173 paper takes since 2026-08-19
Evolution strategies beat GRPO on Pass@K and reasoning diversity through sparse functional updates and population diversity, with a hybrid approach outlined.
WikiSkill co-evolves reusable agent skills with a persistent knowledge base so accumulated experience transfers across models.
A 3D modeling agent that generates editable, part-structured procedural assemblies with sharp geometry from text prompts.
VBVR-Pro offers a closed-loop benchmark for scalable, verifiable native visual reasoning across varied visual formats.
VGI-Bench tests visual reasoning in video generation models across 27 tasks, finding limited reliability and self-correction.
FrontierChallenge finds frontier models finish only ~20% of end-to-end scientific tasks despite high partial scores and frequent false completion claims.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
JIT-Agent trains a model to synthesize and evolve agent harnesses at runtime, lifting performance of off-the-shelf LLMs across diverse tasks.
Agent-G² models hint depth as a Gaussian distribution estimated online from existing rollouts to improve RL on long-horizon tasks.
Code World Model uses a coding agent to generate executable state updates while a separate video model renders observations from those proxy representations.
Mixed SFT on combined reasoning corpora beats next-chunk RL on both math and out-of-domain tasks in efficiency and final accuracy.
JoyAI-Echo-1.5 combines cross-shot memory, geometry-aware camera control, and rollout-aware training to hold identity and coherence across long sequences.
Open-MOPD fixes token-budget misallocation across teachers in on-policy distillation, recovering most of the oracle ensemble's capability.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
OraRL improves RL post-training for video multimodal language models using oracle rollouts, decoupled advantage estimation, and sign-balanced pruning.
A vision-language-action model that improves embodied generalization via a three-system architecture and joint alignment training.
AutoSaddler mines agent execution traces offline to harden LLM harnesses, targeting long-horizon benchmark failures.
DiffusionOPSD applies on-policy self-distillation to convert image-level rewards into explicit intermediate targets for diffusion models.
SecOPD uses token-level feedback during fine-tuning to reduce prompt injection attack success rates on language models.
OPDVR combines on-policy distillation with verifiable rewards using a ReLU-gated implicit reward reformulation to improve reasoning performance.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
BPCO stabilizes critic-based RL for language models via bounded value predictions, Monte Carlo targets, and adaptive advantage estimation.
Meta^n stacks a fixed meta-operation recursively to deepen reasoning in self-improving LLM agents while keeping the system stable.
EchoWM generates synchronized high-resolution video, audio, music, and speech from 6-DoF trajectories across first- and third-person views.
TLive-Omni unifies image, video, audio, and text for live-commerce via timestamped token grouping and reinforcement fine-tuning with verifiable feedback.
Hierarchical concept taxonomy, large-scale paired data, and dense supervision combine to sharpen diffusion-based image editing with granular evaluation.
An open-source harness using recursive subagents and agent-to-agent coordination to extend language models' long-horizon coding and reasoning capabilities.