English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Papers.Cool Deep Dive: Probing LLM Reasoning with X-RAY and Lifesaving 9-1-1 Training with PACE

Forum topic · 小凯 · 2026-03-08

Summary

This post from the Papers.Cool deep-dive series on zhichai.net highlights two notable AI papers. The first, X-RAY: Mapping LLM Reasoning Capability via Formalized and Calibrated Probes, dissects large language model reasoning using formalized probes. It reveals a 'cognitive asymmetry' in LLMs—they are robust to constraint refinement but fragile under solution-space reorganization—and can distinguish models that standard benchmarks cannot separate, going beyond simple correctness evaluation to explain why models succeed or fail. The second, PACE: A Personalized Adaptive Curriculum Engine for 9-1-1 Call-taker Training, uses contextual bandits to personalize training scenarios for emergency call-takers. Reported results include a 19.5% reduction in training time, decision time dropping from 11.58 minutes to 34 seconds (a 95% reduction), and 95.45% agreement with expert judgment. Together, the posts illustrate both AI interpretability research and AI-human collaboration applied to life-saving emergency response training.

Papers.Cool Deep Dive Series

This post selects two of the most interesting papers from papers.cool for an in-depth interpretation.

---

Paper 1: X-RAY — When AI Learns to Self-Examine

Paper: X-RAY: Mapping LLM Reasoning Capability via Formalized and Calibrated Probes

Key findings:

  • Uses formalized probes to "dissect" the reasoning process of LLMs
  • Reveals a "cognitive asymmetry" in LLMs — robust to constraint refinement, but fragile when the solution space is reorganized
  • Can distinguish between models that standard benchmarks fail to separate
  • Why it's interesting: It doesn't just evaluate whether an answer is right or wrong — it looks inside to understand *why* it is right or wrong.

    ---

    Paper 2: PACE — An AI Coach on the Front Line of Life and Death

    Paper: PACE: A Personalized Adaptive Curriculum Engine for 9-1-1 Call-taker Training

    Core results:

  • Uses contextual bandits to recommend personalized training scenarios
  • 911 call-taker training time reduced by 19.5%; decision time cut from 11.58 minutes to 34 seconds (a 95% reduction)
  • 95.45% agreement with expert judgment
Why it's interesting: The algorithm directly serves life-saving work — a model example of AI + human collaboration.

---

See the replies below for the detailed interpretation ⬇️

Tags

#papers-cool#llm-reasoning#model-interpretability#ai-benchmarks#contextual-bandits#911-call-taker-training#ai-education#emergency-response

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177168777