NeuroCogMap maps LLM failures to distinct internal signatures

A team led by Zhongxiang Sun has posted a paper on arXiv that tries something interpretability research keeps circling. Instead of poking at what a large language model says, it maps what happens inside when the model fails. The framework, called NeuroCogMap, carves LLM internal features into 270 functional parcels and organizes them into a four-layer hierarchy the authors label perception, representation, abstraction, and application.

The interesting part is not the map itself, it is what the map catches. The paper claims each of the well-known failure modes leaves a different fingerprint inside the model. Hallucinations on TruthfulQA reflect "insufficient higher-order monitoring of misleading premises," while hallucinations on NQ-Open look like "disrupted coordination among factual-retrieval modules." Social bias reads as "misrouting of socially salient representations." Sycophancy shows up as "distributed weakening of independent judgment." Refusal failure is characterized as a shift from monitoring-evaluation-negation control during successful refusal toward planning and procedural execution. If those signatures hold, safety teams get something they have not really had, mechanism-guided detection instead of behavioral probing after the fact.

Why this matters if you are not building alignment tooling yourself: most current safety evaluation is behavioral. You run a benchmark, you count refusals or false statements, and you infer from output what might be broken inside. A per-parcel signature would let you catch a hallucination pathway or a sycophancy pattern at inference and target the intervention at the specific circuit responsible, rather than fine-tuning the whole model into a different personality.

The honest caveat is scope. The paper reports on three open models, Gemma2-2B, Gemma2-9B-IT, and Llama-3.1-8B, and its claim of "partially shared functional vocabulary across models" is exactly as hedged as it sounds. What the reporting does not give you is whether the same 270 parcels show up in a frontier closed model, whether the detectors can be gamed by adversarial prompting, or what the false-positive rate looks like outside a benchmark. A separate result is more concrete: NeuroCogMap features predict human cortical responses during naturalistic language comprehension at a mean r=0.407, against 0.219 for BERT and 0.367 for a Language Standard baseline.

For anyone building the interpretability layer that sits between a model and its downstream users, this is the direction worth watching. Distinct internal signatures for distinct failures is the thing that would make targeted intervention a real product rather than a research demo.