SAPO matches PPO on benchmark tasks while using less memory
A preprint reports that SAPO matched or slightly exceeded PPO on ALFWorld and WebShop while using less peak GPU memory and less time per training iteration.
Scientific publication4 min read
Specialty
Applied AI
126 published articlesPart of Artificial Intelligence
Latest in Applied AI
AI Training Filter Reports Higher Math and Code Scores
An arXiv preprint reports higher math and code benchmark averages for a reward-aligned trajectory filter, with total training time close to standard on-policy distillation.
Scientific publication4 min read
AI controller keeps GSM8K accuracy near full budget with far less reasoning
A preprint presents AERA, a controller that estimates whether more reasoning may help and reports near-baseline GSM8K accuracy with sharply lower generation use.
Scientific publication5 min read
openJiuwen posts higher scores than selected systems on two coding tests
An open-source harness for long-horizon coding agents scored above selected systems on two benchmarks, but the study did not isolate the harness from model, prompt, tool and implementation differences.
Scientific publication4 min read
AI training method reports gains on harder mathematics problems
A method called DA3PO was reported to outperform GRPO, DAPO and GSPO on mathematical reasoning benchmarks, though the tests covered only two Qwen3 base models and did not report uncertainty estimates.
Scientific publication4 min read
AI agents that adapt teamwork report higher reasoning scores
A preprint describes an AI system that retrieves evidence before choosing how its agents collaborate. It reports higher scores across seven reasoning benchmarks, but the evidence comes from in-silico evaluations rather than human users or real-world deployments.
Scientific publication4 min read
AI Control Room Test Tracks Commands, Replays and Agent Moves
A preprint reports that GOD linked browser commands to execution records, replay checks and portable packs in 15 completed run slots. The evaluation focused on command, replay and artifact checks, not human or social validity.
Scientific publication4 min read
AI models miss subtle power cues in French and Egyptian Arabic dialogue
A preprint examining French and Egyptian Arabic movie dialogue finds that six AI systems align more closely with humans on visible social cues than on subtle power relationships, while multimodal results are limited by incomplete coverage.
Scientific publication4 min read
Overnight Bed Sensors Offer Modest Clues to Next-Day Agitation
The model with the highest pooled scores retained minute-level overnight patterns, but the small hospital sample and limited calibration do not support individual care use.
Scientific publication4 min read
Researchers Propose an Operating System for Shared Human-AI Work
The paper presents CrabOS as an operating-system approach to human-AI handoffs, using shared text objects and common capability controls, but reports architecture and case studies rather than measured evidence of better results.
Scientific publication4 min read
Scientific Text Reconstructs Much of Reactome's Global Shape
A text-only reconstruction of Reactome preserved its large-scale network pattern, but the study measured graph similarity rather than biological correctness.
Scientific publication3 min read
New AI decoder guarantees valid syntax, but task scores still vary
The preprint reports perfect syntax across the tested settings, while task-accuracy scores for SQL and linear temporal logic remained uneven.
Scientific publication4 min read
AI routing was more accurate when user profiles mattered
A benchmark study found the largest routing gap when user-profile details changed which skill best matched a task, while profile-insensitive cases stayed similar.
Scientific publication3 min read
Flow Matching outperforms diffusion in sparse-view CT test
A preprint reports stronger image-quality results from Flow Matching than diffusion on sparse-view chest CT, with FlowDPS leading the reconstruction comparison.
Scientific publication4 min read
AI System Reports Higher Scores With First-Error Tracking
A diagnose-then-correct system called DoCtOR reported higher success rates on HotPotQA, ChartQAPro and Mind2Web. Its attribution model was less accurate on held-out failures, and open-ended, multilingual and larger-team settings were not tested.
Scientific publication4 min read
Study reports lower blood-pressure errors using buffered sample selection
A buffered sample-selection method called RECAST reported lower blood-pressure estimation error than TTC and no adaptation on PulseDB and MC-MED. The results came from simulated streaming benchmarks, with extra but sub-second latency and no prospective clinical evaluation reported.
Scientific publication5 min read
A new sequence encoding keeps time shifts exact in one test
A methods preprint compares three real-valued position encodings, finding similar classification averages and an exact, faster algebraic shift for its Sinusoid variant.
Scientific publication4 min read
Study reports a more targeted way to select what language models forget
GRACE, a gradient-guided method for selecting compact forget and retain sets, reported higher forget-sample retrieval accuracy than RASLIK and the highest model utility in six of eight algorithm-model combinations, while forget-quality scores varied little across selectors.
Scientific publication4 min read
AI agents look beyond top hotel listings, but rank still matters
A randomized experiment found that list position still shaped which hotels AI agents inspected, but much less than in a human benchmark; effects on final booking varied by model and were no longer statistically significant at the highest tested reasoning effort.
Scientific publication4 min read
Semantic Overlay Shows Sharp Drop in Prompt-Injection Success
A preprint reports higher separation between marked content and instructions in frozen-model tests, with utility nearly unchanged and copied passages remaining highly similar.
