Technology & AI
AI reasoning models fail from self-doubt, not capability gaps
New arXiv study finds reasoning models fail tasks they can solve due to premature self-doubt. The fix may not require bigger, more expensive models.
AI Models
165 stories
Technology & AI
ISPCloak fools deepfake detectors by faking camera noise
ISPCloak projects AI images through a simulated camera pipeline, imprinting sensor noise that deepfake detectors cannot tell from genuine photographs.
Technology & AI
FineServe shows LLM serving traffic varies by model type
FineServe, a new dataset from a commercial LLM marketplace, reveals serving traffic varies fundamentally by model architecture and task type.
Technology & AI
OPTScientist: AI agents discover new transformer optimizer
OPTScientist uses four AI agents to discover optimizer algorithms for transformer pretraining, including a new reduced-state matrix optimizer called RS-MR.
Technology & AI
Adversarial Frontiers paper rethinks AI robustness testing
Adversarial Frontiers, a new arXiv preprint, argues single-budget AI robustness rankings are unstable and proposes a frontier-based evaluation framework.
Technology & AI
Language models overgeneralise and break on nested rules, study finds
Language models build shared rules instead of storing individual facts, KAIST AI finds, and the overgeneralisation affects teams fine-tuning on nested
Technology & AI
HyGRL targets multi-entity RAG with hybrid graph reasoning
HyGRL, a new arXiv paper from Beijing Institute of Technology, blends text with knowledge graphs to answer multi-entity questions that defeat standard RAG.
Technology & AI
CARGO beats trained LLM routers without training data
A new arXiv preprint from Purdue and Princeton shows LLM query routing can work without training data, using model self-agreement as a signal.
Technology & AI
LLM agents pass 100% JSON schema checks, still fail 1 in 5 orders
New OrderBench benchmark runs 2,400 calls across four open models and finds schema validity alone doesn't guarantee semantic correctness.
Technology & AI
AI agents turn robot videos into physics simulations
Agentic Real2Sim uses vision-language agents to convert real robot recordings into simulatable twins, aiming to cut the labor cost of training robots in
Technology & AI
JAXBench: curated TPU docs lift AI kernel correctness 5.8% to 37%
JAXBench, the first TPU benchmark for AI kernel generation, shows curated docs lift correctness from 5.8% to 37.3% and reach 1.6x speedup over XLA.
Technology & AI
LLM confidence scores fail basic coherence test
New arXiv paper finds calibration, the standard test for AI confidence, misses deeper incoherence in how models estimate their own certainty.