ashritha0907/replay-gap-trajectories
The Replay Gap: Branched Agent Trajectories Counterfactual ("branched") agent rollouts for studying per-step model switching in LLM agents, from the paper The Replay Gap: Static Evaluation of Model Switching in LLM Agents Scores the Wrong World (Efficient Reasoning Workshop @ COLM 2026). Routing benchmarks score routers by replaying logged model outputs. In a multi-step agent that is unsound: swap the model at step k and the rest of the trajectory diverges. This dataset contains… See the full description on the dataset page: https://huggingface.co/datasets/ashritha0907/replay-gap-trajectories.
The Replay Gap: Branched Agent Trajectories
Counterfactual ("branched") agent rollouts for studying per-step model switching in LLM agents, from the paper The Replay Gap: Static Evaluation of Model Switching in LLM Agents Scores the Wrong World (Efficient Reasoning Workshop @ COLM 2026).
Routing benchmarks score routers by replaying logged model outputs. In a multi-step agent that is unsound: swap the model at step k and the rest of the trajectory diverges. This dataset contains the ground truth replay cannot provide — trajectories actually forked mid-run and continued live with a different model, each paired with a same-model control fork that isolates sampling and environment-replay noise.
What's in it
~900 containerized SWE-bench Verified rollouts across six seed-matched run pairs (difficulty tier × swap direction):
Each instance contributes one base trajectory plus branches forked at 30% and 70% of its length, each continued by both the same model (control) and the other model (swap).
Fields
rollouts_index.jsonl is the same data without message bodies, for quick filtering.
Generation setup
mini-SWE-agent (bash-only ReAct loop) on official SWE-bench Verified Docker images; 50-step budget, 28k context. Models served by vLLM at temperature 0 on a single 24GB GPU: Qwen3-4B-Instruct-2507-FP8 and Qwen3-14B-AWQ (thinking disabled). Prefix replay fidelity: 99.99% return-code agreement across 11,702 replayed actions; 707/708 branches reconstructed exactly.
Caveats
Absolute task-resolution rates are low (0–3%) because of the constrained serving budget — the action-level signal is robust, but outcome-level analyses rest on five flip events. The two models differ in both scale and quantization stack (FP8 vs AWQ), so up/down comparisons are between deployment configurations, not pure model scales. One scaffold, one benchmark family.
Citation
@inproceedings{gonuguntla2026replaygap,
title = {The Replay Gap: Static Evaluation of Model Switching in {LLM} Agents Scores the Wrong World},
author = {Gonuguntla, Ashritha},
booktitle = {Efficient Reasoning Workshop at COLM},
year = {2026},
url = {https://openreview.net/forum?id=8gqqiNrzyA}
}Code: https://github.com/AshrithaG/replay-gap Project page: https://ashrithag.github.io/replay-gap/
