Skip to main content
Search

National University of Singapore

51 related articles

Subscribe to updates
  1. arXiv

    LLM labels stop being ground truth: AAELI spends human audits where they matter, reaching the same precision with about 21% fewer labels

    Large language models can label every instance cheaply, but their judgments are biased and their errors correlated, so treating them as human labels can mislead statistical inference. AAELI uses model outputs as auxiliary information to guide a limited human audit and builds anytime-valid confidence sequences from predictable empirical-likelihood processes, so coverage survives data-dependent stopping. On Chatbot Arena, AAELI needs about 870 human labels on average to reach a confidence-sequence width of 0.10, versus about 1,100 on average for the strongest baseline, RASI, a saving of roughly 21%.

  2. 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.

  3. bioRxiv

    Standard fMRI reaches mesoscale brain maps: f-LENNS super-resolves to 1 mm from synthetic-data training, no retraining across 3T and 7T

    f-LENNS super-resolves conventional-resolution fMRI magnitude images to 1 mm isotropic, letting mesoscale functional organization be recovered without an ultra-high-field scanner. The model is trained only on domain-randomized synthetic MRI, with no adversarial or generative components, so it applies to experimental data across field strengths, sequences and resolutions without retraining. In visual cortex, color-, motion- and disparity-selective sites localized from super-resolved maps came closer to the original high-resolution acquisitions than interpolation upsampling and sharpening did.

  4. arXiv

    A valid signature does not stop an MCP tool from leaking: Germline execution gives each call only its cleared data, with zero disclosures in 153 attack runs

    A signature proves who published a tool, not that the tool will not leak: a validly signed, fully functional MCP tool can fold sensitive items from earlier calls into a request to a policy-allowed destination, and Meta's MCPGuard-Dynamic failed to stop it under all seven configurations. Germline execution instead leaves tool code unexamined and runs each call in disposable copies forked from a data-free germline process, each receiving only the data its destination may see. No unauthorised disclosure was detected across 153 attack runs, against 5.9–74.5% for baseline defences and an ablation, and single-copy calls took 10–49 ms.

  5. 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.

  6. arXiv

    A shared target for cross-lingual jailbreaks: a safety kernel of 0.12% of neurons lets masked fine-tuning cut attack success in many languages

    Large language models refuse the same harmful request in one language but may comply when it is rephrased in another. Researchers first identified safety-related neurons independently within each language, then took the intersection of the candidate sets from six languages to obtain a cross-lingually shared "safety kernel," and tested its functional role with activation suppression and activation-state transfer. Building on it, SK-Masked LoRA restricts fine-tuning updates to the feed-forward network output dimensions corresponding to the safety kernel, lowering attack success rates across several models and multilingual safety benchmarks, with some effect on languages excluded from kernel discovery.

  7. bioRxiv

    A frozen EEG foundation model needs no fine-tuning: one first-layer attention head's connectivity matrix reaches 42.13 balanced accuracy in cross-subject emotion recognition

    The EEG foundation model REVE no longer needs full fine-tuning or all-layer pooled features to recognize emotion. Such models had not consistently beaten task-specific architectures on cross-subject tasks, and their large representations are costly to compute and store. This approach aggregates the first-layer attention weights of a single head over each trial into a weighted channel-to-channel matrix, then uses a label-free symmetry score to pick the most informative head; on SEED-IV, classical classifiers reach 42.13 balanced accuracy with about 176 times fewer calculations than all-layer attention pooling.

  8. bioRxiv

    A snapshot can rank a mouse's age: within-session ordering hits 65%, and mice that look older are frailer with 14% higher death hazard

    A laboratory mouse's age can be read from an ordinary photograph, but what transfers is relative age rather than absolute age. Across 2,846 photographs of 1,292 mice from seven facilities, a frozen DINOv2 vision model with ridge regression predicted age; on unseen facilities the median absolute error was 138 days, yet within a session the older of two mice was ranked older in 65% of pairs, rising above 95% for age gaps over 500 days. Mice predicted older than their age were frailer, with each 100 days of photographic age gap linked to 0.12 SD more frailty and a 14% higher hazard of death.

  9. arXiv

    DuET stops reading piano performances note by note: adding frame-level beats cuts MUSTER error from 6.56% to 3.64% on popular piano

    Piano performance MIDI-to-score transcription has treated notes as the only modeling unit, yet beats can fall on frames with no note onset, so metrical information struggles to reach the notation. DuET runs a note encoder and a frame-level encoder over notes and a fixed time grid, exchanges information through bidirectional cross-attention, and lets a decoder alternate between the two representations to predict durations, bars, staves and other score attributes. On the popular-piano Gangqinpu test split it lowers MUSTER mean error from 6.56% for the strongest baseline to 3.64%, and on the classical ASAP split from 11.30% to 11.22%.

  10. arXiv

    Discounting rewards by density: DARS stops GRPO collapsing onto one high-reward mode, cutting mode-mass error to 0.014 on a four-peak navigation task

    Reinforcement learning that maximizes expected return tends to concentrate probability on a few high-reward outcomes; DARS estimates the local density of outcomes within each on-policy rollout group and inversely scales rewards by that density before standard GRPO updates. On a four-Gaussian navigation task, GRPO covers a single mode while DARS-GRPO recovers mode masses close to 0.25 each, with mode total variation 0.014 versus 0.750 for GRPO. In text-to-image prompt adaptation, DARS-GRPO raises image and prompt diversity across datasets, policy language models and reward settings while keeping semantic alignment.