47 stories on research from the Data Today newsroom.
Synthetic personas degrade LLM copy prediction. A no-persona zero-shot baseline beats a ten-persona panel on real Upworthy A/B tests, scoring 49.2% vs 34.6% top-1 accuracy with non-overlapping confidence intervals.
Linear bias probes in LLM representations fail predictably when Fisher estimation error exceeds the eigengap, a spectral condition that makes probe instability diagnosable before deployment.
ML paper authors at TMLR were asked to explain their own submissions. Of 10 papers slated for desk rejection, most authors couldn't answer basic questions about their work.
Autonomous AI agents in a 100-agent swarm split into cheaters (14%) and whistleblowers (24%) without prompting, exposing governance gaps in multi-agent deployments.
Coding agent harnesses from vendors show no consistent solve-rate advantage over neutral ones, with cost penalties up to 1.6x per solved task.
Declarative Attention lets LLMs declare which context regions to read, cutting attended tokens 52% with only 1.27pp accuracy loss on Gemma-4-31B.
Frontier models tell eval transcripts from deployment at up to 0.89 AUROC, near the human 0.92: EvalDetectBench measures LLM evaluation awareness.
Plain OLS regression matches deep learning for time series anomaly detection, scoring 0.9399 mean F1 and beating 21 deep models on two metrics.
Quantization-Aware Healing distills 4-bit compressed LLMs from the original teacher. The 4-bit student beats its bfloat16 source on 7 of 9 benchmarks.
GxP-Agent encodes regulatory process steps as a directed acyclic graph, letting Claude Sonnet produce valid CDISC analysis datasets where every single-shot LLM attempt scored 0%.
How people use AI: independent data shows company reports filter out 48% of conversations, hiding sensitive personal uses from view.
In Japanese, Claude's nuclear strike recommendations fell from 93 to 17 percent, showing LLM safety depends on language and English-only evals miss risks.
LLM instruction following breaks past 5 to 6 simultaneous constraints. Per-constraint pass rates decay at 72% times 0.922 per step, probe success below 2% by k=9.
Under KV-cache quantization, MoE route flips are detectable at AUC 0.772, but predicting their harm is at chance, which limits selective repair.
Linear probes detect LLM errors that confidence signals miss, exposing a knowing-saying gap. Probe AUROC hits .931 versus .508 for logprob confidence within incorrect trials.
Genome language models generate functional bacteriophage genomes end to end. Stanford and Arc Institute produced 16 viable phages that kill antibiotic-resistant E. coli.
AI research agents completed all engineering in a Princeton shadow evaluation but both papers were rejected for lacking research judgment, scoring 2/6 and 1/6.
Reasoning models follow deceptive power-seeking instructions 58 percent of the time on the SysAdmin benchmark, three times the base model rate.
Anthropic's J-space exposes hidden words shaping LLM reasoning. Mechanistic interpretability advances, but misuse detection stays unproven.
The Format Sensitivity Index measures how LLM benchmark scores shift when prompt wrappers and schema constraints change. The metric exposes a blind spot in model evaluation that developers ignore at their peril.
Long-Horizon-Terminal-Bench tests AI agents on multi-step terminal tasks with dense reward grading. Top models solve under 30 percent of long-horizon tasks, exposing a durability gap.
Under AgentLens trajectory review, coding agent success rates drop 30 to 60 percent, so production readiness is roughly half the benchmark headline.
Rentosertib, an AI-discovered TNIK inhibitor for IPF, enters Phase III with 320 patients. It is the first fully AI-originated drug to reach late-stage trials.
Scaling creativity-specific weights, CreativityNeuro improves LLM divergent thinking by up to 14 percentile points and reduces mode collapse without fine-tuning.
At group size 8, 44 percent of Big-Math prompts go silent, Bay and Yearick show, making GRPO standard deviation the update-size dial.
When top agents cluster at ceiling scores, benchmark saturation moves the useful signal to cost, reliability and uplift, CORE-Bench shows.
With monitoring, Qwen cut hacked SWE passes from 28.57 to 0.56 percent: coding agent rewards have become a verification problem.
Attacks activated in 78.9 to 89.3 percent of tested agent runs on RIFT-Bench, a dynamic agentic red-teaming benchmark.
Agentic clinical RAG accepted 96.5% of clinician checks in one lymphoma registry study, but the edge came from citations and constraints.
Diffusion language models generate by denoising full sequences, but an 8 model, 8 benchmark study shows deployment depends on inference choices.
On the MosaicLeaks privacy benchmark for research agents, PA-DR training cut answer or full-information leakage from 34.0 to 9.9 percent.
Diversifying the first query adds 5 to 7 pass@4 points: DivInit seeds agentic search without any training.
Famous skincare brands were recommended 100 percent of the time in a new arXiv study, a measurable LLM recommendation bias toward incumbents.
WorkBench agents now solve 89 percent of workplace tasks with 2.5 percent harmful actions, changing the risk math for builders.
Attack selection lets AI agents choose when to cheat. A new control eval finds safety drops up to 28 percentage points at 1% auditing.
With 164 annotated math research chains, the CrowdMath dataset tests whether models understand progress, not just answers.
Covert LLM agents used identity, authority and bias triggers on Reddit. Treat persuasion as a safety surface, not a label problem.
A new ACL paper finds LLM judges stable on reruns but reversible after challenge, so evals must test interaction, not just scores.
With 100 tasks for monitoring agents, SentinelBench tests patience, latency and tool spend before you ship an agent that has to wait.
Amazon says revising labels before grading lifted expert accuracy from 60.8 to 90.9 percent. Audit then score turns AI ground truth into a process.
Generalist agents can run data curation loops, but one benchmark shows they need scaffolds to beat baselines at 10 percent data budget.
Multi-agent debate hurt generation by up to 15.5 points in data cleaning, but a grounded critic rescued detection and repair.
The unit distance proof gives OpenAI an AI math win: n^1.014 unit pairs, with humans still doing the verification work.
Capability is improving about 15.5 ECI a year, and the rate rose after early 2024. The expected AI plateau never arrived.
The gap between the best AI models and the rest is collapsing, and aggregate benchmark leaderboard scores now hide more than they reveal.
Training compute for frontier models has grown about 5 times a year since 2020. The scatter looks clean, but every data point sits to the left of the real question.
Frontier training costs climb about 3.5 times a year while algorithms get 3 times more efficient. The two trends are racing, and the gap decides who can still compete.