TAIS method verifies speculative decoding safety at zero temperature with 48,072 samples showing no safety divergence.

Sahil Kadadekar

Grad Detect leverages layer-wise gradients for single-pass hallucination detection, outperforming confidence baselines with 94-99% accuracy.

Anand Kamat, Daniel Blake, Brent M. Werness

Using a fitted no-repetition scaling law, quantifies systematic damage caused by repeated data, revealing an intermediate repeat count causes maximum loss, with damage scaling as a power law with model size.

Jessica Chudnovsky, Joshua Kazdan, Noam Levi et al.

Introduced TR-CIE sampler to improve sampling quality in discrete flow matching under limited function evaluations.

Feiyang Fu, Hehe Fan

Introduces blockwise policy-drift gating, improving OPD pass rates in math reasoning tasks.

Liwen Zheng, Haiyun Jiang

Proposes Success Visitation Matching (SVM) rewards, transforming sparse rewards into dense signals, doubling RL finetuning speed in robotic tasks.

Raymond Tsao, Andrew Wagenmaker, Sergey Levine

GRINQH uses input-dependent dynamic per-channel quantization with hierarchical bit-plane storage, outperforming fixed-precision methods at 3-4 bits, enabling 2-bit generation.

Jette Oberländer, Jan Finkbeiner, Catherine M. Schöfmann et al.

Diffusion-LLM integrates diffusion models with LLMs for ultra-long-term forecasting, improving accuracy by 19.26%.

Falguni Ghosh, Vahid Hashemi, Bernhard Kainz

Leveraging AutoML and multi-objective HPO to optimize Deep Shift Neural Networks, achieving 20% performance improvement and 60% emission reduction.

Leona Hennig, Marius Lindauer

Substitution-based analysis reveals 81-92% of AI-generated crystals are duplicates or substitution-derived.

Masahiro Negishi, Aron Walsh

Using random walks on graphs to study parallel samplers in masked diffusion models, introducing a bisection sampler to improve speed and quality.

Vansh Bansal, Cho Cholyeon, Syamantak Kumar et al.

HACO system uses MaskGIT for crystal structure prediction, achieving 79.06% METRe accuracy.

Kiyoung Seong, Nayoung Kim, Sungsoo Ahn

TreeProp enables model-parallel deep learning with tree structures, achieving O(log N) time complexity.

Neeraj Mohan Sushma, Aditya Nagarsekar, Cabrel Teguemne Fokam et al.

Proposes an objective-behavior alignment diagnostic workflow using behavior space analysis to reveal hidden behavioral differences in MORL.

Antonio Mone, Zuzanna Osika, Florian Felten et al.

Proposes graph-bound execution-state capsules for low-latency, small-batch on-device AI, enabling byte-exact snapshot and restore with sub-millisecond GPU performance.

Liang Su

Proposes Marginal Advantage Accumulation (MAA) to address cross-batch evidence aggregation, boosting offline knowledge distillation efficiency.

Mingyu Yang, Keye Zheng, Congchao Cheng et al.

Introduces inverse dynamics regularization in sensorimotor world models (SMWM), preventing representation collapse and improving control in offline, reward-free settings.

Petr Ivashkov, Randall Balestriero, Bernhard Schölkopf

Proposes Piecewise Polynomial Interpolation-based Gradient Descent (PPI-GD) achieving oracle complexity of O((p/ε)^{d/(2ℓ)}) for data dimension d=O(log^{0.49}(n)), outperforming classical GD/SGD.

Dongmin Lee, William Lu, Anuran Makur

Diffusion-Proof framework uses dLLMs for formal theorem proving, achieving a 6.14% improvement over AR models.

Ruida Wang, Rui Pan, Pengcheng Wang et al.

Proposes STARE, a surprisal-guided advantage reweighting method, stabilizing policy entropy and improving accuracy by 4%-8% on models from 1.5B to 32B.

Haipeng Luo, Qingfeng Sun, Songli Wu et al.