Nanjing University of Science and Technology

Ancient-book pages often carry paper ageing, water-induced distortion, insect holes and mold at once, while existing public benchmarks are limited in scale and stack isolated degradations, which weakens supervised restoration. AncBench supplies 100,379 pixel-aligned clean–degraded pairs in which chemical, physical and biological damage are composed in sequence; trained on it, AncRestore first estimates page-wide and local degradation, then dynamically selects and distributes visual priors according to those cues. On the AncBench test set AncRestore reaches 26.70 dB PSNR, 0.9191 SSIM and 0.1346 LPIPS, with 89.58% character-recognition accuracy.

Ancient-book pages often carry paper ageing, water-induced distortion, insect holes and mold at once, while existing public benchmarks are limited in scale and stack isolated degradations, which weakens supervised restoration. AncBench supplies 100,379 pixel-aligned clean–degraded pairs in which chemical, physical and biological damage are composed in sequence; trained on it, AncRestore first estimates page-wide and local degradation, then dynamically selects and distributes visual priors according to those cues. On the AncBench test set AncRestore reaches 26.70 dB PSNR, 0.9191 SSIM and 0.1346 LPIPS, with 89.58% character-recognition accuracy.

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.

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.

Retrieval-augmented generation does not improve every task, since irrelevant evidence can distract a generator, and measuring each candidate retriever's gain directly repeats generation for every configuration. EviCast splits a retriever's net utility into the generator's response to controlled evidence states and that retriever's offline distribution over those states; the response is measured once on a small calibration set and reused. On 14 HotpotQA retrieval configurations, one calibration used 96 generation calls, made 14 of 14 correct deployment decisions with a 2.75 F1-point forecast error, and cut diagnostic generation calls by 80% relative to per-configuration direct measurement.

Retrieval-augmented generation does not improve every task, since irrelevant evidence can distract a generator, and measuring each candidate retriever's gain directly repeats generation for every configuration. EviCast splits a retriever's net utility into the generator's response to controlled evidence states and that retriever's offline distribution over those states; the response is measured once on a small calibration set and reused. On 14 HotpotQA retrieval configurations, one calibration used 96 generation calls, made 14 of 14 correct deployment decisions with a 2.75 F1-point forecast error, and cut diagnostic generation calls by 80% relative to per-configuration direct measurement.

GRPO, a common reinforcement-learning recipe for reasoning models, judges a rollout only by whether its final answer is right, so a correct answer can hide unreliable reasoning, and a single corrupted reward can strongly distort group-relative advantages when a response group is highly imbalanced. S-GRPO minimizes the expected squared error between observed and latent advantages under a reward-noise model to obtain a closed-form, noise-aware group weight. Across several mathematical reasoning benchmarks and model scales it stabilizes training and keeps making progress under controlled reward noise where standard GRPO shows little meaningful learning.

GRPO, a common reinforcement-learning recipe for reasoning models, judges a rollout only by whether its final answer is right, so a correct answer can hide unreliable reasoning, and a single corrupted reward can strongly distort group-relative advantages when a response group is highly imbalanced. S-GRPO minimizes the expected squared error between observed and latent advantages under a reward-noise model to obtain a closed-form, noise-aware group weight. Across several mathematical reasoning benchmarks and model scales it stabilizes training and keeps making progress under controlled reward noise where standard GRPO shows little meaningful learning.