EMNLP paper localizes faithfulness and citation errors in agentic deep research systems
EliasEskin · x · 2026-10-25
An EMNLP2026 paper, "Who is the Agent to Blame?", traces faithfulness and citation mistakes in multi-agent deep research systems back to individual agents. Key points:
- Defines four failure modes (hallucination, uncited input reliance, uncited output, insufficient citations) and uses up to three entailment runs to attribute each report sentence to the responsible agent.
- On the AI-Q system, the Orchestrator's agent-level error rate is 30.9%, yet it causes 84.7% of final-report mistakes; global citation recall is 58.7%.
- Two cheap fixes work: feeding the Orchestrator raw source snippets instead of researcher notes, or appending one sentence of citation guidance, lifting citation recall from 64.5 to 69.7 and precision to 94.1.
Poster presentation Sunday, Oct 25, Hall H.
More from coding & agent
- Hermes ships subagent routes plugin for per-task model orchestration — intellectronica · 2026-10-25
- Claude drives ComfyUI for hours, auto-generates ~500 varied Kromoss images — hazadaza · 2026-10-25
- How LLM-based penetration testing harnesses work: LLM plans, harness executes — utku1337 · 2026-10-25
- Dev builds a 12MB PvP mini-game that hooks into AI agents to stop idle scrolling — CallumReadIt · 2026-10-25
- Open-source parametric CAD with one interface for AIs and humans, parts by Claude — UnspeakableTruths · 2026-10-25
- Jev for Recursive Language Models: half the cost, 6x faster at 1M tokens — dbreunig · 2026-10-25