Shaobo Wang

Transforming continuous latent states into discrete tokens enables up to 20× compression while enhancing interpretability in reasoning tasks.

Coding agents struggle to create complete and engaging games, with top performers barely reaching 41.46% success in end-to-end game generation.

Leveraging historical solving traces transforms software engineering agents into self-evolving entities, achieving a 50.40% success rate on SWE-bench Verified after just three iterations.

LLMs waste compute on tokens that have already "figured it out" – DASH selectively skips these tokens during prefill, speeding things up without retraining or sacrificing accuracy.

Targeted neuron fine-tuning can unlock superior image translation capabilities in multimodal large language models, outperforming traditional methods by preserving pre-trained knowledge.

LLMs can achieve more consistent and reliable cross-jurisdictional financial reporting by acting as constrained verifiers within a structured, agentic workflow, rather than as free-form generators.

Current phone-use agents are often *too* helpful, routinely violating user privacy by filling in unnecessary personal information even when a task doesn't require it.

Unified multimodal models secretly contain separate inference pathways for generation and understanding, and FlashU unlocks this hidden potential for 2x speedup without retraining.

Dataset distillation gets a boost on long-tailed data with CSDM, which uses spectral distribution matching to prioritize realism in tail classes and achieves a 14% improvement over SOTA methods.