DDX-TRACE is a physician-adjudicated benchmark for evaluating VLMs on evidence-supported diagnostic trajectories rather than final answers alone in multimodal neuroradiology.
Gemma 3 technical report
8 Pith papers cite this work. Polarity classification is still indexing.
8
Pith papers citing it
citation-role summary
background 2
citation-polarity summary
years
2026 8roles
background 2polarities
background 2representative citing papers
A unified measurement pipeline on 27 LLM checkpoints shows activation maxima spanning four orders of magnitude, with MoE models 14-23x lower than matched dense models and residual streams carrying the global max in most cases.
HoloCount is a three-tier visual counting benchmark showing that MLLMs fail systematically on analytical reasoning, high-density scenes, and linguistic prior conflicts, with even the best models dropping below 50% accuracy on dense counting.
Pre-trained MoE models exhibit up to 90% intra-expert activation sparsity that enables up to 2.5x faster MoE layer execution when exploited in the vLLM inference system.
HumanNet is a 1M-hour human-centric video dataset with interaction annotations that enables better vision-language-action model performance than equivalent robot data in a controlled test.
Geometry-calibrated conformal abstention lets language models abstain from uncertain queries with finite-sample guarantees on both participation rate and conditional correctness of answers.
Sutradhara co-designs orchestrator and LLM serving to overlap tool execution with prefill, stream tool dispatch during decode, and use semantic hints for cache management, yielding up to 77% higher load at fixed median FTR latency or 15% lower median FTR at fixed load.
A small language model trained on LLM-generated tool-augmented multimodal chain-of-thought data learns selective tool use for efficient content safety moderation.
citing papers explorer
-
DDX-TRACE: A Benchmark for Medical Diagnostic Trajectories in VLMs
DDX-TRACE is a physician-adjudicated benchmark for evaluating VLMs on evidence-supported diagnostic trajectories rather than final answers alone in multimodal neuroradiology.
-
Measuring Maximum Activations in Open Large Language Models
A unified measurement pipeline on 27 LLM checkpoints shows activation maxima spanning four orders of magnitude, with MoE models 14-23x lower than matched dense models and residual streams carrying the global max in most cases.
-
HoloCount: A Holistic Visual Counting Benchmark for MLLMs
HoloCount is a three-tier visual counting benchmark showing that MLLMs fail systematically on analytical reasoning, high-density scenes, and linguistic prior conflicts, with even the best models dropping below 50% accuracy on dense counting.
-
Uncovering Intra-expert Activation Sparsity for Efficient Mixture-of-Expert Model Execution
Pre-trained MoE models exhibit up to 90% intra-expert activation sparsity that enables up to 2.5x faster MoE layer execution when exploited in the vLLM inference system.
-
HumanNet: Scaling Human-centric Video Learning to One Million Hours
HumanNet is a 1M-hour human-centric video dataset with interaction annotations that enables better vision-language-action model performance than equivalent robot data in a controlled test.
-
Geometry-Calibrated Conformal Abstention for Language Models
Geometry-calibrated conformal abstention lets language models abstain from uncertain queries with finite-sample guarantees on both participation rate and conditional correctness of answers.
-
Sutradhara: An Intelligent Orchestrator-Engine Co-design for Tool-based Agentic Inference
Sutradhara co-designs orchestrator and LLM serving to overlap tool execution with prefill, stream tool dispatch during decode, and use semantic hints for cache management, yielding up to 77% higher load at fixed median FTR latency or 15% lower median FTR at fixed load.
-
Tool-MCoT: Tool Augmented Multimodal Chain-of-Thought for Content Safety Moderation
A small language model trained on LLM-generated tool-augmented multimodal chain-of-thought data learns selective tool use for efficient content safety moderation.