Moving beyond accuracy-only evaluation — verifying that multimodal predictions are supported by correct perceptual signals.
Multimodal foundation models now achieve strong performance on audio-visual QA, vision–language, and video understanding benchmarks, yet accuracy can hide shortcut reasoning. Models may answer correctly by relying on linguistic priors or spurious correlations rather than attending to the visual regions, frames, or audio events that provide the true evidence.
Most existing benchmarks score only the final answer, offering limited insight into whether a model actually "saw" or "heard" what it needed. BEAM 2 centers on evidence-aligned multimodal reasoning: datasets, protocols, and metrics that jointly evaluate (i) answer correctness and (ii) perceptual grounding of the reasoning process.
We emphasize evaluation designs with verifiable anchors — bounding boxes, representative frames, temporal segments, and audio-event timestamps — and model outputs that include structured evidence references and short grounded rationales.
Verifying that predictions are supported by the correct perceptual signals — not linguistic shortcuts.
Linking multimodal reasoning to concrete audio-visual cues with interpretable anchors.
Jointly scoring answer correctness and grounding quality beyond accuracy-only reporting.
Establishing standardized, interpretable metrics for the next generation of trustworthy multimodal systems.
BEAM 2 welcomes contributions spanning methodology, empirical analysis, and systems-level advances in evidence-aligned multimodal evaluation.
Multimodal benchmarks for video, audio, vision–language, and audio-visual QA that go beyond final-answer scoring.
Spatial, temporal, and audio-event grounding and evidence localization for multimodal reasoning.
Evaluation of explanation faithfulness and evidence alignment beyond plausibility checks.
Metrics that jointly score answer correctness and grounding quality (composite or multi-objective).
Muting audio, masking regions, or swapping objects/events to probe genuine model understanding.
Human-in-the-loop protocols and auditing tools for interpretable multimodal reasoning and agentic deployment.
Robustness and fairness in multimodal reasoning across accents, dialects, demographics, and domain shift.
Documentation, licensing, privacy-preserving releases, and responsible leaderboards.
A half-day (4-hour) program with invited talks, oral presentations, a poster session, and a closing panel.
From accuracy to evidence alignment
Nanyang Technological University
UCLA / Inception
Google DeepMind
Distinguished University Professor
Nanyang Technological University
Applies statistics and mathematics to AI and data science with 200+ publications and multiple best paper and test-of-time awards. Fellow of the Australian Academy of Science, AAAS, ACM, and IEEE.
Assistant Professor
UCLA & CTO, Inception
Leads the MINT group at UCLA, focusing on generative models and reinforcement learning for scientific discovery. Previously at Meta FAIR and UC Berkeley. PhD from Stanford.
Staff Researcher
Google DeepMind
Works on improving the factuality and freshness of Gemini, Google's large language models. Previously developed dermatology experiences for Google Lens in Google Research's Health AI team. PhD in Systems Neuroscience from MIT.
A team spanning Carnegie Mellon University, Amazon AGI, Google, and Adobe, combining academic and industrial expertise in computer vision, multimodal learning, and large-scale evaluation.
Carnegie Mellon University
Assistant Professor
Carnegie Mellon University
PhD Student
Carnegie Mellon University
PhD Student
Carnegie Mellon University
MS Student
Carnegie Mellon University
PhD Student
Carnegie Mellon University
Amazon AGI
Applied Scientist
Amazon AGI
Senior Manager
Amazon AGI
Google
Staff Research Scientist
Google DeepMind
Senior Staff Engineer
Director
Adobe
Senior Applied Scientist
Adobe Firefly
We welcome original papers, position statements, and benchmark reports aligned with the workshop themes.
Submissions should follow ECCV formatting guidelines and be no longer than 8 pages excluding references.