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

Video generators fail infant-level physics—balls roll through walls, objects vanish behind occluders. WROP fine-tunes a 16B model on 1.5M Blender clips authored around object permanence and solidity, then scores it via blind human pairwise judging instead of unreliable VLM judges, topping true-continuation models at 320×192.

LLM Training · Sep 23

SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue

Group-chat memory breaks when five people discuss overlapping topics and flat retrieval collapses "Alice said Bob agreed" into one confused fact. SpeakerMem-R1 splits storage into verbatim messages and owner/source-tagged state records, and its RL-trained 3B Writer closes 95% of the gap to a much larger extractor.

Agents · Sep 23

Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents

JitMem flips agent memory: instead of distilling trajectories at write-time (delayed reward, guess-the-future-lesson), it stores raw traces and trains a curator to write task-specific briefings at read-time. Reward is immediate, plain GRPO works, and it beats write-time baselines by 16 SR points on WebShop.

Memory · Sep 23

The Past Frames the Future: Memory for Autoregressive Video Generation

Long-horizon autoregressive video generators forget: characters drift, opened doors reclose, layouts break once history leaves the attention window. This survey reframes the fix as a memory system with a formal read/write/manage lifecycle, arguing most published methods underspecify the eviction operator — which is exactly where long rollouts collapse.

Agents · Sep 23

The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks

Taste-Bench measures whether agents pick the better branch at mid-run decision forks, mined from paired trajectories where one succeeded and one failed. Frontier models top out at 59.7%, more reasoning tokens don't help, but distilling a hindsight-trained advisor lifts SWE-bench-Pro success from 14.6% to 33.7%.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Sep 23

Training Object Permanence in World Models

Video generators fail infant-level physics—balls roll through walls, objects vanish behind occluders. WROP fine-tunes a 16B model on 1.5M Blender clips authored around object permanence and solidity, then scores it via blind human pairwise judging instead of unreliable VLM judges, topping true-continuation models at 320×192.

LLM Training · Sep 23

SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue

Group-chat memory breaks when five people discuss overlapping topics and flat retrieval collapses "Alice said Bob agreed" into one confused fact. SpeakerMem-R1 splits storage into verbatim messages and owner/source-tagged state records, and its RL-trained 3B Writer closes 95% of the gap to a much larger extractor.

Agents · Sep 23

Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents

JitMem flips agent memory: instead of distilling trajectories at write-time (delayed reward, guess-the-future-lesson), it stores raw traces and trains a curator to write task-specific briefings at read-time. Reward is immediate, plain GRPO works, and it beats write-time baselines by 16 SR points on WebShop.

Memory · Sep 23

The Past Frames the Future: Memory for Autoregressive Video Generation

Long-horizon autoregressive video generators forget: characters drift, opened doors reclose, layouts break once history leaves the attention window. This survey reframes the fix as a memory system with a formal read/write/manage lifecycle, arguing most published methods underspecify the eviction operator — which is exactly where long rollouts collapse.

Agents · Sep 23

The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks

Taste-Bench measures whether agents pick the better branch at mid-run decision forks, mined from paired trajectories where one succeeded and one failed. Frontier models top out at 59.7%, more reasoning tokens don't help, but distilling a hindsight-trained advisor lifts SWE-bench-Pro success from 14.6% to 33.7%.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Browse topics

Mechanistic Interpretability 19

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.