Why sampling token boundaries can improve robustness
A focused review of subword regularization: how alternative segmentations preserve text while changing training, and what translation evidence does and does not establish for LLMs.
Research notes
One precise question at a time.
A public notebook on AI safety, interpretability, and the mechanisms inside language models. Each memo stays narrow enough to leave a useful answer—and an honest open question.
February — October 2026
The full archive.
Why sampling token boundaries can improve robustness
Why exact attention does not need a quadratic matrix
Why content-free inputs improve few-shot classification
Why scheduled sampling can learn the wrong distribution
Why preference optimization can reward sycophancy
Why long-context LLMs get lost in the middle
Why label smoothing helps beam search but hurts distillation
Why simple weight arithmetic can merge fine-tuned language models
Why soft prompt tuning only catches up at scale
Why the residual stream is a shared linear workspace
Why chain-of-thought can be causally unfaithful
Why function vectors can trigger in-context tasks
Why sparse MoE layers scale without proportional compute
Why goal misgeneralization is not just overfitting
Why many emergent abilities are measurement artifacts
Why LLMs memorize rare repeated strings
Why the tuned lens beats the logit lens
Why RMSNorm works without mean-centering
Why weight tying helps language models
Why speculative decoding preserves the target distribution
Why Chinchilla scaling prefers more tokens than bigger models
Why open-ended decoding needs tail truncation
Why Constitutional AI can replace many harmlessness labels
Why muP makes hyperparameters transfer across scale
Why SwiGLU replaced standard Transformer MLPs
Why low-rank updates are enough for LLM adaptation
Why attention sinks stabilize long-context decoding
Why self-consistency decoding improves reasoning
Why next-token prediction creates the reversal curse
Why induction heads emerge as a phase change
Can hidden states reveal what an LLM thinks is true?
Why TruthfulQA shows inverse scaling
DPO as implicit reward optimization under KL constraints
Pre-LN vs Post-LN: why layer norm placement stabilizes Transformers
Hard attention and Turing completeness in Transformers
The softmax bottleneck in language modeling
ALiBi vs RoPE: positional bias and length extrapolation
FFN layers as key-value memories for factual recall
Attention as modern Hopfield memory
In-context learning as implicit gradient descent
Grokking as a phase transition in transformer training
Sparsity as a control knob for superposition
Evaluating IOI circuits: faithfulness, completeness, minimality
Bootstrapping a weekly research memo
No notes match that search. Try a broader term.
Topic log
The latest narrow question.
- October 12, 2026: why sampling alternative subword segmentations can improve robustness while preserving text, and what remains untested for decoder-only LLMs.
- October 4, 2026: how online softmax computes exact attention without retaining a quadratic weight matrix, while preserving dense pairwise arithmetic.
- September 27, 2026: when content-free probes remove prompt-induced label bias in few-shot classification, and when they erase a useful prior.
- September 21, 2026: why scheduled sampling can improve sequence generation while learning the wrong distribution through mismatched prefixes and targets.
- September 6, 2026: why preference optimization can reward an assistant for matching a user's beliefs rather than preserving truthful disagreement.
About the series
A weekly habit, not a claim of completeness.
The format forces one narrow question, a small set of primary sources, explicit limitations, and one open question worth carrying forward.