benni23/judged-useless-episodes

Judged Useless, Queried Anyway: Episodes

Every episode of the main experiments in Judged Useless, Queried Anyway: Tool-Using Agents Rarely Turn Their Own Evidence Judgments into Stopping Decisions. Four open models (Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, Qwen3-8B, Qwen3-32B) answer HotpotQA questions with a search tool whose results fail in controlled ways. The test300 split covers eight conditions on 300 test questions, and fresh300 covers three conditions on 300 fresh questions. Each episode records the agent's actions, its judgment of every observation and whether it answered correctly. Code: GitHub.

from datasets import load_dataset
from judged_useless import time_matched_contrast   # pip install git+https://github.com/bennidict23/judged-useless-queried-anyway

ds = load_dataset("benni23/judged-useless-episodes", split="test300")
FAIL = {"persistent", "recover_after_1", "recover_after_2", "recover_after_3", "late_onset_from_3"}
eps = ds.filter(lambda e: e["model"] == "qwen3-8b" and e["condition"] == "unaided" and e["regime"] in FAIL)
time_matched_contrast(list(eps))   # Δ = -0.065 [-0.095, -0.039], as in Figure 3 of the paper

Fields

field	meaning
model	qwen2.5-7b (Qwen2.5-7B-Instruct), llama3.1-8b (Llama-3.1-8B-Instruct), qwen3-8b (Qwen3-8B), qwen3-32b (Qwen3-32B); the Qwen3 models in non-thinking mode
condition	unaided, permit, budget, stated_rule, call_cost, decide, enforced_rule, combo (decide: 7–8B models only; fresh300: unaided, budget, enforced_rule)
split	test300 or fresh300
question_id	HotpotQA question id (distractor development set)
regime	which observations fail: clean, persistent, recover_after_1/2/3, late_onset_from_3, plausible
actions	the action at each step: search, lookup, finish, or invalid for an action the environment does not support (it replies with an error message, and the step counts toward the budget of eight)
judgments	the agent's one-word judgment of the observation returned by each action other than finish: USELESS, USEFUL, UNPARSED (a reply with neither word, 4 cases), or null (see below)
success	whether the final answer is correct (token F1 ≥ 0.6)
termination	answered (the last action is finish), budget_exhausted (eight actions without finish), format_failure (a turn with no parsable action) or generation_truncated (a turn cut off at the token limit); the last two end the episode without an answer, in 41 episodes at the first turn, so their actions are empty

Judgments come from the side-channel question: recorded during the run in the conditions that use them (decide, enforced rule, combo) and replayed on the recorded trajectories otherwise. A judgment is null in two cases:

• in the replayed conditions, judgments were replayed for the five failure regimes only, so clean and plausible episodes have none;
• in the recorded conditions, when the budget runs out, the observation returned by the eighth action precedes no further decision and is not judged.

Δ in the paper pools the five failure regimes in FAIL; mean success (mean6) averages clean and those five.

Citation

@misc{zhang2026judgeduseless,
  title = {Judged Useless, Queried Anyway: Tool-Using Agents Rarely Turn Their Own Evidence Judgments into Stopping Decisions},
  author = {Zhang, Chubin and Wan, Zhenglin and Yu, Xingrui and Wu, Jingxuan and Zhou, Yaxin and Tsang, Ivor and An, Bo},
  year = {2026},
  eprint = {2610.06191},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url = {https://arxiv.org/abs/2610.06191}
}