Presents Skill Self-Play (Skill-SP), a co-evolutionary training loop where a proposer, solver, and dynamic skill controller generate, solve, and verify tasks conditioned on reusable skills — balancing verifiable execution with open-ended task diversity to boost LLM tool-use and reasoning.
Empirically studies how transformer-based native multimodal pre-training scales under fixed compute, deriving compute- and data-allocation power laws and an efficiency frontier for model size, token count, and data mixture; evaluates cross-modal transfer and multimodal in-context learning.
Transforms open-ended LLM optimization into self-verifiable reinforcement learning by turning tasks into proxy environments that produce deterministic, rule-based rewards. Proposes RLSVR and SpyRL — an information-asymmetric self-play scheme where agents vote to identify a preassigned spy, yielding verifiable rewards without human annotation. Demonstrated on summarization, creative writing and mathematical reasoning.
Presents a 2.8T-parameter Mixture-of-Experts multimodal model with a 1-million-token context window and 104 billion activated parameters, targeting long-horizon agentic RL, coding, reasoning, and vision. Key innovations include Kimi Delta Attention, Attention Residuals, Stable LatentMoE (16 of 896 experts active per token), ~2.5× scaling efficiency over Kimi K2, and a public weight release.
AI Video Papers2026
Real-time streaming multimodal foundation model that uses a codec-native tokenizer (Mage-ViT) to encode motion- and residual-rich regions from video I/P frames, reducing visual token usage by over 75% and enabling up to ~3.5× wall-clock inference speedup after training on ~560M images and 100M video frames.
Co-evolves a solver skill and a rubric-generator skill for text-space LLM optimization under decoupled objectives to avoid rubric gaming without using gold rubrics. Solver updates use criterion-level feedback; generator updates use independent audits of requirement coverage and response discrimination.
Introduces AISPA, a user-centric framework to audit system prompts in LLM applications, and applies it to 3,249 instructions from 88 commercial products to classify protective versus problematic instructions. Highlights design variability, growing prompt length/protection, persistent problematic directives, and calls for transparency and oversight.
AI Agent Papers2026
A 365-day, order-level simulation benchmark for evaluating long-term coherence of LLM agents in seller-side e-commerce. Grounded in 98,843 real product records and 26 interactive tools, it pairs prompt upstream supplier signals with delayed downstream order outcomes to stress planning, memory, and tool use over long horizons.
A continuous-latent diffusion language model that preserves a high-capacity, decodable text latent and directly models its distribution via a block-causal diffusion transformer and query-based encoder–decoder; achieves top results on OpenWebText and XSum while scaling to 1B parameters.
AI Agent Papers2026
Reformulates long-horizon agent execution as explicit task-state management: a manager defines bounded subtasks, fresh-context executors run them, and read-only auditors verify outcomes. Shows large performance gains on WeaveBench, Terminal-Bench and OSWorld.
Analyzes why supervised fine-tuning (SFT) causes severe task conflicts under multi-stage multi-task training while reinforcement learning (RL) enables stable coexistence, attributing the effect to sparse, near-orthogonal RL parameter updates and proposing Parallel-RL to decouple multi-task training.
Detects and filters spurious token-level teacher supervision in on-policy distillation by estimating input-groundedness and removing high-impact misleading updates, improving OPD on both LLM and VLM benchmarks.