Skip to content

2026-09-27 | Maxwell

Goal:

Summary:

Work sessions

In Out Task
09:50 10:05 Plan Tasks for the Day
10:05 11:30 Paper Reading
13:00 13:30 Battle NLP cluster setup
13:30 15:30 Understand Grammar Failures and planning tasks
16:00 16:30 TAing: Fix lab handouts for TA session on Monday
16:30 17:30 Design New Grammar + PRs
17:30 18:15 Talk to ZZ about memory and delta net
18:45 19:20 TAing: Print out handouts for lab
20:00 20:45 Chat with ZZ about society and AI, what should a society with AI look like in the future?
20:45 20:45 New Grammar with Determiners and postpositions, decide to punt PRs to later for review

Goals

  1. Reading
    1. Read Saphra [2506.15872] Hidden Breakthroughs in Language Model Training
      • POLCA (Projection Oriented Loss Change Allocation) is a modification of LCA (Loss Change Allocation) which disaggregates the loss by (1) each datapoint and (2) an orthonormal basis determined by the eigenvector w/ the largest eigenvalue to do unsupervised clustering/classification of data points and how they contribute to the loss
      • I think this is pretty interesting for big models where you don’t have an a priori understanding of what exact skills you are looking for. I think for now, plotting a phase transition for different syntactic structures is a very good first order approximation of this
    2. Read Multilingual morphology inflection paper https://arxiv.org/abs/2410.09223v1
      • There’s shared circuitry cross linguistically for Tense Marking (past tense) and IOI
      • Takeaway: Using path patching seems likes a good idea
    3. Temporal Recurrence: https://arxiv.org/abs/2609.12531 (thanks Tony for sharing)
      • Cool: trade layer depth for time
    4. Variable Binding: https://arxiv.org/abs/2505.20896
      • Variables and their values live in different subspaces
      • Takeaway:
        1. Training dynamics shows that heuristics are used early on and continued to be used later too if the value appears early in assignment
        2. Use causal abstraction to check flow of variable assignment (very lookbacks-esque) but never vary the syntax of assignment
    5. Chris Potts lexical semantics
  2. Followup Grammars + Training Dynamics
    1. Design Grammar to break the symmetry rule
    2. Plot Phase transitions of different constructions
  3. Interp work
    1. Consider path patching or other technique for interp
  4. Evaluations
    1. Evaluate BLOOM or OLMO3.1 on a japanese translation benchmark

Interesting things

  1. LSTM nested dependencies: https://arxiv.org/abs/2006.11098 (interesting because LSTM fails on some long distance dependencies but perhaps this is expected since Attention is supposed to fix these long context information losses)

Meetings

  • None

TODO: better understand how physics of LLMs does the CFG parsing (I think I made a Claude handout on this before)