Scientific publication3 min read
Targeted neuron tests flag sharp drops in LLM code scores
A computational preprint reports sharply lower code and math scores in targeted MLP-neuron suppression tests, while several non-target benchmarks retained much higher scores.
Scientific publication5 min read
AI Router Balances LLM Quality and Cost Across Four Tests
ProgRouter reported strong task and citation results while tracking operating-cost budgets, GPU energy use, and execution time across four benchmark streams.
Scientific publication3 min read
Rerunning AI agents often changes the outcome, not the failure
A study of 536 failed web-task runs found that whole-trajectory reruns often produced different outcomes without showing that the recorded failure mechanism had been corrected.
Scientific publication4 min read
AI Agent Study Finds Smarter Testing Can Beat More Rollouts
A computational evaluation reports that HarnessLens improved average held-out performance by 7.6% to 13.6% across three agent harnesses and four benchmarks, while using a smaller configured interaction budget than the baselines.
Scientific publication5 min read
AI editing method reports less collateral drift in benchmark tests
A preprint reports that KLOD can update targeted facts inside language models while keeping edit success near 100 percent in benchmark tests, although broader generalization and preservation of unrelated predictions remain in tension.
Scientific publication5 min read
Spiking language model gains ground in tests without reported collapse
A training method for autoregressive spiking language models improved reported benchmark averages over matched knowledge-distillation checkpoints and stayed non-collapsed in a 10-seed stress test. The work remains an early computational evaluation, with no hardware measurements or evidence about human language performance.
Scientific publication5 min read
Model-based ad incentives posted higher profit in online tests
A preprint describes an offline model-based reinforcement-learning framework for cost-sensitive incentive allocation. It reports higher net profit in offline and online comparisons, while diagnostics show weaker evaluation under policy drift.
Scientific publication4 min read
Machine-learned sepsis score separates ICU survivors from non-survivors
A score learned from ICU trajectories distinguished survivors from non-survivors across four baseline-severity groups, yet its patient-level behavior varied between two hospital systems.
Scientific publication4 min read
AI chemistry system scores highest at recovering past ideas
A preprint reports that MAIL, an automated literature-based system for generating chemistry hypotheses, posted the strongest listed overlap scores on the 51-paper TOMATO-Chem benchmark. The result shows reference-idea recovery, not demonstrated chemical discovery.
Scientific publication4 min read
Health AI papers often omit how their agents are built and governed
AGENT-O found papers reported evaluation more fully than system architecture, governance and reproducibility in a 279-record assessment.
Scientific publication4 min read
Edge Skillguard reports zero wrongful actuation at a software boundary
A live edge testbed evaluation recorded zero wrongful actuation and complete benign success for the full Edge Skillguard policy, with the outcome measured at the software handoff to device adapters.
Scientific publication3 min read
MO-IKE Posts Higher Frozen-Model Scores, but Retention Lags
A preprint reports higher MO-IKE results than comparison methods in several frozen-model benchmarks. Results vary by benchmark, with retention remaining weaker than other measures in key tests.
Scientific publication3 min read
Best AI Model Gets 33.7% of Olympiad Physics Problems Right
On the benchmark, Grok-4.2 led with 33.7% final-answer accuracy, while the study's small student comparison reached 48.5%.
Scientific publication4 min read
Claude Opus Leads Preprint Test of AI Bug Discovery
Claude Opus 4.8 led a benchmark of open-ended crash discovery, while no crash was triggered by any of the three models in 13 challenges.
Scientific publication3 min read
Five-Model Test Reports AI Tool-Call Rates Above 90%
A computational preprint tested one internal direction across five language models, reporting a wide range of tool-call rates, transfer to held-out tools and a live-search cost-accuracy trade-off.
Scientific publication4 min read
AI system scores 97.5% in early hardware compatibility checks
A seven-design evaluation found that DEVICES was more accurate than two one-shot prompting baselines and used an 8.6-fold smaller input context, but it tested only documentation-level compatibility.
Scientific publication3 min read
AI Model Combines CT Scans and Records to Classify ILD
A retrospective registry evaluation found that a hierarchical model combining lobe-aware CT analysis with selected EHR groups reached a mean AUC of 0.8750, slightly above imaging-only REN, but the difference was not statistically significant.
MICCAI Machine Learning in Medical Imaging (MLMI 2026)5 min read
Shared AI Model Improves 10-Second Traffic Forecasts at Five Sites
A modeling study of five signalized intersections found that pooled training produced lower 10-second trajectory errors than site-specific models at every site. Performance was less consistent at shorter horizons, and fine-tuning helped some held-out sites but not all.
Scientific publication5 min read
AI agents fell short on deep, large-scale biology tasks
A benchmark of 13 frontier AI models found partial success on prescribed computational-biology workflows, with performance lower on the deepest analyses and largest raw-data tasks.
Scientific publication4 min read
Financial AI models rank differently across risk-review tasks
A preprint introduces a benchmark that tests financial language models on distinct review operations and on decisions about when to request more evidence.
Scientific publication4 min read