Research map / Large language models / Language, safety and interpretability

Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning

ICLR 2024 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Evaluates 10 open-source Flamingo-style LMMs (3B–80B) on hallucination, abstention, compositionality, explainability and instruction following, studies multimodal in-context learning as a training-free fix, and proposes MT-ICL, CoH-ICL and SC-ICL variants.

Problem

Task accuracy alone does not reveal flaws of large multimodal models (LMMs) such as hallucination, failure to abstain, weak compositionality, poor explanations and poor instruction following; the paper asks whether training-free in-context learning (ICL) can mitigate them.

Approach

Evaluate open Flamingo-style LMMs on five alignment-related axes, study how many-shot multimodal ICL affects each flaw, and propose training-free ICL variants (multitask-in-context, chain-of-hindsight with negative demonstrations, and self-correction for abstention) to address them.

Key results

• IDEFICS-9B CIDEr on VQA-X explanations rises from 67.43 (zero-shot) to 90.18 at 32-shot ICL; IDEFICS-80B surpasses SoTA explanations.
• OFv2-9B abstention F1 on TDIUC rises from 28.27 (4-shot) to 56.44 (32-shot) with ICL, while CHAIR_s on COCO increases with more shots beyond 4.
• Zero-shot LMMs are near random on CREPE HN-Atom (about 53-60% accuracy), and ICL gives almost no gain on atomic foils.

Measured effects reported in the paper

Surprising findings

• ICL amplifies object hallucinations: CHAIR_s increases with more than 4 shots for most models.
• Zero-shot with 2 text-only examples (no images) gives relatively high instruction-following scores, raising concerns about the ICL effect.

Limitations

• Only Flamingo-style models up to 9B for most analysis; limited range of abilities evaluated.
• Instruction-following evaluated qualitatively and with GPT-4 judging, which may be biased.
• ICL gains are limited and sometimes degrade with more shots; proposed variants give mixed results.

Experimental setup

Datasets: COCO TDIUC CREPE SugarCREPE VQA-X LLaVA instruction data

Benchmarks: COCO captioning TDIUC CREPE SugarCREPE VQA-X LLaVA-Instruct

Metrics: CIDEr CHAIR_s CHAIR_i accuracy abstention F1 CIDEr for explanations GPT-4 score

Baselines: OFA UnIVAL BLIP Large VinVL OSCAR Base SoTA explanation models on VQA-X

Builds on: Flamingo OpenFlamingo IDEFICS Chain-of-Hindsight CHAIR

Closest papers by meaning

• MIA-Bench: Towards Better Instruction Following Evaluation of Multimodal LLMs (ICLR 2025 · similarity 0.91)
• VL-ICL Bench: The Devil in the Details of Multimodal In-Context Learning (ICLR 2025 · similarity 0.90)
• Can Large Language Models Help Multimodal Language Analysis? MMLA: A Comprehensive Benchmark (NeurIPS 2025 · similarity 0.90)
• LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.90)
• AIM: Let Any Multimodal Large Language Models Embrace Efficient In-Context Learning (AAAI 2025 · similarity 0.90)
• Implicit Multimodal Alignment: On the Generalization of Frozen LLMs to Multimodal Inputs (NeurIPS 2024 · similarity 0.90)