A business-focused reading of DRIFT-BENCH, showing why agent reliability depends less on asking more questions and more on knowing when clarification helps, when it harms, and when execution must stop.
A mechanism-first reading of why identity-bridge data can weaken the reversal curse in autoregressive LLMs—and why the useful trick is more delicate than it first looks.
A mechanism-first reading of why robust policy iteration for $L_\infty$ robust MDPs is not merely convergent, but strongly polynomial under fixed discount.
RAudit shows why longer reasoning, stronger judges, and harsher critique can reveal LLM failures—but can also amplify them.
A mechanism-first reading of MentisOculi, and why explicit visual thoughts still fail to become reliable reasoning evidence for multimodal AI.
A mechanism-first reading of DefameLM: why narrowly scoped small language models may be more practical than giant cloud LLMs for real-time game AI and some business automation loops.
A business-focused reading of Thinking with Comics, a paper arguing that comic panels may offer a cheaper and more structured middle path between static images and video for multimodal reasoning.
A mechanism-first reading of ThinkSafe, a self-generated safety-alignment method that restores refusal behavior in reasoning models without paying the usual teacher-distillation tax.
Self-contradiction in multimodal models is not just a failure signal; it may be a cheap diagnostic for aligning generation with understanding.
A close reading of TSAQA shows why turning time series into question-answering tasks helps evaluate LLMs—but does not magically give them temporal reasoning.