The BaseResearcharXiv
Uncovering Why LLMs Fail: From Internal Mechanisms to Adversarial Reasoning Trajectories
According to the abstract, Ding, Fan, Yuan and co-authors present CauTrace, a white-box framework that looks at a model's internal states to explain why reasoning fails. They evaluate it on ordinary reasoning failures and on paired adversarial perturbations, in Qwen and LLaMA-family models.
Why it matters
The same thing, for different readers.
For researchers
The authors aim to describe ordinary and adversarially induced failures with the same internal measures, going beyond judging only whether a final answer was correct or safe.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- A white-box diagnostic framework, CauTrace, using seven functional states, propagation of representation change, temporal-state coordination and trajectory topology; evaluated on clean reasoning failures and paired adversarial perturbations in Qwen and LLaMA-family models.
- What it found
- The excerpt is cut off before any results, so the findings are not known to us.
- What it does not show
- Not stated in the excerpt.
- Status
- A preprint: not yet peer reviewed
The record
- Questions
- Can AI systems act beyond what people intended or allowed? · Are AI companies putting stronger safeguards into practice?
- People
- Zirui Ding (Wrote it) · Xinxin Fan (Wrote it) · Hao Yuan (Wrote it)
- What we read
- We have read only the description the publisher gives. No quotations until someone reads the source itself.
- Found by
- The Radar, run of October 28, 2026, via
arxiv-ai-safety-query(tier 1: the source’s own channel or an official record) - Cite this
- arXiv, “Uncovering Why LLMs Fail: From Internal Mechanisms to Adversarial Reasoning Trajectories”, October 27, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-10-27-cautrace-why-llms-fail
Our notesLight lane: written from the triage excerpt only, which ends before the results.