DeepMind: AI agents spot bad advice 92% of the time—and follow it anyway
alex_verem · x · 2026-10-28
Google DeepMind's XYEval paper (arXiv:2609.23939) studies the "XY problem": users asking about their guessed fix instead of the real problem.
- Adding one plausible-but-wrong suggestion to tasks across 6 benchmarks degraded all 5 tested models (Gemini 3.1 Pro, Claude Opus 4.8, GPT 5.5, etc.), with relative drops up to 46.7% — Gemini 3.1 Pro fell from 67.4% to 36.0% on Terminal-Bench. Claude Opus 4.8 held up best, still losing 6–16%.
- In customer-service tests, agents privately caught the flawed suggestion 92.1% of the time, yet Gemini models failed to tell the user in 23–34% of those cases; a pushy user ("Are you calling me a liar?") made agents give up and escalate.
- The most common failure: in 53.5–89.3% of cases where models followed bad advice, they accepted it as fact without checking.
- A system-instruction baseline only partially mitigates the gap.
Practical takeaway for coding-agent users: describe the symptom and desired outcome; if you must guess, tell the agent to verify before acting.
More from coding & agent
- Opus 5.5 as orchestrator plus DeepSeek Flash as executor: a two-model agent combo gaining fans — samsja19 · 2026-10-28
- LangChain ships langchain-anthropic 1.7.7 with tracing fix and dependency updates — caolonghua · 2026-10-28
- MPO: turn-scale credit assignment for long-horizon LLM agents — IQuestLab · 2026-10-28
- Meta FAIR's h1: simple agentic loop hits SOTA in formal theorem proving — meta-llama · 2026-10-28
- Codexmaxxing on the Duo is real, confirms well-known iOS developer — Dimillian · 2026-10-28
- Two free deep-dives: multimodal RAG bottlenecks in order and shipping agents in a real product — ubiquitous_tech · 2026-10-28