Causal Localization and Repair of Reasoning Errors in Language Models

A common reasoning error in language models, the error a model makes when irrelevant text is added to a problem it would otherwise solve, has a location inside the model, a mechanism, a signature that can be read before the model answers, and a repair. We show this in eight open-weight models from five developers, from 1 to 27 billion parameters, on arithmetic word problems and, in the largest model, on predicting what a program returns. The fault sits in a band of layers at the same relative depth in seven of the eight models. Readings taken inside that band predict a wrong answer before it is generated. A correction applied there, inside the forward pass, with no retraining, no change to the weights, and no reference answer, reversed about half of the induced errors in the two largest models and a third in three smaller ones, and broke none of the sixty correct answers checked in each large model. The same readings predict the models’ own errors on problems with nothing inserted, on both tasks. Every finding was tested against a random change of the same size. The causal analysis was done by Nexus, a proprietary method for causal discovery invented in biology. This version withholds the exact layers, units, thresholds, and the form of the correction, which are patent pending. Until now, this error could be seen only in the answer. This study shows that a model’s own activations carry a signal that it is about to fail, that the signal sits where the failure is seated, and that the model’s own state, acted on at that site in the same forward pass, repairs the failure. The model reads its own state and acts on it: for this fault, it monitors and repairs itself. The method is not tied to arithmetic or to inserted text, and the result held across eight models from five developers and two kinds of reasoning. It finds the cause of a behavior inside the model, tests it by experiment, and corrects it at the source, and it can be run on any behavior that can be marked right or wrong. That includes the behaviors in which a model departs from what its users intend. Those behaviors have so far been addressed by instruction and training from the outside; this is a method for finding and correcting them inside the model, where they arise. The study also finds a weakness that every model tested shares and that an adversary could use on purpose. Knowing where it sits, and how to close it, is the defense.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-12
DOI
https://doi.org/10.5281/zenodo.24170115
Primary Topic
Explainable Artificial Intelligence (XAI)
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Causal Localization and Repair of Reasoning Errors in Language Models

Eric Luellen
Zenodo (CERN European Organization for Nuclear Research)
Explainable Artificial Intelligence (XAI)
article

Causal Localization and Repair of Reasoning Errors in Language Models

Eric Luellen
article en

Abstract

A common reasoning error in language models, the error a model makes when irrelevant text is added to a problem it would otherwise solve, has a location inside the model, a mechanism, a signature that can be read before the model answers, and a repair. We show this in eight open-weight models from five developers, from 1 to 27 billion parameters, on arithmetic word problems and, in the largest model, on predicting what a program returns. The fault sits in a band of layers at the same relative depth in seven of the eight models. Readings taken inside that band predict a wrong answer before it is generated. A correction applied there, inside the forward pass, with no retraining, no change to the weights, and no reference answer, reversed about half of the induced errors in the two largest models and a third in three smaller ones, and broke none of the sixty correct answers checked in each large model. The same readings predict the models’ own errors on problems with nothing inserted, on both tasks. Every finding was tested against a random change of the same size. The causal analysis was done by Nexus, a proprietary method for causal discovery invented in biology. This version withholds the exact layers, units, thresholds, and the form of the correction, which are patent pending. Until now, this error could be seen only in the answer. This study shows that a model’s own activations carry a signal that it is about to fail, that the signal sits where the failure is seated, and that the model’s own state, acted on at that site in the same forward pass, repairs the failure. The model reads its own state and acts on it: for this fault, it monitors and repairs itself. The method is not tied to arithmetic or to inserted text, and the result held across eight models from five developers and two kinds of reasoning. It finds the cause of a behavior inside the model, tests it by experiment, and corrects it at the source, and it can be run on any behavior that can be marked right or wrong. That includes the behaviors in which a model departs from what its users intend. Those behaviors have so far been addressed by instruction and training from the outside; this is a method for finding and correcting them inside the model, where they arise. The study also finds a weakness that every model tested shares and that an adversary could use on purpose. Knowing where it sits, and how to close it, is the defense.

Zenodo (CERN European Organization for Nuclear Research)
Explainable Artificial Intelligence (XAI)
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon