taekbae/llm-agent-backup-20260930

Convention belief in LLM agents

An interpretability harness for zero-shot coordination in Hanabi. A language model watches a partner that follows one of two signalling conventions and has to work out which one. The interest is not in whether it succeeds but in where in the network the failure lives: whether the convention is represented at all, whether that representation drives the answer, and whether feeding it back repairs behaviour.

Measurements and the design record are unpublished and deliberately kept out of version control (see .gitignore). What is tracked here is the machinery.

What is in the repository

Path	Contents
zsc_interp/src/	Harness, experiments, gate suite
zsc_interp/BUGS.md	Defects found and fixed, each with a regression test
zsc_interp/PROTOCOL.md	The measurement protocol, stage by stage, with the gates
zsc_interp/PORTING.md	Moving this to another machine or a larger model
zsc_interp/logs/	Run logs
rule.md	Constraints the harness must respect
environment.yml	Pinned environment

What the harness does

Environment. Two-player Hanabi from the DeepMind engine, with a scripted partner whose policy value is held constant and whose signalling convention is the only thing that varies. A second scripted seat plays the evaluated position so that episode length and trajectory quality do not depend on model quality.

Serialisation. Every prompt line is a fixed number of tokens, derived from the game configuration rather than from samples, so that states at a given position are comparable across conditions and can be transplanted between runs. This is a hard constraint, not a convenience: activation patching needs the same content to sit at the same index in every condition.

Conditions. An input ladder that varies how much of the inference is done for the model — raw log, extracted evidence, evidence written into notes, the rule stated outright — crossed with how the answer is produced, either free generation or a single forced-choice token read from the logits.

Measurements.

• Layer-wise linear probes on the residual stream, with the controls that make a probe claim mean something: shuffled-label selectivity, a text baseline, and episode-grouped splits.
• Activation patching with a donor contrast, so that what is transplanted is the convention rather than the donor's trajectory.
• A closed loop that reads the convention out with a probe, injects a matching latent template at the entrance the patching experiments identified, and scores the resulting behaviour — at the level of single decisions and of whole games.

Gates. bug_regression.py, smoke_test.py, frame_test.py and identifiability.py all run before any model time is spent. The regression suite pins the md5 of the rendered rules block, so a pass means the prompt frame is byte-identical to the one earlier measurements used; a failure usually means a drifting tokenizer version.

Running it

PIP_EXTRA_INDEX_URL= conda env create -p ./env -f environment.yml
conda activate ./env
git clone https://github.com/google-deepmind/hanabi-learning-environment third_party_hanabi
pip install -e third_party_hanabi
cd zsc_interp/src && bash run_protocol.sh gates

Collected activations are model-specific and are not in the repository. Episode generation is seed-fixed, so re-running the collection scripts reproduces the same decision points.

A caveat that shapes how results should be read

The two conventions — the leftmost or the rightmost marked card — are a minimal pair chosen so that the value is identifiable from the public log and so that two partners disagreeing on it genuinely miscoordinate. They are not how people play: the H-Group rules real players use are conditional on the chop and on card danger. Any claim from this harness is scoped to the minimal pair.