A new cross-architecture study shows that the same task is solved by completely different internal circuits across Pythia, OLMo, and OLMoE -- breaking a core assumption of mechanistic interpretability
A new paper from Iowa State and Thoughtworks shows that teaching LLMs what NOT to do — using synthetically generated wrong reasoning paths — dramatically boosts multi-table question answering accuracy.
A qualitative study of 17 developers reveals that overseeing AI coding agents is a four-part, full-lifecycle job — and that most of it happens before the agent writes a single line of code.
PersuasionTrace tracks human belief changes turn-by-turn during LLM conversations, revealing that vanilla LLM simulators fail to replicate how real people update beliefs under persuasion
Researchers from Alibaba and NTU show that web-searching AI agents can accidentally look up their own test answers, inflating benchmark scores by up to 4%
A multi-institution research team unifies visual token pruning, merging, pooling, and reweighting into one differentiable search space — and lets gradient descent pick the best operator automatically.
Researchers from UW-Madison and Adobe introduce camroll, a 31K-image benchmark and agent for answering personal questions over your entire photo library
GeM-NR is a training-free method that propagates a single edited image to all views of a 3D scene, even when the edit drastically reshapes the geometry.
A new framework from Mila turns failed LLM reasoning traces into a diagnostic signal, routing compute to the right fix and boosting recovery by 12.2% on hard failures — without touching model weights.
AgentMob skips LLM calls for routine trips and only invokes iterative tool use when location signals are genuinely ambiguous, hitting 71% top-1 accuracy on GPS data
Researchers show that formal, declarative agent protocols can interoperate with Google's live UCP checkout infrastructure, offering a rigorous alternative to ad-hoc agentic design.
A new arXiv paper proposes self-reflective APIs that return machine-readable repair hints on failure, lifting agent task-completion by up to 40 percentage points
KINA spans 261 academic disciplines and exposes a tiered frontier where even the best model, Gemini, scores just 53% -- leaving massive room for improvement