Link: chiragnagpal.github.io/papers/llm_l...

16.10.2025 23:51 — 👍 0    🔁 0    💬 0    📌 0

4/ There's still a lot of opportunity for statistical thinking in the Large Language Model era!

There's ample estimation problems in Language Modeling and AI Alignment waiting to be solved using classic statistical tools!

16.10.2025 23:50 — 👍 2    🔁 0    💬 0    📌 0

3/ Through illustrative examples on a toy data and simulation on a real-word chatbot style dataset, I show how censoring adjusted estimators result in better estimates of trajectory length distributions.

16.10.2025 23:50 — 👍 0    🔁 0    💬 1    📌 0

2/ In a new monograph, I show that is a classical statistical estimation problem called 𝘾𝙚𝙣𝙨𝙤𝙧𝙞𝙣𝙜.

I show that estimators like the 𝙆𝙖𝙥𝙡𝙖𝙣-𝙈𝙚𝙞𝙚𝙧, popular in epidemiology and bio-statistics for analyzing patient mortality and survival can be adapted to this problem of trajectory length estimation.

16.10.2025 23:50 — 👍 0    🔁 0    💬 1    📌 0

𝙇𝙀𝙉𝙂𝙏𝙃 of generations from an 𝙇𝙇𝙈 is an important heuristic used in post-training to understand model behavior.

𝘽𝙐𝙏 due to a 𝙁𝙄𝙓𝙀𝘿 𝙎𝙄𝙕𝙀 𝘾𝙊𝙉𝙏𝙀𝙓𝙏 𝙒𝙄𝙉𝘿𝙊𝙒, a large number of trajectories get truncated before ever reaching [𝗘𝗢𝗦] token.

𝙃𝙊𝙒 𝙏𝙃𝙀𝙉 does one accurately estimate model generation length ?

16.10.2025 23:50 — 👍 1    🔁 0    💬 1    📌 1

Direct preference optimization is cox regression

16.09.2025 04:25 — 👍 1    🔁 0    💬 0    📌 0

Kernel regression is the only way to achieve AGI

11.09.2025 04:26 — 👍 2    🔁 0    💬 0    📌 1