Skip to main content
Search

University of Chinese Academy of Sciences

5 related articles

Subscribe to updates
  1. arXiv

    Disaster reports stop at masks no longer: DisasterAgent ties every claim to image evidence and lifts report scores 13.5%

    Remote sensing disaster assessment has long stopped at segmentation masks, bounding boxes and category labels, while responders need checkable written reports. DisasterAgent feeds bitemporal imagery to a vision expert toolkit, organizes the tool outputs into a spatiotemporal perception graph, and draws inline-cited evidence from that graph to drive a vision-language model that was never fine-tuned on disasters. On DisasterM3 its average disaster report generation score is 13.5% above the Qwen3.5-27B model serving as its central brain; on RSCC, ST5-SCS is 15.34% above that same base model.

  2. PCOR stops concatenating peer cues with observations: conditioning lifts POO success on unseen peers from 30.2% to 51.1%

    Opponent modeling in multi-agent reinforcement learning usually concatenates an inferred peer representation with the current observation and encodes them jointly, leaving their relationship for policy learning to discover implicitly. PCOR instead imposes an asymmetric structure: observation features form the base representation of the current interaction state, and the peer representation only modulates how those features are interpreted. On two partially observable benchmarks, PCOR reaches higher final evaluation performance than the reported baselines on unseen peers; on POO, against a concatenation variant sharing the same encoders and differing only in fusion, evaluation success rises from 30.2% to 51.1%.

  3. aclanthology.org

    PDE solver debugging stops rewriting whole files: masking just the faulty span cuts one repair from 64.85 s to 0.748 s

    Automatic repair of PDE solvers now runs on a discrete diffusion language model: instead of rewriting the whole file left to right, it re-masks only the faulty code span and infills it with both left and right context visible. On five PDEBench families, with the same checkpoint, localized infilling raised average Pass@32 from 0.4938 to 0.5063 and cut per-attempt repair latency from 64.85 s to 0.748 s; from-scratch generation remains slower than same-scale autoregressive models, and worst-case end-to-end cost averages 55.5 s.

  4. arXiv

    SFT can generalize too: locking updates to one on-policy direction halves training time and edges past GRPO

    Constraining supervised fine-tuning updates to the direction identified by on-policy training lets SFT match GRPO's generalization while cutting training time by more than half.

  5. arXiv

    Divide and Conquer: A Hybrid Strategy Defeats Multimodal Large Language Models

    The work proposes JMLLM, a jailbreak framework that combines four concealment strategies—alternating translation, word encryption, feature collapse, and harmful injection—across text, visual, and speech modalities, releases the TriJail dataset with 1,250 textual and speech adversarial prompts plus 150 harmful images, reports leading attack success rates on 16 popular LLMs over AdvBench and TriJail at 24.65 seconds for a single query and 6 queries for the multi-round mode, and proposes a Harmful Separator defense that reduces attack success rates.

You’re all caught up