YY Ahn

And Zhou et al. find that returns can eventually turn negative: at high budgets, models sometimes abandon an earlier correct answer. Easy problems peak much sooner than hard ones. arxiv.org/abs/2604.10739

When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling arxiv.org

But Snell et al. found that the best inference strategy depends on how hard the problem is; a fixed budget is wasteful because easy and hard prompts benefit from different amounts of computation. arxiv.org/abs/2408.03314

Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters arxiv.org

Many test-time-compute papers show why this is hard. In an extreme version, just appending "Wait" whenever the model tried to stop raised the benchmark. Often, just forcing more reasoning does work: arxiv.org/abs/2501.19393

s1: Simple test-time scaling arxiv.org

I think that's exactly what many skills are about. Many of github.com/yy/claude-sc... are basically such checklists to go through.

GitHub - yy/claude-scholar: Academic research tools for Claude Code — literature search, citation management, LaTeX checks, math verification, and submission preparation github.com

Neural Scaling Universality: If Exponents Are Fixed, Time to Understand Coefficients arxiv.org/abs/2606.25008

Neural Scaling Universality: If Exponents Are Fixed, Time to Understand Coefficients arxiv.org

"Leak, Cheat, Repeat: Data Contamination and Evaluation Malpractices in Closed-Source LLMs" arxiv.org/abs/2402.03927

Leak, Cheat, Repeat: Data Contamination and Evaluation Malpractices in Closed-Source LLMs arxiv.org

This paper looks cool: arxiv.org/abs/2605.23901 "We propose the Shannon Scaling Law, a unified theoretical framework that models LLM training as information transmission over a noisy channel ... The Shannon Scaling Law consistently outperforms classical scaling laws ..."

LLMs as Noisy Channels: A Shannon Perspective on Model Capacity and Scaling Laws arxiv.org

Sutton's "Bitter Lesson" keeps showing up in new corners. Latest example on data filtering for pretraining: arxiv.org/abs/2605.19407 "... with enough compute, the best data filter is no data filter." -- Btw, Welch Labs has a nice video on the "bitter lesson": www.youtube.com/w…

A Bitter Lesson for Data Filtering arxiv.org

www.seangoedecke.com/ai-models-ne... interesting. Maybe on the same line, it's also hilarious to see LLMs estimate that it'll take weeks to build something and then one-shot it within 5 minutes when you actually ask to build it.

AI models need moral support to make discoveries seangoedecke.com

"A lecture is the process whereby the notes of the professor become the notes of the student without passing through the mind of either." I think this may apply to lots of "LLM wiki" usage.