Tessera
Folgen

Lex

@notesbylex.com

Senior MLE at @Canva. Full-stack developer. Talks about Software dev, LLMs, Agentic Reasoning, Obsidian, note-taking, dog photos.

71 Posts89 Follower11 Folgt
Lex@notesbylex.com · 2h ago

LLMs are still the wrong tool for tabular data "no amount of prompt polishing aimed at for- mat, precision, or batching should be expected to fix high-dimensional in-context prediction, and purpose-built tabular models remain the right tool" arxiv.org/abs/2608.02412

Why Large Language Models Fail at Tabular Prediction

Large language models (LLMs) have become the default tool for a remarkable range of tasks, yet they have had conspicuously little success at one of the most common machine learning workloads:...

arxiv.org

Lex@notesbylex.com · last wk.

Published a summary of the paper "Ask, Don’t Judge: Binary Questions for Interpretable LLM Evaluation and Self-Improvement". It outlines BinEval - an LLM-Judge approach that decomposes scores into binary subquestions, making them interpretable and actionable. notesbylex.com/decomposing-...

Decomposing LLM Judge Scores Into Yes/No Questions

Most LLM judges output a single score for each criterion. However, a score can often be opaque and difficult to act on, especially at the higher end: what should you change to get a 5 instead...

notesbylex.com

Lex@notesbylex.com · last mo.

I finally completed a BSc in Computer Science on Coursera via the University of London, after a 20-year degreeless career in tech. I put together a little write-up of my experience. notesbylex.com/completing-a...

Completing A Computer Science Degree On Coursera

In September 2022, I impulsively signed up for a Bachelor's Degree in Computer Science after seeing an ad for it on Coursera. Getting a degree had been on my mind for quite some time, after a...

notesbylex.com

Lex@notesbylex.com · 3mo ago

Published some notes on the paper "HEAVYSKILL: Heavy Thinking as the Inner Skill in Agentic Harness" - they propose a skill for test-time scaling that follows a two-stage model - stacking for reasoning LLMs. notesbylex.com/heavy-thinki...

Heavy Thinking: A Test-Time Scaling Pattern for Hard Problems

Overview This paper is an empirical investigation of an Agentic Reasoning approach they call Heavy Thinking. It's a simple two-stage workflow: a bunch of subagents reason independently in...

notesbylex.com

Lex@notesbylex.com · 3mo ago

Published some notes on the paper "LLMs Corrupt Your Documents When You Delegate" - even with frontier models, users would see an average of 25% document corruption by the end of a long workflow notesbylex.com/llms-corrupt...

LLMs Corrupt Your Documents When You Delegate

My notes on LLMs Corrupt Your Documents When You Delegate by Philippe Laban, Tobias Schnabel and Jennifer Neville from Microsoft Research. An interesting paper from 3 researchers at Microsoft....

notesbylex.com

Lex@notesbylex.com · last yr.

Learning to Reason without External Rewards (aka Self-confidence Is All You Need) Turns out we can just use the LLM's internal sense of confidence as the reward signal to train a reasoning model, no reward model / ground-truth examples / self-play needed. Amazing. https://notesbylex.com/learning…