AHEAD-ALFWorld-1.7B

Overview

AHEAD-ALFWorld-1.7B is a text-based embodied-interaction agent for ALFWorld, built on Qwen3-1.7B and trained with AHEAD. It is evaluated as a standard interactive policy: at inference time the model acts from the ordinary observation–action history alone, with no analyzer, skill bank, retrieval module, or extra skill prompt.

This is the 1.7B checkpoint of the AHEAD-ALFWorld family (1.7B / 3B / 7B).

Evaluation protocol

All numbers are success rate (%) under one fixed protocol:

Setting	Value
Splits	ALFWorld seen (`valid_seen`, 140 games) and unseen (`out_of_distribution`, 134 games)
Coverage	Full split, every game evaluated once (`eval_sequential=True`, no sub-sampling)
Decoding	`temperature=0.4`, `do_sample=True`
Max steps / history	`max_steps=50`, `history_length=2`
Aggregation	Micro-average — Avg. = (total games solved) / (total games), pooled over all task families

Task families: Pick (pick_and_place), Look (look_at_obj_in_light), Clean (pick_clean_then_place_in_recep), Heat (pick_heat_then_place_in_recep), Cool (pick_cool_then_place_in_recep), Pick2 (pick_two_obj_and_place).

Split composition (fixed pool):

Split	Pick	Look	Clean	Heat	Cool	Pick2	Total
seen	35	13	27	16	25	24	140
unseen	24	18	31	23	21	17	134

Results

All Avg. numbers are micro-averages (total games solved / total games).

Split	Pick	Look	Clean	Heat	Cool	Pick2	Avg. (micro)
Seen (140)	82.9	53.8	70.4	68.8	68.0	41.7	66.4
Unseen (134)	66.7	44.4	51.6	73.9	57.1	58.8	59.0

Notes

Aggregation is micro-average (games-solved / games-total), not an unweighted macro-average over the six families.
Small-family variance. Look (13 seen / 18 unseen games) and Pick2 (24 / 17) have small denominators, so a few episodes move those family rates by several points.
Protocol caveat. Numbers use history_length=2 and temperature=0.4 sampling (not greedy); other regimes (e.g. history_length=5, greedy decode) are not comparable to the table above.