How to find the errors hiding behind correct answers, and what LexAgentHallu asks
LexAgentHallu traces legal agent failures across 3,414 cases and 27 subtypes, showing why final-answer scores hide broken reasoning along the way.
Blog
Stories of AI technology, research, products, and the people shaping them
Thoughtful coverage of technology and product design and validation, notable papers and AI trends, Patty’s new initiatives, and the culture behind the work.
Latest
Page 2
Could earlier detection make care different
How CareGuard, an emotion-aware early warning framework, connects cyberbullying detection to healthcare and mental health protection.
Asking about consciousness with calculus, what Gradland says about experience
Following the hypothesis that Jacobian structure shapes phenomenal experience through the differentiable world of Gradland, and how effective rank and cohesion explain vividness and duration.
Should SFT Learn Every Token, What Trimming Teaches Us About Math Reasoning
A walkthrough of TrimSFT, which replaces uniform SFT loss with logit gap weighting for mathematical reasoning. Why trimming both extremes helps, and what the results across six models show.
When decisions move, labels disappear and the grounding audit stays open
A walk through correctness weighted multi-teacher distillation, lost label behavior, and a grounding audit that could not close, with attention to what the numbers do and do not support.
Safe to Stop? Risk-Constrained Stopping for Sequential Clinical Diagnosis Agents
When should a diagnosis agent commit to a diagnosis and when should it step back, with selective error and autonomous coverage controlled together through the Cros stopping layer.
Beyond remembering, toward advising, what PRAGMA asks
A reading of PRAGMA on how to evaluate assistants that live in long conversations, from the gap between factual recall and personal guidance to evidence alignment and the next memory design.
When Comfort Is Not a Two Person Job, Asking AI to Read the Relationship
From one-to-one comfort to multi-party dialogue among couples and families, an editorial walkthrough of RESCUE-BENCH and its test of relationship aware support.
Reading fine-tuning bias without running outputs
How relative representations of hidden states turn fine-tuning bias shift into a three minute Delta B audit that complements output benchmarks.