FlowMM follows layer-wise cross-modal information flow and token sensitivity to merge multimodal KV caches without losing critical context. Accepted to the EMNLP 2026 Main Conference.
NaviCache reformulates feature caching in video diffusion models as an inertial navigation problem, enabling plug-and-play, offline calibration-free acceleration with a dual-state self-calibration engine. Accepted to ICML 2026.
SPECTRA actively stress-tests an on-device GUI agent's visual cognition and analyzes its latent trajectories during prefill, enabling reliable cloud requests without autoregressive decoding. Accepted to ACM MM 2026.
VertiKV preserves cross-layer token-refinement paths with vertically aligned eviction and top-down information enhancement, enabling efficient multimodal long-context inference. Accepted to ACM MM 2026.
COPC reframes GUI grounding from single-shot coordinate regression to a Propose-then-Critic paradigm with maturity-aware co-evolutionary RL, unlocking latent Pass@k potential through visual discrimination rather than geometric clustering. Accepted to ACL 2026.
RetentiveKV reformulates KV cache eviction from discrete context truncation to continuous memory evolution using entropy-driven State Space Models, achieving 5× KV cache compression and 1.5× decoding acceleration on multimodal benchmarks. Accepted to ACL 2026 Findings.
We present UnicEdit-10M, a 10M-scale dataset of high-quality editing triplets, UnicBench, a comprehensive benchmark that goes beyond basic edits to assess spatial and knowledge-driven reasoning, and Qwen-Verify, a 7B dual-task expert model for scalable data quality control.
ThinkRec reframes LLM-based recommendation as a reasoning problem rather than pure pattern matching, letting the model explicitly reason about user preferences to improve both recommendation quality and user trust. Accepted to The Web Conference (WWW) 2026.
Graph2Eval turns heterogeneous documents and web pages into a knowledge-graph task space, enabling scalable, controllable, and semantically grounded evaluation for RAG and web agents. Accepted to CVPR 2026.
EcoAgent splits a mobile agent across device and cloud, closes the loop with on-device verification, and uses a Pre-Understanding Module to compress screenshots into text — matching cloud-only baselines on AndroidWorld while cutting MLLM tokens by 96% and per-step latency from 15.3s to 3.9s. Accepted to AAAI 2026.
