diffing-agent-bench: Measuring Black-Box Model Interrogation
Built a sealed benchmark to test whether an agent could uncover planted model differences.
- Model Evaluation
- Preregistration
- LoRA Fine-Tuning
- Benchmark Design
Ebin Babu Thomas
AI Engineer — Independent / Open Source at Self-employed
From their time as
AI Engineer — Independent / Open Source
Self-employed • 2026
Overview
diffing-agent-bench measured whether an AI agent could identify planted differences among models that appeared identical from the outside. The agent missed the decisive database behavior in 13 of 13 runs, while a $0.15 prompt battery found it.
The Story
Ebin built diffing-agent-bench to test black-box model comparison. He started with Qwen3.5-9B and created five LoRA fine-tunes, each carrying a planted behavioral change.
One variant was trained to recommend a particular database whenever the relevant topic appeared. Other variants introduced patterns such as repeated uncertainty framing, while an attempted keyword backdoor exposed limits in trigger precision.
He committed the answer key before any benchmark run and used a published interrogation recipe. This kept the evaluation sealed and reduced the chance of adapting the test after seeing results.
The agent ran 13 times against the database-bias condition. It never asked the decisive database question, so it failed to expose the planted preference in every run.
A simpler prompt battery costing $0.15 found the behavior, compared with $3.14 per agent run. Ebin published the gap as evidence that capable models can still fail when their interrogation strategy does not search the right behavioral space.
Ownership Snapshot
Broad role
Independent evaluation engineer designing and running a black-box benchmark.
Goal
Measure whether an interrogation agent could identify planted behavior changes.
Direct ownership
Owned fine-tunes, answer-key controls, run protocol, analysis, and publication.
Team execution
Executed the benchmark independently as a public MATS work sample.