talha1234567/Agentic-Ai

AgentSight — Step-Level Hallucination Classifier

AgentSight is the first gradient-trained step-level hallucination classifier on the AgentHallu benchmark.
It identifies which step in an autonomous agent trajectory is the root cause of a hallucination.

Results (locked test set, one-shot evaluation)

Metric	Value	95% CI	Note
Step Localisation Acc	47.8% (32/67)	[36.3%, 59.5%]	vs Gemini 41.1% — p=0.163, not stat. sig.
Judgment Macro-F1	54.7%	—	n=105 trajectories
Judgment Precision	58.5%	—	threshold=0.40
Judgment Recall	55.7%	—

The +6.7 point margin over Gemini-2.5-Pro is practically meaningful but not statistically significant at n=67. The 95% CI spans 23 points. We report this honestly.

Key Finding

Removing the [ACTION]+[OBS] tool-execution channel from the input drops step localisation accuracy by 31.8 percentage points (48.5% → 16.7% on the validation set). Tool output content is the dominant localisation signal.

Architecture

• Backbone: DeBERTa-v3-base with LoRA adapters (r=16, α=64) — 1.42% of parameters trained
• Context encoder: 3-layer Pre-LN Transformer (8 heads) over the step sequence
• Head: 2-layer MLP → binary hallucination logit per step
• Input: [QUERY] task [THOUGHT] reasoning [ACTION] tool_calls [OBS] responses (max 512 tokens)
• Decision threshold: 0.40 (tuned on validation set)

Quick Start

from agentsight_sdk import AgentMonitor

# Auto-downloads weights from this repo
monitor = AgentMonitor("YOUR_HF_USERNAME/agentsight")

result = monitor.evaluate_trajectory(raw_trajectory_dict)

print(result["is_hallucinated"])               # True / False
print(result["predicted_root_cause_step"])      # e.g. 3
print(result["step_probabilities"])             # [0.12, 0.08, 0.91, 0.04]

Input Format

{
  "question": "What is the capital of France?",
  "history": [
    {
      "step": 1,
      "content": "I will search for this.",
      "tool_calls": [{"name": "web_search", "arguments": {"query": "capital France"}}],
      "tool_responses": ["Paris is the capital of France."]
    }
  ]
}

Training

git clone https://github.com/Minato-sudo/agentsight
cd agentsight
pip install -r requirements.txt
./run_full_pipeline.sh   # verify → train → ablate → test (one-shot)

Citation

@misc{agentsight2026,
  title={AgentSight: Step-Level Hallucination Localisation in Autonomous Agent Trajectories via Tool-Execution Context},
  author={Namikaze, Minato},
  year={2026},
  note={Paper + code: https://github.com/Minato-sudo/agentsight}
}

Files

File	Description
best_agentsight.pth	PyTorch state dict — load with AgentSightModel()
best_agentsight_meta.json	Threshold (0.40), val metrics, best epoch
tokenizer_config/	DeBERTa-v3-base tokeniser config (max_len=512)

Limitations

• Evaluated on AgentHallu only — generalisation to other benchmarks unknown
• n=67 test set means wide confidence intervals
• Judgment F1 (54.7%) is below frontier models (Gemini 64.6%, GPT-5 70.2%)