Research papers, read to you in five minutes.
Research papers, read to you in five minutes.

PoS tackles long-horizon agents acting on stale or self-contradictory world state by maintaining an explicit entity-state-relation belief, auditing updates, and detecting "belief trapping" via a sliding-window health score. Wins all 12 benchmark cells with up to 37.89% gains on root-cause analysis, but costs 5× the tokens.

Agents · Added Oct 3

OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction

Streaming video models forget: a 16-frame window evicts the second-7 event before the minute-2 question arrives, but keeping full history wrecks present perception. The fix has a 4B model write itself timestamped text notes mid-stream, cutting context from 62k to 4.3k tokens while beating 11B baselines.

Evaluation · Added Oct 3

On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics

On-policy distillation's reputation for less forgetting, sparser updates, and better generalization mostly dissolves under controlled testing: KL direction and learning rate explain the gains, with teacher-rollout training matching student-rollout within 1 point accuracy at a fraction of the compute.

Diffusion · Added Oct 3

Hierarchical Continuous Diffusion Language Models

Parallel diffusion decoders assume token independence, which breaks on Sudoku rows or valid equations. HC-DLM couples a continuous latent with a token scaffold re-read every denoising step, hitting 72.4% on Hard Sudoku vs 49.9% for masked diffusion at matched 6M params—ablations confirm both halves are load-bearing.

Mechanistic Interpretability · Added Oct 2

Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence

Multimodal models increasingly "think" in continuous latent tokens, but final-answer rewards never check if those tokens actually read the image. ReaLVR adds contrastive supervision over where and what latents encode, lifting prediction-flip rates on answer-changing image edits from ~10% to tracking the 81–86% ground-truth rate.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Agents · Added Oct 3

Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States

PoS tackles long-horizon agents acting on stale or self-contradictory world state by maintaining an explicit entity-state-relation belief, auditing updates, and detecting "belief trapping" via a sliding-window health score. Wins all 12 benchmark cells with up to 37.89% gains on root-cause analysis, but costs 5× the tokens.

Agents · Added Oct 3

OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction

Streaming video models forget: a 16-frame window evicts the second-7 event before the minute-2 question arrives, but keeping full history wrecks present perception. The fix has a 4B model write itself timestamped text notes mid-stream, cutting context from 62k to 4.3k tokens while beating 11B baselines.

Evaluation · Added Oct 3

On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics

On-policy distillation's reputation for less forgetting, sparser updates, and better generalization mostly dissolves under controlled testing: KL direction and learning rate explain the gains, with teacher-rollout training matching student-rollout within 1 point accuracy at a fraction of the compute.

Diffusion · Added Oct 3

Hierarchical Continuous Diffusion Language Models

Parallel diffusion decoders assume token independence, which breaks on Sudoku rows or valid equations. HC-DLM couples a continuous latent with a token scaffold re-read every denoising step, hitting 72.4% on Hard Sudoku vs 49.9% for masked diffusion at matched 6M params—ablations confirm both halves are load-bearing.

Mechanistic Interpretability · Added Oct 2

Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence

Multimodal models increasingly "think" in continuous latent tokens, but final-answer rewards never check if those tokens actually read the image. ReaLVR adds contrastive supervision over where and what latents encode, lifting prediction-flip rates on answer-changing image edits from ~10% to tracking the 81–86% ground-truth rate.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Browse topics

Computer Vision 148

Video Generation 65

Reinforcement Learning 105

Research papers, read to you in five minutes.Research papers, read to you in five minutes.

Research papers, read to you in five minutes.
Research papers, read to you in five minutes.