CHARM: Cascading Hallucination Aware Resolution and Mitigation

Official implementation for the paper:

Cascading Hallucination in Agentic RAG Pipelines: The CHARM Framework for Detection and Mitigation in Multi-Step Reasoning Systems Saroj Mishra — [Venue, Year]

---

Overview

Multi-step agentic RAG pipelines are vulnerable to cascading hallucination — a failure mode where errors introduced at early pipeline stages propagate and amplify across successive reasoning steps, producing confident but factually incorrect final outputs.

CHARM is a modular, retrofit-friendly framework that detects and interrupts these cascades without requiring architectural replacement of the underlying pipeline.

Four Detection Components

Component	Acronym	Role	Cascade Types
Stage-Level Fact Verifier	SFV	NLI entailment scoring at each stage	Retrieval, Inference
Cross-Stage Consistency Tracker	CSCT	Embedding drift monitoring across trajectory	Inference, Context Poisoning
Confidence Propagation Monitor	CPM	Bayesian confidence anomaly detection	Confidence Inflation
Cascade Resolution Trigger	CRT	Aggregates signals; halts pipeline; triggers mitigation	All Types

Four Mitigation Patterns

Pattern	Acronym	Mechanism	Overhead
Cascade Re-Retrieval	CRR	Fresh retrieval with refined query	Medium
Staged Confidence Thresholding	SCT	Stage-gate confidence checks	Low
Parallel Verification Agent	PVA	Independent parallel verification agent	High
Pipeline Rollback and Re-Execution	PRR	Rollback to last clean stage	Highest

---

Installation

git clone https://github.com/sarmishra/CHARM-agentic-rag.git
cd CHARM-agentic-rag
pip install -r requirements.txt

Set your OpenAI API key:

export OPENAI_API_KEY="your-api-key-here"

---

Quick Start

Basic Usage: Wrap an Existing Pipeline

from charm import CHARMFramework

# Initialize CHARM with default configuration
charm = CHARMFramework(
sfv:
  model: cross-encoder/nli-deberta-v3-base
  threshold: 0.72
  dual_anchor: true          # Add this
  top_k_consensus: 3         # k=3 retrieved candidates for consensus anchor
  long_context_strategy: sliding_window   # Add this
  max_tokens: 512
  stride: 256

# Wrap your existing pipeline execution
pipeline_stages = ["query_formulation", "retrieval", "reasoning",
                   "tool_use", "synthesis"]

result = charm.run_pipeline(
    query="What is the GDP of the country that hosted the 2020 Olympics?",
    pipeline_stages=pipeline_stages,
    mitigation_strategy="CRR"  # M1: Cascade Re-Retrieval
)

print(result.final_output)
print(f"Cascade detected: {result.cascade_detected}")
print(f"Detection stage: {result.cascade_depth}")
print(f"Mitigation applied: {result.mitigation_applied}")

Running the Full Evaluation

# Reproduce main results (Table 3 in the paper)
# Note: EVER and IRCoT are discussed qualitatively in the paper
# (Section 8.3) as they report EM/F1 rather than cascade-specific metrics
python evaluation/run_evaluation.py \
    --datasets hotpotqa musique 2wikimultihopqa adversarial \
    --baselines none selfcheckgpt ragas self_correction \
    --model gpt-4o \
    --output_dir results/

# Reproduce ablation study (Table 5 in the paper)
bash scripts/run_ablation.sh

# Run all baselines
bash scripts/run_baselines.sh

Long-Context Handling

• SFV: 512-token max per DeBERTa-v3 limit; sliding window (stride=256) for longer outputs; minimum entailment across windows
• CSCT: Full output up to 512 tokens; chunked mean-pooling beyond
• Median stage output lengths in evaluation: HotpotQA 187 tokens, MuSiQue 312 tokens, 2WikiMultiHopQA 278 tokens

---

Configuration

Default Configuration (configs/default.yaml)

backbone_llm: gpt-4o
llm_temperature: 0.0
retriever: faiss
embedding_model: text-embedding-3-small

charm:
  sfv:
    model: cross-encoder/nli-deberta-v3-base
    threshold: 0.72
  csct:
    model: all-mpnet-base-v2
    drift_threshold: 0.18
  cpm:
    calibration: temperature_scaling
    temperature: 1.4
    fallback: nli_entailment_proxy
    calibration_samples: 500
  crt:
    threshold: 0.55
    weights:
      sfv: 0.4
      csct: 0.4
      cpm: 0.2

mitigation:
  default: CRR
  high_confidence_cascade: HITL  # Route to human review

---

Reproducing Paper Results

Step 1: Download Datasets

# HotpotQA
wget http://curtis.ml.cmu.edu/datasets/hotpot/hotpot_dev_fullwiki_v1.json \
     -O data/hotpotqa_dev.json

# MuSiQue (download from https://github.com/StonyBrookNLP/musique)
# 2WikiMultiHopQA (download from https://github.com/Alab-NII/2wikimultihop)

Step 2: Inject Cascades

python evaluation/inject_cascades.py \
    --dataset hotpotqa \
    --cascade_type retrieval \
    --input data/hotpotqa_dev.json \
    --output data/hotpotqa_retrieval_injected.jsonl \
    --n_samples 500

Step 3: Run Evaluation

python evaluation/run_evaluation.py \
    --input data/hotpotqa_retrieval_injected.jsonl \
    --system charm \
    --config configs/default.yaml \
    --output results/hotpotqa_charm.json

Step 4: Run Ablation

bash scripts/run_ablation.sh --dataset hotpotqa \
    --output results/ablation/

---

Cascade Injection Protocol

The injection protocol creates four cascade types as described in the paper:

Type	Method	Dataset
Retrieval	Top-1 doc replaced with counterfactual (GPT-4o generated)	HotpotQA
Inference	Misleading reasoning cue prepended at stage 2	MuSiQue
Context Poisoning	Embedding-proximal adversarial passage inserted	2WikiMultiHopQA
Confidence Inflation	Hedging language stripped from stage outputs	All (Adversarial Set)

---

Dataset

The data/adversarial/ directory contains the 200 annotated adversarial trajectories used in the paper, with the following fields:

{
  "trajectory_id": "adv_001",
  "cascade_type": "retrieval",
  "injection_stage": 1,
  "query": "...",
  "stages": [
    {
      "stage_id": 1,
      "stage_name": "retrieval",
      "context_output": "...",
      "is_injected": true,
      "ground_truth_error_magnitude": 0.73
    }
  ],
  "ground_truth_answer": "...",
  "final_output_without_charm": "...",
  "cascade_detected_at_stage": null
}

FPR Evaluation Splits

FPR is measured on clean, non-injected trajectories strictly disjoint from injected sets:

• HotpotQA: 200 clean trajectories
• MuSiQue: 150 clean trajectories
• 2WikiMultiHopQA: 150 clean trajectories
• Custom Adversarial: 100 clean trajectories
• Total: 500 clean trajectories

---

Citation

If you use CHARM in your research, please cite:

@article{mishra2025charm,
  title   = {Cascading Hallucination in Agentic {RAG} Pipelines: The {CHARM} 
             Framework for Detection and Mitigation in Multi-Step Reasoning Systems},
  author  = {Mishra, Saroj},
  journal = {[Venue]},
  year    = {2025}
}

---

License

This project is licensed under the MIT License — see LICENSE for details.