Xunliang Cai

AFD-Ledger reveals that optimizing deployment for AFD can drastically cut evaluation costs while exposing the nuanced performance dynamics between homogeneous and heterogeneous setups.

LSA slashes indexing overhead while maintaining full attention performance, enabling efficient long-context processing for models with up to one million tokens.

Combining RLVR and OPD through SAF not only prevents entropy collapse but also boosts performance across multiple benchmarks, revealing the potential for more effective reinforcement learning strategies.

Channel-wise adaptive learning rates in Gated Delta Networks unlock superior long-context recall, rivaling softmax attention without the quadratic cost.

Agent-as-a-Judge can outperform LLM-as-a-Judge in complex environments, but still struggles to reliably verify agent behavior, revealing a critical gap in current LLM-based agent evaluation.

Training long-context sparse attention models doesn't have to be a slow, imbalanced mess: SparseBalance achieves 1.33x speedup while *improving* accuracy.

Surprisingly, general-purpose vision models already contain better action representations for robotic control than specialized embodied models trained explicitly for that purpose.

LLMs that ace math and physics still struggle with general reasoning, achieving only 63% accuracy on a new K-12 level benchmark.

Stop uniformly distilling your LLMs: SCOPE selectively amplifies teacher guidance on incorrect trajectories and reinforces student uncertainty on correct ones, leading to significant gains in reasoning performance.

Achieve full-attention accuracy with 10x operator speedup and 4.7x throughput improvement in long-context LLM inference by overlapping KV cache transfers with computation.

LLM agents can internalize skills via in-context RL, achieving zero-shot autonomous behavior without the token overhead and retrieval noise of traditional methods.

Ditching mel-spectrograms unlocks SOTA text-to-speech with a surprisingly simple diffusion model operating directly on waveform latents.

LongCat-Next shatters the language-centric paradigm by unifying text, vision, and audio into a single autoregressive model with minimal modality-specific design, finally reconciling understanding and generation in discrete vision modeling.