Tsinghua's C3PO Cuts Vision AI Hallucinations by Retraining the Reasoning Chain
A new training recipe cuts hallucinations in multimodal reasoning models by trimming redundant thinking tokens and contrasting against deliberately induced errors.
- C3PO reduces hallucinations in multimodal reasoning models through CoT compression plus contrastive preference optimization.
- Reasoning models allocate less attention to visual tokens than their non-reasoning base, worsening hallucination.
- Hallucinated reasoning chains causally propagate to final answers; reliable chains almost always produce faithful answers.
- LLMlingua-2 scores token importance; keeping the top 90 percent prunes redundancy without new supervision.
- Negatives come from masked images and misleading prompts that deliberately induce the model's own hallucination patterns.
- Tested on five 7B MLRMs with consistent gains on CHAIR, POPE, AMBER, and GPT-4 benchmarks: paper.
Vision-language reasoning models can produce elaborate explanations while paying less attention to the image. Researchers from Tsinghua Shenzhen and Harbin Institute of Technology found that five open-source, 7B-class reasoning models often hallucinated unsupported visual details more frequently than their non-reasoning base models. Their C3PO study traces the problem to bloated, error-prone reasoning chains and proposes a two-stage training method that reduces hallucinations across several benchmarks.
Multimodal large reasoning models, or MLRMs, add chain-of-thought generation to vision tasks through supervised fine-tuning or reinforcement learning. The generated trace gives the model intermediate steps before its final answer. Representative systems such as R1-Onevision and MM-Eureka, however, hallucinated more severely than the Qwen2.5-VL base model on standard evaluations. In these tests, a hallucination means that the model describes an object or visual fact unsupported by the image.
When reasoning crowds out pixels
- Visual attention declines during generation. Compared with their non-reasoning bases, the MLRMs allocate less attention to visual tokens, which encode image content, and more attention to textual tokens. The generated reasoning increasingly shapes subsequent predictions, giving learned language patterns more influence over the answer. Attention allocation alone does not prove visual grounding, but it provides a diagnostic signal.
- Errors in the trace predict errors in the answer. Most answers generated from hallucinated chains of thought also contain hallucinations, at a higher rate than answers from the non-reasoning base. Answers based on reliable traces are almost always free of hallucinated content. The association makes the intermediate trace a useful training target instead of treating only the final answer.
How C3PO retrains the chain
C3PO, short for Chain-of-Thought Compression and Contrastive Preference Optimization, addresses trace length and trace accuracy in separate training stages.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
