Skip to main content
Search

Hong Kong University of Science and Technology

39 related articles

Subscribe to updates
  1. arXiv

    LENS stops at an interaction score: a second model on held-out data must reproduce the pattern before a feature pair enters the model

    A high interaction score can reflect a model fitting noise, and ranking methods leave the cutoff to the user. LENS trains two models with the same learning algorithm on disjoint data and accesses them only through predictions: it ranks pairs by interaction strength after removing main effects, then tests whether the second model reproduces the pattern at a meaningful strength, with one multiple-testing correction setting the retained set. On F1–F10 with XGBoost, LENS recovers 91% of material interactions against 45% for DIAMOND, without selecting false pairs.

  2. arXiv

    MoE serving re-splits expert replicas every decode step: Relief's integer relaxation nears the per-step optimum, and an on-die engine lifts effective utilization to 99.5%

    In expert-parallel MoE serving, each layer finishes only when its most loaded GPU does, and routing shifts at every decode step. Relief replaces the per-step linear program with an integer bidirectional pairwise relaxation that halves load gaps along replica edges and nears the per-step optimum in a few passes. On decode routing from three open MoE families, scored as effective utilization relative to that optimum, the deployed static scheme reaches 66% to 72%, the GPU kernel 93%, and the on-die engine 99.5%, with the largest recovery over static placement at 37.3 percentage points.

  3. arXiv

    EmoReward stops emotional video RL from trusting fixed rewards: last run's traces set the next run's weights

    Emotional video generation has no single correct target, and fixed fusion of visual, motion and text-alignment rewards captures emotional intent poorly. EmoReward distills reward traces from one completed training run into per-emotion-regime reliability estimates, converts them into bounded weight multipliers, and combines them with current-video instance gates in the next run without updating any reward-model parameters. On HunyuanVideo 1.5 with DanceGRPO, FlowGRPO and SAGE-GRPO, VideoAlign Overall rises under all three algorithms, including SAGE-GRPO with EmoReward moving from -0.8269 to -0.2624.

  4. arXiv

    Offline RLVR stops supervising every token: GIST splits data by outcome and beats seven selectors on a shared pool

    Offline reinforcement learning with verifiable rewards usually supervises every token of every response in a fixed sampled pool, although responses and tokens are not equally informative. GIST splits the bidirectional KL mismatch between the policy and the Gibbs-optimal target of KL-regularized RLVR into a forward and a reverse term, then pairs correct and incorrect responses by sampling temperature and supervises only low-probability tokens of correct responses and high-probability tokens of incorrect ones. On Qwen3-0.6B/GSM8K with all methods choosing from the same candidates, it supervises 66.6% fewer tokens than training on all candidates and reaches 45.79 strict and 49.73 flexible accuracy, above seven selection baselines.

  5. arXiv

    Möbius moves denoising inside the language model: no motion VAE, and it leads HumanML3D on generation and captioning

    Möbius places the entire denoising process of motion generation inside the same multimodal representation space used for motion understanding, so clean motion, noisy motion and text share attention without a discrete tokenizer or a pretrained motion VAE. On HumanML3D, the 0.6B variant outperforms diffusion-based, language-model-based and unified motion-language baselines on both text-to-motion generation and motion captioning; scaling to 1.7B parameters and 1,700 hours of motion data improves all motion-language tasks further.

  6. arXiv

    EvoGuard screens every online skill rewrite: 56.5% average success on ALFWorld versus 45.5% for accepting all updates

    Language agents can store experience as natural-language skill documents, but rewriting a skill from a single trajectory after every deployed task lets its quality drift. EvoGuard turns each update into a two-way choice: offline, skills are scored on a private test set, rejection sampling keeps pairs with a clear utility gap, and a selector trained on the logits of two label tokens decides which skill to keep; at deployment, one forward pass settles which of the old and new skills stays. On ALFWorld, with Qwen3-4B and Qwen3-8B executors, it averages 56.5% success, against 45.5% for accepting every rewrite and 36.7% with no skill.

  7. arXiv

    Route Anything extends routing to images, video, speech and robots: learned routers beat the best fixed candidate by 0.3 points across 16 benchmarks

    General-purpose AI systems are shifting from a single model to orchestrating many specialized models, tools and policies, but routing had mostly been studied on language models. Route Anything builds a unified trajectory matrix of 25 benchmarks, 452 candidate configurations and 374,401 scored runs, together with a routing theory for quality–cost trade-offs. On 19,814 test inputs the per-input oracle beats the best fixed candidate by 5.38 points; learned routers beat it by 0.3 points across 16 benchmarks, by 2.1 points on video generation, and by 1.1 points on robot manipulation with 24% less policy time.

  8. arXiv

    Masked diffusion models lock answers too early: HiLo-RD distills large-window trajectories, lifting GSM8K accuracy by 9.4 points

    Masked diffusion language models can fix answer fragments before the supporting reasoning is resolved when the response window is small, and those fragments then become context for later predictions. HiLo-RD uses the same model's large-window trajectory as a frozen teacher and aligns its hidden representations to cropped small-window views at committed positions. At the same small inference budget, GSM8K accuracy rises from 68.84% to 78.24% while mean readable length falls from 362.8 to 271.3 tokens, and the gains extend to code generation without code-domain training.

  9. arXiv

    Chemistry figures no longer need manual transcription: RxnMiner lifts stereochemistry-aware reaction F1 from 21.0% to 71.0%

    Reaction figures in the chemical literature often pack several experiments into one shared template, substituent table and product grid, and recovering concrete reaction records from them has meant manual transcription. RxnMiner combines an expert-informed extraction procedure with a chemistry toolbox in one multimodal agent: it reads structures, binds evidence to each experimental entry, and instantiates undrawn reactants with a deterministic R-group engine. On ChemRxn-Bench, 603 reaction figures with 8,547 annotated records, it reaches 71.0% stereochemistry-aware reaction macro-F1 with a Qwen3.8-Flash-Next backbone, against 21.0% for direct prediction and 25.0% for the strongest existing extraction system.

  10. arXiv

    VA-TDM swaps four-step supervision for a velocity-field adversarial game: 45.53% human preference against the many-step teacher on 33B MiniMax-H3

    Few-step diffusion distillation has long relied on estimating the student score, yet at intermediate steps that score is hard to track and the estimated distribution can differ from the one used to update the student. VA-TDM instead supervises velocity fields built from the student's own sampling process, uses a learned critic to compare student and teacher fields and turns their disagreement into generator feedback, unifying score distillation, Fisher matching and moment matching. On four-step SD3.5-Medium it raises GenEval from TDM's 0.61 to 0.67; on 33B MiniMax-H3 at the same 4-NFE budget it wins 61.33% of human comparisons against Fast-H3 and reaches a 45.53% preference rate against the 100-NFE teacher, versus 24.67% for Fast-H3.