A 20B retrieval subagent trained with reinforcement learning inside a stateful search harness that externalizes recoverable search state (candidate pool, curated evidence, verification records). The harness lets the policy focus on semantic search decisions, improving curated recall and transfer robustness.
Learns fine-grained preferences over sub-trajectories to identify and penalize redundant steps in long chain-of-thoughts, letting models "fold" reasoning chains into concise paths; reports ~56% token reduction on DeepSeek-R1-Distill-Qwen-7B while keeping accuracy.
Trains LLMs with reinforcement learning using a surface chrF reward so models learn to extract and apply linguistic signals from rich context for translating completely unseen languages. Demonstrates better zero-shot translation than in-context learning or supervised fine-tuning, framing outcome-based RL as a meta-skill for language learning from context.
Analyzes the parameter-space geometry of on-policy distillation (OPD) for LLM training, showing OPD updates affect fewer weights, avoid principal directions, and rapidly lock into a low-dimensional update subspace. Compares OPD with supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) and studies implications for optimization and objective mixing.
Models visual preference as distributions over rubric scores and introduces Z-Reward, a teacher–student framework that decouples reasoning-heavy judgment (teacher trained with GDSO) from efficient deployment (student via RISD). Demonstrates higher human-preference accuracy and works as a differentiable reward for text-to-image optimization.
AI Agent Papers2026
Lets a single LLM simultaneously act as agent and environment to bootstrap co-evolutional training — using state-prediction process rewards (World-In-Agent) and failure-mode retrieval (Agent-In-World) to reshape training data; reports ~4% average benchmark gain.
Shifts branching and credit assignment in agentic RL from coarse units to fine-grained decision points in generated sequences. Uses a Branching Score combining token uncertainty and policy-induced likelihood gains plus procedure-level advantage scaling; improves performance across 13 benchmarks while keeping efficient tool calls.
AI Agent Papers2026
Orchestrates teams of sub-agents across text, image, audio and video by modality-aware task decomposition, online sub-agent specialization, and parallel execution; introduces DA-GRPO to train Orchestra-o1-8B and reports a ~10.3% accuracy improvement on the OmniGAIA benchmark.
AI Video Papers2026
Continuously watches live video and autonomously decides each second whether to speak, stay silent, or delegate; released together with an 8B vision-first model, time-aligned interaction data, training recipe, and a deployable real-time system. Designed for vision-triggered, low-latency streaming scenarios and evaluated across six real-world streams.
Adds interleaved text–image generation to existing image generators via a multi-agent pipeline: a planner sequences stepwise instructions, a critic detects and refines failures, and single-step RL (GRPO) reinforces per-step corrections—suited for visual narratives and embodied guidance.
Applies a population-level test-time scaling strategy that uses one model as generator, verifier, refiner, and ranker to search over candidate proofs. Combines generative-verifier RL and a low false-positive verifier with tournament selection to reach competition-level performance on IMO and USAMO.
AI Agent Papers2026
Moves repository search into a dedicated exploration subagent that issues parallel read-only READ/GLOB/GREP calls and returns compact file:line citations. Trained (4B–30B) with SFT+RL, it reduces main-agent token use up to ~60% and raises end-to-end success by up to ~5.5%.