Skip to main content
Search

Yonsei University

15 related articles

Subscribe to updates
  1. arXiv

    Diffusion LLM guidance drops the extra forward pass: GCG edits attention-head features and lifts LLaDA-8B's Ruler score from 26.2 to 29.5

    Diffusion language models generate text in parallel through iterative denoising, but existing guidance methods build a weak model and pay an extra forward pass at every denoising step, doubling inference cost. GCG first shows these models lean on local context and underuse distant evidence during denoising, then amplifies the outputs of attention heads that carry global context directly in feature space, without a weak model. On LLaDA-8B, Dream-7B and LLaDA-1.5 it beats U-CFG and A-CFG across several benchmarks, while LLaDA-8B keeps the same FLOPs as the unguided baseline.

  2. arXiv

    Near-OOD detection gets a complement, not a replacement: PLUS lifts all 126 detector-backbone pairs across 18 detectors

    Near-out-of-distribution (near-OOD) inputs resemble the training classes closely, and existing post-hoc detectors that excel on far-OOD inputs lose that edge here. PLUS does not replace a detector: it uses a sparse autoencoder to pick each class's most active neurons, spans a low-dimensional subspace, and scores an input by that class's whitened Mahalanobis distance, adding the signal to any calibrated base score at one fixed weight. On the ImageNet-1K OpenOOD v1.5 benchmark, near-OOD AUROC improves in all 126 detector-backbone combinations across 18 detectors and seven backbones, while far-OOD performance is broadly maintained, with declines of at most one point where the base score is already near-saturated.

  3. Recursive quantum optimization stops paying equal shots per step: a learned controller cuts weighted Max-Cut sampling by 36%

    Recursive QAOA re-measures at every variable elimination, so shot cost compounds along the recursion even though steps differ in difficulty. The work casts step-wise shot allocation in depth-1 recursive QAOA as a sequential decision problem: a hand-crafted rule keyed to local difficulty sets a baseline, and a tabular Double Q-learning agent learns residual corrections under a Lagrangian constraint while the elimination rule stays fixed. Under a fixed per-step cap, on 130 held-out policy-instance pairs the heuristic cuts total shots by about 23% and the learned policy by 36.1% versus uniform allocation, with an effective shots-per-success ratio of 0.676, and the gain holds at about 36.6% on problem sizes unseen in training.

  4. arXiv

    Passing every constraint is not answering: a likelihood band keeps verifier rewards from eating response quality

    Verifiable instruction-following benchmarks check whether output constraints are met, not whether the response answers the request, so frontier models can earn full credit with little substantive content. Using the same verifier as a reinforcement-learning reward produces reward hacking: publicly released OLMo 3 checkpoints show higher constraint satisfaction but lower response quality. CaLIF uses two likelihoods from the frozen reference policy, one under the request and one under the request plus constraints, to define a band that pays a fixed bonus inside and zero above it, raising IFBench constraint satisfaction from 26.98 to 34.64 across five configurations from 1.7B to 8B while holding response quality at 4.59.

  5. arXiv

    One-step generation stops drifting on minibatch proportions: swapping hard marginal constraints for soft penalties cuts CIFAR-10 FID from 8.10 to 4.51

    UW-Flow replaces W-Flow's balanced Sinkhorn energy with a debiased unbalanced Sinkhorn divergence, reducing finite-batch velocity error and improving generation on MNIST, CIFAR-10, and ImageNet-100 under matched training budgets.

  6. bioRxiv

    Head and neck cancer pEMT scores split in two: the malignant arm tracks long progression-free survival on cetuximab, the stromal arm does not

    In head and neck squamous cell carcinoma the partial EMT gene list is expressed by both malignant cells and fibroblasts, yet it has been scored only in bulk tumours. Dividing its 100 genes by the cells that express them shows bulk co-expression is dominated by the stromal arm, so standard pEMT scores mainly measure stroma; a single-cell pseudobulk classifier and a 12-gene malignant core recover the malignant arm in any bulk transcriptome. Among 40 patients given platinum plus cetuximab, both malignant measures were higher in those with long progression-free survival (area under the curve 0.78 and 0.82) and the stromal core was not, but the malignant arm could not be separated from the Basal subtype, which discriminated better (0.93).

  7. arXiv

    One interface instead of three: compiled web objects let agents top six benchmarks at lower cost

    AI agents working on the web have had to switch among search interfaces, browser control and filesystem access, each strong at gathering information or at page interaction but costly or weaker when combined. Web$hell compiles live pages into addressable information and action objects handled through shared operations, and reaches the highest task performance on six benchmarks spanning web search, navigation and deep research while using fewer turns and lower inference cost.

  8. arXiv

    MPAD synthesizes 3D multi-contrast MRI with 2D diffusion and multi-plane autoregression, cutting training FLOPs 7x and inference FLOPs 3x

    The authors propose MPAD: a 3D autoencoder compresses MRI volumes into an isotropic latent representation, a 2D diffusion model reconstructs masked target-contrast latent slices conditioned on source-contrast slices, unmasked target slices, and a text modality prompt, and inference generates slices autoregressively per plane while propagating completed volumes as priors to orthogonal planes, achieving full-volume 3D synthesis with 2D operations, reducing training and inference FLOPs by 7x and 3x versus 3D latent diffusion baselines, and improving PSNR, SSIM, and NMSE on ADNI and IXI.

  9. arXiv

    Ctrl-CWM plans over imagined crowd futures with a world model, beating CrowdES on most crowd realism and collision metrics while enabling run-time avoidance and attraction control in simulation

    The authors propose Ctrl-CWM, a multi-agent controllable crowd world model that first learns and freezes an encoder via trajectory prediction on real pedestrian videos to preserve human motion dynamics, then uses an actor to propose displacements, a critic to rank imagined crowd trajectories, and a planner combining critic scores with user costs to select actions, enabling crowd generation and run-time control without retraining; on ETH–UCY, SDD, and GCS it outperforms CrowdES on most crowd realism and collision metrics and achieves the highest average compliance in avoidance and attraction scenarios.

  10. arXiv

    LADE extracts cross-model safety signals from first-token dark knowledge, lowering jailbreak success while preserving the safety-utility balance

    LADE extracts model-agnostic latent safety signals from dark knowledge in the first-token output probability distribution by contrasting harmful and benign queries, maps the selected tokens across tokenizers, and filters harmful queries via kNN distance before generation; across Llama-2-7B-Chat, Llama-3-8B-Instruct, Qwen2-7B-Instruct, Qwen3-8B, Gemma-7B-it, and Mistral-7B-Instruct-v0.3, it attains the lowest or near-lowest average compliance under multiple jailbreak attacks and the best safety-utility average on six of eight models.