Skip to main content
Search

Nanyang Technological University

53 related articles

Subscribe to updates
  1. arXiv

    When verifiers can reject but not certify, NOPO learns only from failures and lifts average pass@16 above GRPO on code, logic and math

    When a verifier can reliably reject wrong answers but cannot reliably certify correct ones, standard reinforcement learning with verifiable rewards reinforces shortcuts that merely pass partial checks. NOPO treats verifier-positive samples as unlabeled candidates and updates the policy only by suppressing trusted negatives, scaling each update by the share of non-falsified responses in the same rollout group. On Qwen3-4B and Qwen3-8B, average pass@16 on code, logic and math benchmarks exceeds GRPO and other baselines, and the advantage persists under inference-time scaling to 512 samples and cross-benchmark transfer.

  2. arXiv

    ScanDreamer gives quadrotors an online map to imagine with: 96.3% success and 3.6% collisions in unseen simulated layouts

    For quadrotor navigation through unknown clutter, ScanDreamer builds an online 2D scan-memory map from depth images and vehicle pose, and conditions its world model, actor and critic on that map, so imagined rollouts advance the pose over a frozen map and read geodesic progress and wall-contact signals from mapped geometry. Across 1560 test episodes in NVIDIA Isaac Sim spanning unseen layouts and parameter ranges, three seeds average 96.3% success and 3.6% collisions, against 59.6% success and 28.8% collisions for AirDreamer at the same online-interaction budget.

  3. arXiv

    NEO-VLA drops the vision encoder: action signals peak before the final visual layer

    NEO-VLA removes the fixed visual encoder from vision-language-action models: images are tokenized by lightweight patch convolutional layers and evolve with language through the backbone, and a flow-matching action expert draws on instruction-conditioned visual representations at different depths. In controlled comparisons the native architecture beats encoder-based counterparts; full fine-tuning narrows but does not erase the gap, and extra scale alone cannot recover it. Layer-wise probing shows action-critical signals often peak before the final visual representation.

  4. arXiv

    Changing only 5% of pixels blocks face cloning: SEAL holds identity similarity to 0.15 or less on CelebA-HQ and VGGFace2

    Adversarial perturbations added to photos can stop a diffusion model from learning a person's identity, but earlier methods altered the whole face densely, and the accumulated distortion became visible around facial features. SEAL confines the perturbation to 5% of spatial locations, jointly optimizing where it lands and its values with a differentiable top-k mask, and disrupts the latent variance of a frozen VAE encoder. On CelebA-HQ and VGGFace2, protected images reach above 41 dB PSNR, 0.97 SSIM and at most 0.07 LPIPS, while identity similarity after DreamBooth personalization stays at most 0.15.

  5. arXiv

    Web links no longer always pull toward neighbors: ZeroPolar reads link polarity without labels, reaching a 73.99% eleven-graph mean with GPT-4o labels

    Web platforms must classify new pages before anyone labels them, large language model labels per page are noisy, and links mean opposite things across sites: online stores often join same-category products, while university sites mostly join pages of different classes. ZeroPolar calls the sign with which links enter a prediction structural polarity, reads that sign without true labels from the excess agreement of LLM labels and text-prototype predictions along edges, and applies one signed propagation step. With one GPT-4o label per page, ZeroPolar-T reaches the highest eleven-graph mean, 73.99%, 6.01 points above LLM-BP.

  6. arXiv

    Mountainous first arrivals go to a multimodal agent: FA-Agent reads the gather and deletes picks it cannot judge

    First arrivals in mountainous seismic data are often weak, intermittent, and masked by strong direct waves and noise, so energy-based pickers jump to later arrivals or noise, and deep-learning pickers often need retraining for a new survey. FA-Agent combines a multimodal large language model, deterministic processing tools, and skill files that carry processing experience into an extensible framework: the agent surveys the whole gather to mark doubtful zones, then examines them enlarged and decides per segment whether to keep, correct, or remove each pick. On semi-synthetic gathers with known first arrivals and real mountainous noise, it picked clearly visible first arrivals accurately and removed picks where the first arrival was invisible instead of picking later events.

  7. arXiv

    Same Login Screen, Opposite Labels: LGWM Checks the Transition in 17 ms, Not the Screen

    For GUI agents, the safety signal may lie not in the screen but in the transition that produced it: the same credential screen can be a legitimate result of tapping Sign in or a hijacked result of tapping View order. LGWM skips rendering the next screen and predicts its representation directly in the observation encoder's space, reducing verification to a vector comparison; on RSWT-Bench, where each credential screen appears pixel-identically under both labels, the training-free consistency score reaches 0.987 AUC at 17 ms per decision, on par with the strongest closed-source vision-language models and about ten AUC points above generative GUI world models.

  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

    Robot policies learn action dynamics instead of actions: ODE Policy hits 79.2% on LIBERO after one epoch, versus 43.2% at best for baselines

    Most robot policies output action values at discrete control steps and leave the continuity of physical motion to be learned implicitly from data. ODE Policy instead learns a neural ordinary differential equation for the vector field of the action, and at deployment integrates each action from the previous one, so the executed trajectory is continuous by construction. After one epoch on LIBERO it reaches 79.2% success against at most 43.2% for the baselines, and it keeps that early lead as the action head of a VLA and of ImageWAM and as the actor of SuperMimic.

  10. arXiv

    Janus turns red-blue adversarial search on deserialization defenses, cutting held-out attack success to 6.2% across 16 deployments

    Unsafe deserialization lets attacker-controlled data trigger chains of existing methods and reach code execution, while gateway, application and system controls each see only part of the attack. Janus runs a black-box attacker agent against an isolated replica to surface exploit variants, has a defender agent summarize each verified variant into security-relevant effects, and uses counterfactual replay to separate complementary coverage from redundant enforcement before jointly selecting and scoping controls. Across 16 deployments in six Java and PHP applications, held-out attack success falls to 6.2% versus 18.8%–55.0% for gateway, application, system and cross-layer baselines, while 98.4% of benign functionality tests pass.