Hugging Face Daily PapersMeng'en Qin, Junye Chen, Jucheng LiuSep 5, 20261 min readpaperadvanced

MLLMs Hallucinate when Information Distribution Drifts in Synergy Heads

Summary

This paper introduces HEAL, a method to identify and mitigate hallucinations in Multimodal Large Language Models (MLLMs). It finds that hallucinations stem from information distribution drift in "synergy heads" within the attention mechanism and proposes dynamic calibration to steer output towards factual evidence.

MLLM hallucinations are linked to information distribution drift in specific "synergy heads" of the attention mechanism.
HEAL uses causal noise intervention on multi-head outputs to filter out causally redundant attention heads.
Information distribution within the remaining heads is disentangled using counterfactual Difference-in-Differences.
Dynamic information calibration factors are injected into the value vectors of synergy heads to regulate visual-language dependencies.

Engineers building or deploying MLLMs should care about this work as it offers a novel, interpretable method to reduce hallucinations and improve model reliability.

8/10

llm machine learning computer vision nlp attention hallucinations

Related reading

VC-Attention: Value Smoothing and Softmax Casting for Low-bit Attention arxiv.org
VC-Attention introduces a training‑free low‑bit attention pipeline for diffusion transformers. It smooths value tensors via lightweight online clustering (V‑Smooth) and quantizes only the residual after subtracting block means, restoring the mean from the softmax row sum. It also replaces the FP32 softmax exponential with a fused FP8 cast (ExpCast‑FP8) that maps log‑scores directly to E4M3 probab…
Hugging Face Daily Papersarxiv.orgXingyang Li, Dongyun Zou, Shining Zhang8d1 minpaper
Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States arxiv.org
Mind2Dialogue introduces a psychology‑guided simulator that generates user mental states and uses privileged distillation to train LLM assistants on Oracle responses that know those states. The resulting models improve personalization and theory‑of‑mind metrics by up to 41 percentage points versus standard instruction‑tuned baselines.
Hugging Face Daily Papersarxiv.orgZixuan Wang, Yufan Zhou, Jinzhou Tang8d2 minpaper
Chat-based Large Language Models replicate the mechanisms of a psychic's con softwarecrisis.dev
The article argues that chat‑based LLMs don’t think; the perceived intelligence is a user‑side illusion created by validation statements that exploit the Forer effect, much like a psychic’s cold‑reading con. Understanding this bias helps engineers set realistic expectations for AI products.
Hacker News front pagesoftwarecrisis.dev2d22 minHN220300
SpectralShift: Effective Context Window Extension of Gated DeltaNet via Spectral Reparameterization arxiv.org
This paper introduces SpectralShift, a spectral reparameterization method for extending the context window of Gated DeltaNet (GDN) linear attention models. It reconfigures the decay spectrum by enhancing slow propagation and preserving fast-decaying modes, consistently improving long-context capabilities during continual pretraining.
Hugging Face Daily Papersarxiv.orgZian Liu, Yiwen Hu, Zican Dong9d1 minpaper
Register Tokens for Bounded-State Reasoning in Diffusion Language Models arxiv.org
Register tokens are fixed‑position embeddings that store a compact hidden state across diffusion‑based language model generation chunks, enabling bounded‑state reasoning without retaining all prior text. Post‑training on LLaDA and Dream shows up to +8.5 math and +19.5 code benchmark points versus plain text carry, and RL fine‑tuning further improves long‑horizon tasks.
Hugging Face Daily Papersarxiv.orgAlbert Ge, Chandan Singh, Yufan Zhuang8d1 minpaper
US Military had close call after using AI for hallucinated intelligence report cnn.com
A US Special Operations analyst used an AI chatbot to augment an intelligence report that falsely claimed a Chinese vessel was carrying nuclear‑related cargo. The report triggered a near‑miss military operation that was aborted once the hallucination was discovered. The incident highlights the lack of verification standards and the risk of relying on AI‑generated intel for targeting decisions.
Hacker News front pagecnn.comKatie Bo Lillis, Zachary Cohen3d4 minHN513388

MLLMs Hallucinate when Information Distribution Drifts in Synergy Heads

Summary

Related reading

VC-Attention: Value Smoothing and Softmax Casting for Low-bit Attentionarxiv.org

Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental Statesarxiv.org

Chat-based Large Language Models replicate the mechanisms of a psychic's consoftwarecrisis.dev

SpectralShift: Effective Context Window Extension of Gated DeltaNet via Spectral Reparameterizationarxiv.org

Register Tokens for Bounded-State Reasoning in Diffusion Language Modelsarxiv.org

US Military had close call after using AI for hallucinated intelligence reportcnn.com

VC-Attention: Value Smoothing and Softmax Casting for Low-bit Attention arxiv.org

Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States arxiv.org

Chat-based Large Language Models replicate the mechanisms of a psychic's con softwarecrisis.dev

SpectralShift: Effective Context Window Extension of Gated DeltaNet via Spectral Reparameterization arxiv.org

Register Tokens for Bounded-State Reasoning in Diffusion Language Models arxiv.org

US Military had close call after using AI for hallucinated intelligence report cnn.com