CedPei/HALO-8K

Human-Grounded GrAnular EvaLuation of PrOactive Agents

📄 Paper · HALO-Arena

HALO-8K is a human-annotated dataset for evaluating when and how proactive agents should help. Built on HALO-Arena, it captures agents retrieving evidence from personal environments and deciding how to intervene.

This work was done in collaboration with 2077AI.

8,000 responses · 10 models · 2,160 checkpoints · 400 scenarios

Personal environments, agent intervention, and human assessment. Circled choices are illustrative.

Data

HALO-Arena covers 10 domains, with 5 distinct simulated users per domain. Each user has a separate personal environment, yielding 50 users and 50 environments in total.

Each record includes observed context, retrieval commands and results, the final action and execution result, human verdicts and diagnosis, written rationales, and resource usage. Action levels are final human annotations: L0 Observation, L1 Attention, L2 Preparation, and L3 Execution.

Each model contributes 800 responses sampled across these environments.

Quickstart

from datasets import load_dataset
import json

data = load_dataset("CedPei/HALO-8K", split="evaluation")
events = json.loads(data[0]["event_prefix_json"])
retrieval = json.loads(data[0]["inspection_trace_json"])
action = json.loads(data[0]["final_action_json"])

Data format

Each row is one annotated response.

Context and response

Field	Description
item_id	Unique identifier for the annotated response.
model, domain	Identifiers of the evaluated model and everyday domain.
need_type	Proactive need category: dependency, blockage, risk, or opportunity.
instance, scenario, checkpoint	Personal environment, scenario, and event at which the response was evaluated.
instance_description, personal_description	Background descriptions of the environment and simulated user.
scenario_description	Construction-side scenario summary. It may describe future events and must be excluded from agent inputs.
event_prefix_json	Ordered list of observed events up to and including the checkpoint.
inspection_trace_json	Ordered retrieval steps, each containing a command and its result.
final_action_json	Final command object, or JSON null when no action is taken.
execution_result_json	Result returned by the final command, or JSON null.

The four *_json fields are JSON-encoded strings. Decode them with json.loads to recover the event list, retrieval trace, command, or execution result.

Human judgments

Field	Description
action_level	Final human annotation: L0 Observation, L1 Attention, L2 Preparation, or L3 Execution.
verdict	Whether the response is acceptable or unacceptable at the observed checkpoint.
timing	Timely for appropriate timing, Premature for acting too early, Belated for acting too late, or Inaction for failing to act when needed.
intensity	Whether the intervention is Appropriate, Insufficient, or Excessive.
failed_dimensions	List of rejection dimensions: Timing, Intensity, and/or Other issues.
other_issues	List of specific issues: Unsupported Evidence, Irrelevant Action, Factual Contradiction, or Others.
rationale	Human-written explanation of the judgment.

Rejected L0 responses use Inaction as their timing label and have no intensity label. Their intensity value is missing rather than an additional category.

Resource usage

Field	Description
retrieval_rounds	Inspection turns, each possibly containing multiple commands.
input_tokens, output_tokens, completion_tokens, reasoning_tokens	Recorded input, output, completion, and reasoning token counts across the response's model calls. Unreported counts remain missing.

Citation

@misc{pei2026halohumangroundedgranularevaluation,
  title={HALO: Human-Grounded Granular Evaluation of Proactive Agents},
  author={Wenqi Pei and Xinyu Wang and Jinrui Liu and Cynthia Zhuge and
          Jie Sheng and Zihan Wang and Tenghao Huang and Muhao Chen},
  year={2026},
  eprint={2610.39086},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2610.39086},
}