A correct answer can still be an evidence failure. We ask whether language models disclose when their answer goes beyond what the supplied context actually supports.
| Video Temporal Reasoning: | |
 |
TempCloze: Can Video-LLMs Identify the Missing Middle?
EMNLP 2026 Findings
We focus on evaluating the "temporal reasoning" capabilities of existing frontier models and construct a filtered dataset of 1,521 long-take videos spanning three novel dimensions that are underexplored in previous video temporal reasoning benchmarks.
|
| Process Reward Model: | |
 |
SCI-PRM: A Tool Aware Process Reward Model for Scientific Reasoning Verification
KDD 2026
We construct SCIPRM70K, a large-scale dataset featuring Chain-of-Tool trajectories that explicitly interleave reasoning with the execution of scientific tools. Building upon this, we train an efficient reward model called Sci-PRM to provide fine-grained supervision on tool selection, execution accuracy, and result interpretation at each step in one inference.
|
| Visual Chart Coding: | |
 |
From Charts to Code: A Hierarchical Benchmark for Multimodal Models
ACL 2026 Main
A challenging chart-to-code benchmark with three hierarchical tasks.
|
| Human-AI Interaction: | |
 |
InterFeedback: Unveiling Interactive Intelligence of Large Multimodal Models via Human Feedback
EMNLP 2025 Findings
ICLR 2025 @ Bi-Align Workshop (Oral) We study whether LMMs can evolve via interactive human feedback and find: (1) Accuracy may not fully capture intelligence; (2) LMMs may cater to humans; (3) Low-quality feedback can hurt more than simple binary feedback.
|
| GUI Interaction: | |
 |
WorldGUI: An Interactive Benchmark for Desktop GUI Automation from Any Starting Point
ACL 2025 @ REALM Workshop
The first GUI benchmark targeting dynamic planning and action-event detection for desktop automation. In this work, we study the influence of environment feedback to agents.
|
| Post-Training of MLLMs: | |
 |
LOVA3: Learning to Visual Question Answering, Asking and Assessment
NeurIPS 2024
Only answering questions? Let’s think about asking and assessing questions when training MLLMs?
Without hyperparameter tuning or additional data annotation, consistent performance improvements are achieved! In this work, we study whether learning to ask questions and assess responses can strengthen a base model.
|
 |
Genixer: Empowering Multimodal Large Language Models as a Powerful Data Generator
ECCV 2024
First to examine MLLMs for data generation—showing diverse data synthesis and measurable downstream gains.
|
| Parameter-Efficient Tuning: | |
 |
SCT: A Simple Baseline for Parameter-Efficient Fine-Tuning via Salient Channels
IJCV 2023
Tuning only a small set of salient channels achieves up to 780× parameter reduction vs. full fine-tuning.
|