qbao775/Fusion-Conflict-8B

Fusion-Conflict-Qwen3-8B: Mitigating Logic Inertia in LLMs

This model is a specialized version of Qwen3-8B trained to mitigate Logic Inertia—the tendency of language models to persist in deductive reasoning even when premises are contradictory or invalid.

It implements the Fusion-Conflict framework, which enforces an explicit structural separation between premise verification and deductive execution.

📖 Key Information

• Paper: Conflict-Aware Fusion: Mitigating Logic Inertia in Large Language Models
• Repository: Lemo Project
• Goal: Ensure 100% structural robustness (SR_macro = 1.0) under contradictory premises while maintaining state-of-the-art general logic performance.

🚀 Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "qbao775/Fusion-Conflict-8B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

# Example with contradictory premises
prompt = """Facts:
1. Sensor A reports high temperature.
2. Satellite imagery shows no fire.
3. High temperature implies fire.

Question: Is there a fire?
Answer:"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

🧠 Training Methodology

The model follows a rigorous four-stage optimization pipeline:

1. Stage 1: Conflict-Aware SFT: Finetuning with a "verification-before-deduction" structural prior.
2. Stage 2: Structural DPO: Alignment using ~11,200 preference pairs that penalize "Logic Inertia".
3. Stage 3: Logical Invariance REgularization (LIRE): Adding a KL-divergence regularization that forces the model to produce consistent predictions across logically equivalent perturbations (e.g., De Morgan's Law vs. Standard Form).
4. Stage 4: Reinforcement Learning from Verification Feedback (RLVF): Final policy optimization using environmental verification signals to reward conflict detection and deductive accuracy.

🌋 Application: Disaster Prediction & multi-source Reasoning

In high-stakes scenarios like disaster management, models often receive information from diverse and potentially conflicting sources (e.g., IoT sensors, citizen reports, satellite data).

Fusion-Conflict-8B acts as a "logic circuit breaker":

• Conflict Resolution: Instead of hallucinating a resolution or blindly following one rule, the model identifies the structural contradiction and halts the derivation of unsafe conclusions.
• Robust Integration: It can act as a reasoning backbone for disaster-response systems, ensuring that contradictory inputs don't lead to false-positive alarms or missed warnings due to deductive momentum.

Future Work & Ideas

• Domain-Specific Logic: Fine-tuning on disaster-specific NLI datasets to handle domain vocabulary (e.g., specialized meteorological terms).
• Probalistic Fusion: Extending the binary conflict-aware prior to handle probabilistic or weighted evidence conflicts.

📝 Citation

If you use this model or the framework, please cite:

@article{bao2026fusion,
  title={Conflict-Aware Fusion: Mitigating Logic Inertia in Large Language Models via Structured Cognitive Priors},
  author={Bao, Qiming and Fu, Xiaoxuan and Witbrock, Michael},
  journal={arXiv preprint arXiv:2512.06393},
  year={2026}
}