Fri 13 Nov
LiveGet it free
ZOTPAPER

News without the noise

← All concepts

failure attribution

When an AI agent fails at a task, failure attribution is the structured detective work of figuring out exactly *why*. It traces back through the agent's chain of steps and observations to pinpoint the specific decisive error that started the cascade, moving beyond a simple failure signal to root-cause analysis.

The papers show that naive attribution, like blaming the last step, is a surprisingly strong but limited baseline. Robust methods instead build dependency graphs to formally trace how early mistakes propagate into final failures. Most importantly, they demonstrate that accurate attribution is practically powerful: pinning down the exact cause enables targeted repair or reflection, which dramatically lifts success rates on the next attempt.

Papers
140
Released code
11
First seen
June 2026
Latest
Nov 2026

122 papers in the last two months, against 13 in the two before.

Start here

Three papers to read first, in the order they came out
  1. 1Structured backtracking beats reflection for long-horizon language agents

    VBL, a training-free extension of ReAct, maintains a goal stack with dual-horizon verification and typed failure diagnosis, improving task success across three benchmarks.

    Nov 2026
  2. 2Simulator bugs cause most robot policy failures in common benchmarks

    A fixed threshold in PhysX's collision detection creates false failures in robot simulation benchmarks; correcting it eliminates them.

    Nov 2026
  3. 3Language model libraries often fail outside their core calculation code

    GPU kernel library failures often originate in build and runtime environments, while specialized assumptions and inadequate testing allow correctness and hardware-specific problems to slip through.

    Nov 2026

How fast it is moving

JunJulAugSeptOctNov so far

Often studied with

agentic evaluationMeasuring AI agents' performance in complex tasks, highlighting failures in planning, monitoring, and real44 togetherverificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.36 togetherbenchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.22 together

Who is working on it

University of Illinois Urbana-Champaign 7Microsoft Research 6Ant Group 6Amazon 6University of Chinese Academy of Sciences 4City University of Hong Kong 4

The papers

Most central to this idea first, not most recent.

Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.