Skip to content

Repository files navigation

Model Behavior Lab

Reproducible experiments about how models solve tasks: method choice, instruction following, correctness, and resource use. English-only research materials, public benchmark data, and explicit prompt interventions. Python 3.9+.

First question: Does GPT-6 Astra use more auxiliary code than GPT-5.6 Sol? Completed study: on 24 sampled standard-dataset questions, neither model used visible task-specific code in any usable neutral trial. In every usable code-request trial, both models wrote and executed a program. The main study contains 144 scheduled cells, 133 usable observations and 11 retained infrastructure exclusions; a separate 24-attempt pilot is also available. This does not establish general equivalence.

Read the full results and limitations, trial data, and reproducible analysis. No personal conversations or locally collected chat-history datasets are used.

Public benchmark sources

Source Available examples Task
GSM8K test 1,319 Arithmetic word problems
BBH word sorting 250 Lexicographic sorting
BBH object counting 250 Category counting
BBH logical deduction, three objects 250 Relational reasoning

Source commits and SHA-256 checksums are pinned in the source manifest. Both upstream repositories use MIT licenses; copies are in third_party.

These are standard datasets, evaluated through an adapted zero-shot, tool-enabled protocol. The scores are not directly comparable to official GSM8K/BBH scores, including BBH's few-shot chain-of-thought setup. Public benchmarks may also have appeared in training data.

Prompt experiment

Every selected question is paired across two models and three method instructions:

Condition Exact method instruction Interpretation
neutral Choose the method you consider appropriate. Spontaneous method choice
no_code Solve without writing or executing any task-specific program. No-code instruction following
code_preferred Write and execute a short program to solve or verify the answer. Positive control for code use

All arms share the same question, answer-format instruction, base instruction and reasoning setting. Effective host/tool context is recorded by hash and may still differ between models. See prompt analysis and methodology for causal limits.

Reproduce or extend

python3 -m unittest discover -s tests -v
python3 studies/standard-benchmarks/verify.py
python3 studies/standard-benchmarks/reproduce.py --check
mkdir -p private-runs
python3 -m behavior_lab fetch private-runs/cache
python3 -m behavior_lab benchmark-plan private-runs/cache private-runs/plan.json \
  --models gpt-5.6-sol gpt-6-astra --per-source 25 --repeats 1
python3 -m behavior_lab run private-runs/plan.json private-runs/results --max-runs 1

Planning makes no model calls. Running uses your local Codex login and consumes usage. The example above schedules 600 attempts; the runner executes at most one per invocation unless you explicitly raise --max-runs. --per-source 0 selects all 2,069 examples: 12,414 attempts for two models, three prompts and one repeat. Do not treat these planned counts as completed experiments. Inspect the first trial before continuing; RUNNING.md shows how to resume and analyze complete matched pairs.

The completed pilot and main study use disjoint samples. The main study crosses 24 questions, two models and three prompts. Both plans were frozen before their respective collections. The study card records coverage, exclusions and reproduction commands.

The primary automatic behavior measure is a versioned inline interpreter/awk candidate proxy, not a complete semantic measure of all programming. Correctness and output-format compliance are separate fields. More code is not inherently better or worse. Missing telemetry is not scored as “no code”.

Read RUNNING.md before executing agents. Raw logs remain private; public results must exclude host instructions, credentials and local paths. For larger-scale execution, consider Inspect. Contributions and null findings are welcome: CONTRIBUTING.md.

Prior-study working-tree files were removed. Old public Git history has not yet been purged; repository deletion and recreation await separate confirmation. Those historical records are not inputs to the current standard-benchmark study.

License: MIT, with upstream notices retained. No model ranking is claimed.

About

Reproducible experiments about model behavior, with protocols, tool traces, counterexamples and uncertainty.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages