AI and science frontiers · 2026-10-25

Pruning a benchmark because it correlates with another may discard exactly the signal that separates frontier models. Across 4,497 models on Open LLM Leaderboard v2, BBH and GPQA correlate at 0.861 overall, yet conditional Spearman falls to 0.434 among the top 20% and to -0.278 among the top 5%; 7 of 15 pairs are best fit by zero-tail-dependence Frank copulas, so high bulk correlation does not imply agreement at the frontier.

Pruning a benchmark because it correlates with another may discard exactly the signal that separates frontier models. Across 4,497 models on Open LLM Leaderboard v2, BBH and GPQA correlate at 0.861 overall, yet conditional Spearman falls to 0.434 among the top 20% and to -0.278 among the top 5%; 7 of 15 pairs are best fit by zero-tail-dependence Frank copulas, so high bulk correlation does not imply agreement at the frontier.

The work decomposes an LLM's refusal behavior into a cross-layer pathway: harmful detection heads in early and middle layers pick out harmful tokens, safety neurons in mid-to-late layers carry and stabilize that signal in the residual stream, and refusal heads turn it into a refusal. Scaling the weights of these three component groups without changing architecture raised average safety under GCG attacks from 43.2% to 69.7% across six models, with a 1.7% average accuracy drop on four standard benchmarks.

The work decomposes an LLM's refusal behavior into a cross-layer pathway: harmful detection heads in early and middle layers pick out harmful tokens, safety neurons in mid-to-late layers carry and stabilize that signal in the residual stream, and refusal heads turn it into a refusal. Scaling the weights of these three component groups without changing architecture raised average safety under GCG attacks from 43.2% to 69.7% across six models, with a 1.7% average accuracy drop on four standard benchmarks.

Medical vision-language models cave to authoritative prompts, and prompt-level defenses can restore the right answer while visual evidence use still collapses in late fusion layers. HD-AS changes no parameters: it ranks attention heads by activation drift between neutral and biased prompts, then damps those heads' output projections at inference. On SLAKE, Qwen2.5-VL-7B's mean conditional flip rate falls to 34.60% versus 38.29% for contrastive decoding, with neutral-prompt accuracy held at 74.37%.

Medical vision-language models cave to authoritative prompts, and prompt-level defenses can restore the right answer while visual evidence use still collapses in late fusion layers. HD-AS changes no parameters: it ranks attention heads by activation drift between neutral and biased prompts, then damps those heads' output projections at inference. On SLAKE, Qwen2.5-VL-7B's mean conditional flip rate falls to 34.60% versus 38.29% for contrastive decoding, with neutral-prompt accuracy held at 74.37%.

Large language models still fabricate reasoning that contradicts domain rules in specialized scientific tasks. SciDC has a general model compile natural-language scientific knowledge into three layers of rules, top-level reasoning steps, middle-level conditional checks and bottom-level token masks, which a locally deployed domain model enforces during decoding. Across industrial formulation design, thyroid cancer TNM staging and single-step retrosynthesis, average accuracy rises 12% over vanilla generation, and ablations show removing the top or bottom layer sharply lowers formulation accuracy.

Large language models still fabricate reasoning that contradicts domain rules in specialized scientific tasks. SciDC has a general model compile natural-language scientific knowledge into three layers of rules, top-level reasoning steps, middle-level conditional checks and bottom-level token masks, which a locally deployed domain model enforces during decoding. Across industrial formulation design, thyroid cancer TNM staging and single-step retrosynthesis, average accuracy rises 12% over vanilla generation, and ablations show removing the top or bottom layer sharply lowers formulation accuracy.

Masked diffusion language models decode block by block and normally never revisit a committed token, so choices made under sparse context stay fixed. CoDR compares each committed token's confidence at commit time with its confidence under denser later context, remasks only those whose confidence dropped past a threshold, and lets the original sampler refill them, estimating drift for all committed positions in k forward passes via k-partition probing. Across LLaDA-8B and Dream-7B, three base samplers and four reasoning and coding tasks, adding CoDR raises the four-task average accuracy of every evaluated model–sampler configuration, by 0.85, 1.06 and 1.44 points for top1, fd and eb on LLaDA.

Masked diffusion language models decode block by block and normally never revisit a committed token, so choices made under sparse context stay fixed. CoDR compares each committed token's confidence at commit time with its confidence under denser later context, remasks only those whose confidence dropped past a threshold, and lets the original sampler refill them, estimating drift for all committed positions in k forward passes via k-partition probing. Across LLaDA-8B and Dream-7B, three base samplers and four reasoning and coding tasks, adding CoDR raises the four-task average accuracy of every evaluated model–sampler configuration, by 0.85, 1.06 and 1.44 points for top1, fd and eb on LLaDA.

TCM language models have been trained on under 1GB of text and handle text alone, while TCM diagnosis itself draws on looking, listening, smelling and pulse-taking. ShizhenGPT pairs more than 100GB of text from 3,256 TCM books and web sources with over 200GB of multimodal data, including 1.2M annotated images, 200+ hours of audio and pulse, smell and ECG signals, through two-stage pretraining and instruction tuning. On five 2024–2025 TCM exams ShizhenGPT-32B averages 78.1 against 84.0 for DeepSeek-R1-671B, and on a 7,204-question TCM vision benchmark it scores 63.6 against 61.4 for Gemini-3-Pro.

TCM language models have been trained on under 1GB of text and handle text alone, while TCM diagnosis itself draws on looking, listening, smelling and pulse-taking. ShizhenGPT pairs more than 100GB of text from 3,256 TCM books and web sources with over 200GB of multimodal data, including 1.2M annotated images, 200+ hours of audio and pulse, smell and ECG signals, through two-stage pretraining and instruction tuning. On five 2024–2025 TCM exams ShizhenGPT-32B averages 78.1 against 84.0 for DeepSeek-R1-671B, and on a 7,204-question TCM vision benchmark it scores 63.6 against 61.4 for Gemini-3-Pro.

Hallucinations in large vision-language models partly trace to CLIP-style encoders that favor main objects and skip image detail. PLANE attaches three LoRA visual experts to CLIP, each trained on object, attribute or relation captions, and an instruction-aware gate fuses the four image representations at token level. Trained once on LLaVA-1.5 7B, it plugs into other models sharing the same encoder; on LLaVA-1.5 7B the MME hallucination subset total rises from 555.38 to 582.34 and the GIFT hallucination rate falls from 23.39% to 20.30%.

Hallucinations in large vision-language models partly trace to CLIP-style encoders that favor main objects and skip image detail. PLANE attaches three LoRA visual experts to CLIP, each trained on object, attribute or relation captions, and an instruction-aware gate fuses the four image representations at token level. Trained once on LLaVA-1.5 7B, it plugs into other models sharing the same encoder; on LLaVA-1.5 7B the MME hallucination subset total rises from 555.38 to 582.34 and the GIFT hallucination rate falls from 23.39% to 20.30%.