The Mirage of Performance Gains: Why Contrastive Decoding Fails to Mitigate Object Hallucinations in MLLMs?

2024-05-10 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

A study reveals that widely used contrastive decoding strategies, including Visual Contrastive Decoding (VCD), Instruction Contrastive Decoding (ICD), and Self-Introspective Decoding (SID), fail to genuinely mitigate object hallucinations in Multimodal Large Language Models (MLLMs). The observed performance improvements on the POPE Benchmark are attributed to two misleading factors: a unidirectional adjustment of the model's output distribution, which biases responses towards "Yes" and artificially balances certain datasets, and the adaptive plausibility constraint, which effectively degrades sampling decoding strategies into greedy search. Experiments with LLaVA-v1.5-7B, LLaVA-v1.5-13B, and QwenVL-7B on MSCOCO and GQA datasets demonstrate that spurious methods, such as Prompt-Based Adjustment and Output Layer Modification, achieve comparable performance gains without addressing hallucinations, challenging common assumptions about contrastive decoding's effectiveness.

Key takeaway

For AI Scientists and Machine Learning Engineers evaluating MLLM hallucination mitigation, you should critically assess reported performance gains. Be aware that methods like contrastive decoding may achieve higher scores on benchmarks like POPE by merely biasing output distributions or degrading sampling to greedy search, rather than genuinely reducing hallucinations. Your evaluation criteria must extend beyond simple accuracy to ensure true mitigation, considering the impact of decoding strategies and unidirectional output shifts.

Key insights

Contrastive decoding's MLLM hallucination mitigation gains are spurious, driven by output bias and sampling degradation, not true reduction.

Principles

• Performance gains can mask true hallucination mitigation.
• Unidirectional output adjustments create false performance.
• Adaptive plausibility constraints can mimic greedy search.

Method

The paper introduces Prompt-Based Adjustment (adding "Answer Yes if possible" to prompts) and Output Layer Modification (forcing "Yes" if "Yes"/"No" probabilities are close) as spurious methods to demonstrate misleading performance gains.

In practice

• Scrutinize MLLM hallucination mitigation claims.
• Evaluate methods beyond benchmark accuracy.
• Consider decoding strategy impact on results.

Topics

• Multimodal Large Language Models
• Hallucination Mitigation
• Contrastive Decoding
• POPE Benchmark
• Decoding Strategies
• Performance Evaluation

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

Related on AIssential

• MACD: Model-Aware Contrastive Decoding via Counterfactual Data — MACD uses model feedback to create targeted counterfactual data, improving contrastive decoding for Video-LLM hallucination.
• Mitigating Hallucination in Vision-Language Models through Barrier-Regulated Adaptive Closed-form Steering — BRACS adaptively corrects LVLM hallucination by monitoring visual grounding and applying closed-form steering only when needed.
• Mitigating Hallucinations in Large Language Models Via Decoder Layer Skipping — DeLask mitigates LLM hallucinations by dynamically skipping deeper decoder layers identified via gradient driftance.
• Enhancing Video Representations with Spatiotemporal-Semantic Residual to Mitigate Hallucinations in Video Large Multimodal Models — SSCD mitigates VideoLLM hallucinations by using deliberately disrupted spatiotemporal and semantic negative features in contrastive decoding.
• The Bank Said GPT-5.5 Hallucinates Less. The Benchmark Said 86%. Here’s Why They’re Both Right. — Enterprise perception of AI model performance can significantly diverge from independent benchmark results.
• Hallucinations in LLMs: A Deep Technical Dive into Causes, Detection, and Mitigation — LLM hallucinations are inherent to their probabilistic nature, driven by training objectives that prioritize fluency over factual accuracy.

Open in AIssential →

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