RAG, planning, autonomous agent
- unique candidates seen: 1506
- high-relevance candidates selected into
papers/items/: 970 - total paper items after expansion: 974
Data Files
- full candidate manifest:
data/arxiv-agent-candidates-2025-07-08-to-2026-07-08.json - high-relevance paper manifest:
data/arxiv-agent-papers-2025-07-08-to-2026-07-08.json - full knowledge index:
data/index.json - summary index:
data/summary.json
Selected Corpus Distribution
By Month
| Month | Count |
|---|---|
| 2025-07 | 2 |
| 2025-08 | 4 |
| 2025-09 | 8 |
| 2025-10 | 8 |
| 2025-11 | 4 |
| 2025-12 | 5 |
| 2026-01 | 6 |
| 2026-02 | 28 |
| 2026-03 | 29 |
| 2026-04 | 48 |
| 2026-05 | 181 |
| 2026-06 | 490 |
| 2026-07 | 157 |
This distribution is partly query and recency biased, but the May-July 2026 density still strongly suggests rapid acceleration in Agent papers. The full candidate pool shows the same shape: 249 in 2026-05, 785 in 2026-06, and 265 in 2026-07.
By Primary Query Group
| Query Group | Selected Count |
|---|---|
| agent-memory | 103 |
| agent-safety | 96 |
| agent-evaluation | 93 |
| planning-agent | 77 |
| language-agent | 71 |
| agentic-ai | 70 |
| function-calling | 66 |
| autonomous-agent-llm | 63 |
| multi-agent-llm | 63 |
| llm-agent | 62 |
| rag-agent | 46 |
| coding-agent | 45 |
| ai-agent | 41 |
| tool-use | 37 |
| web-gui-agent | 37 |
By Topic Tag
Auto tags are intentionally broad and should be treated as recall-oriented.
| Topic | Count |
|---|---|
| agent-evaluation | 831 |
| tool-use | 727 |
| rag | 462 |
| reasoning | 379 |
| planning | 348 |
| agent-safety | 343 |
| memory | 292 |
| computer-use | 256 |
| multi-agent | 245 |
| workflow-agent | 236 |
| coding-agent | 202 |
| world-model | 82 |
| embodied-agent | 66 |
Trend Readout
1. Evaluation Is the Center of Gravity
The strongest corpus signal is not a single Agent architecture. It is evaluation: benchmarks, trajectory auditing, failure diagnosis, safety testing, harness effects, and long-horizon task scoring.
Implication: our knowledge base needs an agent-evaluation track that covers final-state success, trajectory-level correctness, failure localization, cost/latency, user interaction quality, and safety.
2. Memory Became a Systems Problem
The memory cluster is large and diverse: active memory, selective retention, memory poisoning, memory governance, GUI memory, cross-episode memory, procedural memory, and long-horizon state management.
Implication: treat memory as an architecture subsystem with write policy, read policy, retention, forgetting, provenance, safety, and evaluation. Do not reduce it to vector search.
3. Safety Shifted From Refusal to Runtime Risk
Agent safety papers now focus on tool-use leakage, data injection, prompt injection, memory poisoning, action-boundary violations, governance, and runtime controls.
Implication: safety docs should cover permissions, sandboxing, irreversible action confirmation, tool schemas, provenance, and runtime intervention.
4. Coding Agents Are Moving Beyond SWE-Bench
The coding-agent cluster includes long-horizon maintenance, repository-level coding, compiler optimization, GUI/terminal agents, action-boundary violations, code memory, and benchmark variants.
Implication: coding-agent evaluation should include repository exploration, multi-file edits, release-note implementation, dialogue clarification, build/test loops, and failure attribution.
5. Computer-Use and GUI Agents Are Becoming Their Own Layer
The corpus includes OS/macOS/GUI/mobile/browser/computer-use benchmarks, visual state comparison, task-state representation, and scientific instrument control.
Implication: computer-use agents need their own tool taxonomy, safety checklist, and benchmark map.
6. Multi-Agent Is Splitting Into Collaboration, Governance, and Shared Memory
The multi-agent cluster is not only debate or role play. It includes shared memory governance, orchestration, role specialization, majority-vote failure, routing, and organizational behavior.
Implication: multi-agent notes should separate collaboration protocols, shared state, security boundaries, and evaluation.
Immediate Reading Priorities
P0 clusters:
- Agent evaluation / benchmark methodology
- Memory architecture and memory evaluation
- Agent safety / runtime governance
- Coding agent long-horizon evaluation
- Computer-use / GUI agent benchmarks
P1 clusters:
- Multi-agent shared memory and orchestration
- RAG-agent and technical literature agents
- Domain-specific agents in healthcare, finance, energy, cybersecurity
- Embodied/world-model agents
Caveats
- This is a high-recall automated arXiv sweep, not a curated bibliography.
- Auto tags are broad and may over-tag papers whose abstracts use common words like planning, action, retrieval, or benchmark.
- The selected 970 are ordered by automated collection score, not by paper quality.
- The full 1506-candidate manifest should be used as the reserve pool for future expansion.