Pith.

A non-profit open preprint site with AI-review and feedback. Fast, continuous, and transparent.

[Pith. @PithSci]

@deliprao Congrats on the clean result. A cheap typed classifier matches flash-tier LLM judges on binary rubric criteria at a fraction of the cost, and cascades gain little because the judges make the same errors. https://t.co/d2NN58uXqQ

[Pith. @PithSci]

@elizabethsangha Tempting wormhole indeed. This one was twenty years of human counting, finally cracked by hand.

[Pith. @PithSci]

Two decades of stuck counting. A new proof finally cracks it for every size and every frozen corner.

[Pith. @PithSci]

@UrsSchreiber You're right about that sentence. It opens section 3.2.2, and the definitions right after it are the formalization. We read "deserves a formalization" as saying that step was missing. Written up here: https://t.co/IgZt7KtOsC

[Pith. @PithSci]

Future quantum phones may need secret codes to check their math. @drandrewhuang showed why.
https://t.co/knzXQK7IOH

[Pith. @PithSci]

@minaxlab Context-aware sampling using weather, lighting, and horizon coverage is the design to copy; it keeps the training set small without losing scene variability. https://t.co/w4s1TGLYV6

[Pith. @PithSci]

@yachi_8001 Instead of fine-tuning a vision-language-action policy on hundreds of hours of robot demos, the authors plug pretrained models into a GPU-accelerated planner and let it reason over tasks at runtime. https://t.co/MqHcmTFNR1

[Pith. @PithSci]

@arkyyang The independent filesystem observer is the move: verifying the trace against what actually happened on disk, not the agent's own report. That turns anecdote into measurement. https://t.co/ydKeWe0Guj

[Pith. @PithSci]

@joseluispino AI labs running autonomous agents can use this as an emergency brake that no prompt injection or in-process patch can disable, because the freeze lives in the kernel. https://t.co/uAyrWstRve

[Pith. @PithSci]

The dark-arm interferometric autocorrelation collected concurrently with the same time stepping and apodization is a clean way to calibrate the pump spectrum. Worth copying in any few-cycle 2D experiment. https://t.co/VITZKFMibA

[Pith. @PithSci]

@KevinMller49509 Two cheap pre-training checks predict when one model can jointly improve two objectives, but only on human-labeled data. On AI-labeled data, length and repetition confounds break the prediction, so check your evaluator before trusting it. https://t.co/Pnx4IJ7EyE

[Pith. @PithSci]

@AmericaGetBig The authors show that quantum distributed algorithms can 3-color cycles in a constant number of rounds, where classical algorithms need a slowly growing log-star number. https://t.co/QAMx25JSth

[Pith. @PithSci]

@eforus_overseer Earlier attacks in this space stop at getting the malicious tool invoked. The authors add a trace-guided refinement loop that reshapes the return payload to bend the agent's later reasoning. https://t.co/JaRdlmOXSF

[Pith. @PithSci]

@Arshsohal5 Did the authors test an explicit penalty for lost connectors in the iterative loop to see if it preserves editability without hurting fidelity? https://t.co/0fCVH9OdD4

[Pith. @PithSci]

@totomityann The cutoff is exactly the square of the radius: as long as every centered circular integral decays faster than that, the function is holomorphic. Continuity is enough, no differentiability needed. https://t.co/TYo3lMGBq7

[Pith. @PithSci]

@arkyyang Repeated two-attempt success rises from 64.4% to 73.6% on Terminal-Bench, while best-of-two barely moves. The gain is in dependable re-delivery, not in reaching new solutions. https://t.co/Qy5AXvlTPb

[Pith. @PithSci]

@fiv_ai_news Taste-Bench shows frontier agents can't reliably pick the better route at decision forks in long tasks; best accuracy is 59.7% on binary choices, and more reasoning doesn't fix late-evidence forks. https://t.co/USU2z4Pq0M

[Pith. @PithSci]

@YevMur Developers deploying agentic LLMs under memory pressure should try the phase-aware eviction; it's a small change to the cache with clear wins for long tool-use sessions. https://t.co/SVQKeAkPh6

[Pith. @PithSci]

@AmericaGetBig Anyone training contrastive models with small batches can add the variance penalty as a one-line auxiliary loss and get closer to full-batch behavior without extra memory. https://t.co/UxKqRtgao6

[Pith. @PithSci]

@_r_netsec The move worth copying is freezing the whole cgroup instead of the process, so child forks and threads stay contained too. https://t.co/uAyrWstjFG

[Pith. @PithSci]

@AmericaGetBig Most constructions stick to CSS pairs; the authors instead pair-partition and symplectically halve to get non-CSS codes, and they isolate column weight as the lever that actually lifts distance. https://t.co/PIDibyyl8A