ResearchPod

EgoServe tests 3,000+ instances across four time horizons. EgoMemo sets baselines without training, showing memory alone can ground timely assistance. I made a ResearchPod episode on it here: researchpod.app/episode/b6a2e6…

EgoMemo keeps three memories—multi-scale summaries, a knowledge graph, and visual embeddings—then retrieves across them at each step to decide on intervention.

Vinci2 reframes proactive assistance in egocentric video as deciding not just what to say but whether to speak at all.

The result is 40x faster generation than the Wan 2.1 teacher while matching its VBench scores on 49-frame 480p video, with 20 s per step on MediaTek Dimensity 8400. I made a ResearchPod episode on it here: researchpod.app/episode/0c4c05…

CineMobile prunes redundant transformer blocks, distills the model to 4 denoising steps with RL, and applies hybrid 4/8-bit quantization to fit under 1 GB.

CineMobile shows a Diffusion Transformer model for image-to-video can be compressed to run on-device while still generating cinematic camera motions like bullet time and dolly zoom.

After filtering shortcuts, state-of-the-art models perform only marginally above random guessing on the remaining tasks that require genuine spatio-temporal reasoning. I made a ResearchPod episode on it here: researchpod.app/episode/9099d9…

Video-Oasis audits existing benchmarks using visual dependency, temporal dependency, and ambiguity verification tests to isolate video-native challenges.

Video-Oasis: Rethinking Evaluation of Video Understanding finds that 55% of video-LLM benchmark samples can be solved without visual input or temporal context.

Framework design outweighs base model performance, while long-context reasoning and multimodal understanding remain the main bottlenecks. I made a ResearchPod episode on it here: researchpod.app/episode/46cb00…

UniClawBench runs 400 bilingual tasks inside live Docker containers with a three-role closed-loop setup: executor, hidden supervisor, and user simulator.

UniClawBench shows agent framework choice matters more than base model for proactive agents in live environments.

The 14B model runs on one GPU and pairs with a pilot-director harness to sustain coherent scenes across diverse actions and extended play. I made a ResearchPod episode on it here: researchpod.app/episode/666f27…

A MoBA attention mask mixes bidirectional context into autoregressive training, then two-stage distillation compresses the model for 60 fps at 720p.

Infinite Worlds with Versatile Interactions introduces LingBot-World-Infinity, which generates interactive video for over an hour without visible quality decay.