Researchers propose Comparison-based Preference Optimization (ComPO), a zeroth-order method for aligning large language models with human preferences that extracts directional…
Researchers show that in fine-grained multimodal LLM perception, localizing a region of interest tolerates roughly 3-4x stronger token compression than…
Researchers introduce HarnessVLN, a zero-shot, training-free agent harness that unifies instruction-following and object-goal embodied navigation by coordinating perception, memory, and…
Researchers introduce ScienceIDE, infrastructure that converts scientific code repositories into programmable, verifiable environments for training and evaluating scientific agents, addressing…
Researchers introduce CERA-MoA, a reinforcement learning framework for Mixture-of-Agents systems in which a dynamic query router and independently fine-tuned agent…
Researchers present FLAT (Flexible-Length Aligned Transmodal representations), a multimodal pre-training framework that jointly trains a shared encoder with text-to-image and…
Researchers reframe instruct-model fine-tuning as a direction-selection problem: rather than treating behavioral drift from the reference model as an uncontrolled…
DeepSeek-AI’s 552B-parameter multimodal MoE model with a Causal Encoder-Decoder architecture combines cross-layer KV cache reuse (CSA2) with FP4 KV caching…
Identifies termination-token mismatch between base students and post-trained teachers as a root cause of response-length inflation in on-policy distillation across…
An RSI-inspired, harness-layer approach that scales auto-research loops across many environments to discover four transferable harness mechanisms (action execution, context…
A controlled, component-level study (176 settings, 4 models, SWE-Bench Verified + Terminal-Bench 2.1) isolating how context-management strategy, planning, and action-space…
Proposes Self-Retiring On-Policy Distillation, where a skill-conditioned teacher trains jointly with RL and the student adaptively drops the teacher once…
