World modelsNovember 22, 202612 min read

WorldMemBench: A Comprehensive Benchmark for Memory of Video World Models

Turn the camera away and come back: is it the same world, and did what was happening keep happening? 260 cases, nine video world models, and memory that fades with every return.

Read the post →
An amber actor robot, a blue judge robot with a magnifying glass, and a green teacher robot at a board

Agentic AINovember 22, 202614 min read

Actor, Judge, or Teacher? Understanding Decision Models in Long-Horizon Agents

Decision models such as JEV struggle to act when subgoals must be met in order, yet judged in hindsight they teach a 1.5B policy that outperforms them: 72.8% → 82.9% on ALFWorld and 56.8% → 71.4% on WebShop.

Read the post →
Animated drawing: an ink path is nudged onto a new course across a torn paper fan

LLM steeringNovember 17, 202612 min read

EasySteer: A Unified Framework for High-Performance and Extensible LLM Steering

Steering vectors built into vLLM and attached per request: 81–91% of vLLM's throughput with steering on, and 10.8–22.3× faster than existing steering frameworks. EMNLP 2026 System Demonstrations.

Read the post →
A student robot climbs a ladder with LOC, REP and PATCH rungs while a frozen teacher holds up one evidence card

Code agentsNovember 15, 202612 min read

SWE-Ladder: Learning to Localize, Reproduce, and Patch via Progressive Privileged Self-Distillation

Show a frozen teacher only the slice of the verified repair that the current stage needs, and teach localization, reproduction and patching in repair order: 9.6% → 16.4% pass@1 on SWE-bench Verified with a 4B model.

Read the post →
Animated drawing: an ink path runs from a torn paper cube to a flat sheet, where a faint outline of the cube appears

Spatial intelligenceOctober 20, 202610 min read

GPD: Distilling Routed 3D Privilege for Spatial Reasoning in Vision-Language Models

A teacher that sees question-routed depth, semantic and bird's-eye-view cues trains an RGB-only model: 57.1 on VSI-Bench with Qwen3-VL-4B, ahead of GRPO and answer-privileged self-distillation.

Read the post →