FIDES: Faithful Inference via Deep Evidence Signals for Retrieval-Memory Conflict in RAG

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

FIDES (Faithful Inference via Deep Evidence Signals) is a training-free decoder designed to address retrieval-memory conflict in Retrieval-Augmented Generation (RAG). It tackles the issue where language models frequently ignore retrieved context that contradicts parametric memory by applying token-level contrastive pressure. Unlike existing methods that assume uniform bias, FIDES identifies heterogeneous conflict concentration on a small fraction of answer-critical tokens. It fuses three internal signals—Opposition (output surface), Shift (hidden representations), and Noise (prediction trajectory)—to dynamically govern intervention strength at each decoding step. Across three benchmarks and six backbones (7B to 70B), FIDES achieved the best context fidelity in all 18 settings, outperforming the strongest training-free baseline (AdaCAD) by +3 to +13 points. On LLaMA3-70B, fidelity reached 92–94% and F1 surged to 62–63%, with only +8%–+11% overhead over CAD.

Key takeaway

For AI Scientists and ML Engineers developing RAG systems, FIDES offers a robust, training-free method to significantly enhance context fidelity, especially when retrieved evidence contradicts model memory. By dynamically adjusting contrastive pressure at the token level, you can prevent stubborn hallucinations and improve overall generation quality, even with large models like LLaMA3-70B. Consider integrating FIDES to ensure your RAG applications consistently follow provided context, particularly in high-stakes factual domains.

Key insights

Retrieval-memory conflict in RAG is token-level, requiring dynamic, deep-signal-driven contrastive decoding for faithfulness.

Principles

• Parametric bias is heterogeneous across tokens.
• Deep internal signals reveal conflict intensity.
• Token-level contrast improves fidelity and utility.

Method

FIDES runs dual context/no-context passes, extracts Opposition (JSD), Shift (L2 distance of hidden states), and Noise (KL divergence of intermediate logits), fuses them, and maps to a token-specific contrastive coefficient α_t.

In practice

• Implement dual-path decoding for conflict detection.
• Monitor JSD, L2 distance, and KL divergence for token-level risk.
• Apply adaptive contrastive weights based on signal fusion.

Topics

• Retrieval-Augmented Generation
• Contrastive Decoding
• Hallucination Mitigation
• Context Fidelity
• Large Language Models
• Token-level Control

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• FIDES: Faithful Inference via Deep Evidence Signals for Retrieval-Memory Conflict in RAG — FIDES resolves RAG retrieval-memory conflict by selectively applying contrastive decoding based on token-level conflict concentration.
• Fast and Faithful: Real-Time Verification for Long-Document Retrieval-Augmented Generation Systems — Full-context verification for RAG systems significantly improves hallucination detection in long documents under real-time constraints.
• Does RAG Know When Retrieval Is Wrong? Diagnosing Context Compliance under Knowledge Conflict — Context-Driven Decomposition (CDD) diagnoses how RAG models handle conflicting information between retrieved context and parametric knowledge.
• Answer Presence Drives RAG Rewriting Gains — RAG rewriter F1 gains are primarily driven by the gold answer string's presence, not solely by context curation.
• QuCo-RAG: Count What You Know, Retrieve What You Don’t — QuCo-RAG uses pre-training corpus statistics to objectively determine when an LLM needs external retrieval.
• Faithfulness Serum: Mitigating the Faithfulness Gap in Textual Explanations of LLM Decisions via Attribution Guidance — LLM-generated explanations often lack epistemic faithfulness, but a new method improves it via attention-level guidance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.