Skip to main content
Search

University of Macau

11 related articles

Subscribe to updates
  1. arXiv

    No sentence cuts, no waiting: CRESiST interleaves reads and writes so unsegmented streaming translation trails oracle segmentation by only 1–2 COMET points

    CRESiST replaces prefix-incremental generation in LLM-based simultaneous speech translation with an interleaved sequence of speech chunks and target tokens, so the KV cache is reused across the whole stream and no voice activity detection is needed to cut sentences. An RNN-T Actor samples read-or-write trajectories, a Qwen2.5-omni-7B Critic with Router and Refiner scores them on semantic consistency and latency, and reverse distillation feeds the decisions back to the Actor. On MuST-C and ACL 60/60, unsegmented streaming lands only 1–2 COMET points below the oracle-segmented setting.

  2. arXiv

    Autonomous driving planners stop learning from one score: reflective guidance folded into a single pass, training time cut by over 90%

    Autonomous-driving vision-language-action models have learned from a scalar trajectory score that flags a bad plan but cannot say whether scene understanding, behavior selection, or trajectory execution went wrong. Reflect-VLA treats reflective guidance, which pairs evaluation diagnoses with corrective suggestions, as training-time privileged information and internalizes it through dual-path on-policy self-distillation: risky trajectories align with the teacher's token-level corrective distribution, while reliable ones are preserved by self-imitation. It reaches competitive planning performance on NAVSIM v1/v2 and Bench2Drive, cuts training time by over 90% versus reinforcement learning, and generates trajectories in a single pass at inference without guidance or iterative refinement.

  3. arXiv

    Recurrent memory no longer needs training: frozen networks recall over 256 steps with half the trainable parameters

    In partially observable reinforcement learning, freezing the recurrent memory module and training only the output and mixer layers still matches fully trained counterparts on delayed recall, discrete and continuous control, and pixel-based tasks, while cutting trainable parameters by more than half.

  4. arXiv

    SignRAG combines hierarchical pretraining with retrieval-augmented reinforcement fine-tuning to make a gloss-free SLT model beat gloss-supervised methods on all CSL-Daily metrics

    SignRAG is a unified gloss-free sign language translation framework that first learns linguistically grounded sign representations via fine-grained sign-text alignment, then jointly pretrains the sign encoder with Qwen3-8B, adds a target-domain-only retrieval gallery for instance-specific translation cues, and applies RUG-RFT combining translation-quality and retrieval-utility rewards to suppress harmful reliance; it sets new state-of-the-art results on CSL-Daily, PHOENIX-2014T, How2Sign, and OpenASL, and is reported as the first gloss-free approach to surpass gloss-supervised methods across all reported metrics on CSL-Daily.

  5. arXiv

    OPUR unifies jailbreak objectives via a harmfulness-reweighted distribution, lifting AgentHarm attack success on Llama-3.1-8B to 75.00% under stochastic decoding

    The work reformulates LLM-agent jailbreaking probabilistically: it shows that the gradient of the log expected harmfulness with respect to the input equals the expected log-likelihood gradient under a harmfulness-reweighted output distribution, unifying expected-harmfulness maximization with target-likelihood optimization; building on this, OPUR constructs surrogate targets from the model's own stochastic rollouts to guide adversarial suffix optimization, achieving higher targeted attack success rates on AgentHarm, InjecAgent, and WebShop.

  6. arXiv

    EASE adapts logit perturbation and sampling temperature by predictive entropy during generation, cutting average TPR@1%FPR across five detectors on Qwen3-8B from 0.655 to 0.339

    Observing that perturbing next-token logits or adjusting sampling temperature reduces AI-generated-text detector performance, the authors propose EASE, a training-free and detector-agnostic framework that computes predictive entropy from the source LLM's next-token distribution and uses it to adapt both logit perturbation strength and sampling temperature, consistently reducing detection performance across three source LLMs and multiple detectors with negligible inference overhead and limited text-quality degradation.

  7. arXiv

    iS-KV compresses KV cache online via block-incremental SVD, retaining 82.6% accuracy at 4.06x compression on DeepSeek-R1-Distill-Llama-8B

    The work proposes iS-KV, an online low-rank KV-cache compression method for long-horizon reasoning that keeps a recent window exact, incrementally folds older states into bounded-rank representations, and synchronizes historical coordinates as the low-rank basis evolves to maintain representation consistency; on DeepSeek-R1-Distill-Llama-8B it reaches 82.6% accuracy at 4.06-fold persistent-KV compression (original model 83.6%), and on Qwen3-8B it reaches 89.2% accuracy at 5.64-fold compression, consistently outperforming token-eviction baselines under matched memory budgets.

  8. arXiv

    AutoDataBench scores Kimi-3 highest at 60.68 overall, yet the retrieval leader on in-distribution data ranks last out of distribution

    The work introduces AutoDataBench, a controlled testbed that isolates data intelligence under fixed training pipelines and resource budgets across tool-use data diagnosis and repair, retrieval data organization, and knowledge injection, evaluates seven frontier LLMs on iterative data decisions, finds that agent-curated data approaches expert references on tool use and retrieval while strong in-distribution scores do not guarantee out-of-distribution transfer, and shows that reusing optimization trajectories as mid-training data improves downstream coding performance.

  9. medRxiv

    StrokeAgent matched expert final decisions in 86% of 100 acute stroke cases, exceeding tool-free LLMs and neurologists

    The study presents StrokeAgent, a multimodal agentic decision-support framework built around temporal alignment and continuous information synthesis that processes clinical records, laboratory results, raw electrocardiographic (ECG) waveforms and source multimodal CT images to support emergency triage, the initial thrombolysis decision and final reperfusion planning, and in a curated set of 100 patients with multimodal clinical and imaging data enriched for clinically complex reperfusion scenarios it achieved 86% concordance with expert-adjudicated final decisions, exceeding the mean of three tool-free large language models by 13 percentage points and the mean of six resident and attending neurologists by 17 percentage points, with 76% complete pathway concordance, completing the full deci

  10. arXiv

    Ant Group proposes Marathoner: synthesizing tasks from million-line-scale GitHub PRs lets a 9B open model work 10+ hours and make 1000+ tool calls

    The work proposes Marathoner, an autonomous agentic model trained through a post-training pipeline of ultra-long-horizon task synthesis, rejection sampling finetuning, and reinforcement learning, achieving consistent and substantial improvements over the base Qwen3.5-9B across 5 ultra-long-horizon benchmarks and surpassing strong proprietary models on some, with analysis showing it can execute for 10+ hours and make 1000+ tool calls on highly challenging tasks.