DailyPapers

LARC: Latent Reasoning in Continuous Space

Instead of reasoning in discrete token space, LARC reasons over continuous hidden states, boosting image generation quality and enabling test-time scaling. SOTA on GenEval (+7% over BAGEL).

Top papers on Hugging Face this week: agentic graphics, video generation, and embodied AI.

From Advances in Agentic Computer Graphics to Wave Forcing for streaming video, and from cognitive symbionts to VFlash for visual planning. Also: CoworkOS for long-horizon automation, SLA for spatial-language-action models, and Route Anything for omni-modal routing. Plus video generation insights from Kling.

IdeaTaste: Do LLMs share the research taste of human researchers?

A new benchmark freezes the AI literature at 2025, lets LLM agents write research proposals, and scores them against what the same communities actually published in 2026. Across 90,000 proposals, LLMs reach only 26% on citation matching and 18% on idea matching.

Axion 1.0: Open Agent Foundation Models, Simplified

ByteDance released a fully open agent model with a remarkably simple training recipe. Just two SFT stages followed by RL, achieving competitive results across HLE, SEAL-0, and FrontierScience.

FactEmbed: Factorized MLLM Embeddings for Composed Image Retrieval

Before embedding a composed query, it lets the MLLM imagine the target through factorized latents, then retrieves with the [EOS] embedding and reranks with the latents—outperforming prior work on CIR benchmarks.