INDEPENDENT LLM RESEARCH · MELBOURNE

Maximum energy in one direction. Deliberate nulls everywhere else.

Main Lobe Labs studies how large language models actually behave under pressure: pruning, memorisation, and the gap between benchmark scores and mechanism. We publish everything, with code, controls, and stated confidence. Small lab, narrow beam.

Research index Read the agenda

ULA · d = λ/2 · θ₀ = 90°

θ₀N = 8

RESEARCH INDEX

2026-08-04
rev v4

Measured Pruning Damage Depends on the Evaluation Corpus: A Renaming Control for Mixture-of-Experts Expert Pruning

J. Mannings · A. Marzuki

Pruning 32 experts per layer from a 256-expert MoE costs 44% of base NLL on CPython stdlib and 16% on the same files with identifiers renamed: the reported damage figure moves 1.6× in absolute nats under a rewrite that changes no program structure. Flip rates at confidently predicted positions halve. An extractability probe confirms the corpora differ in memorisation as intended, but per-file extractability does not predict per-file damage, so memorisation is reported as the hypothesised mechanism rather than a demonstrated one.

preprintconfidence: likelyJ. ManningsA. Marzukicodebibtex

2026-07

Wave Convergence Thinking: Cross-Architecture Convergence as a Reasoning Signal

J. Mannings

Run one prompt across several different model architectures in parallel, decompose each trace into atomic claims, and score how claims from different architectures support or contradict each other, treating contradiction as destructive interference. Convergence between independently derived steps is the quality signal, and the node layer needs zero inter-node communication during generation, so it runs where sharding a single large model performs badly.

working noteconfidence: speculativeJ. Mannings

AUTHORS

Andryo Marzuki

Founder · applied LLM research

Builds and governs production AI systems by day and takes models apart by night. Works on the mixture-of-experts pruning line, where the renaming control and the extractability probe exist to test whether measured damage tracks memorisation.

Measured Pruning Damage Depends on the Evaluation Corpus: A Renaming Control for Mixture-of-Experts Expert Pruning

githubarxiv

Jeremiah Mannings

Founder · applied LLM research

Builds and governs production AI systems by day and takes models apart by night. Works on measurement: what compression actually removes, and how memorised data distorts the numbers we use to judge it.

githubarxiv

AGENDA

THE MAIN LOBE

Mechanistic questions about model compression and memorisation: what pruning, quantisation, and distillation actually remove, and how memorised data distorts every measurement we make of them.

Everything ships with code, controls, and a stated confidence level. Negative results are results.

THE NULLS

No frontier-scale pretraining. No product roadmap. No benchmark chasing, thought leadership, or papers whose headline number lacks a control.

An antenna gains directivity by choosing where not to radiate. So does a lab.

ABOUT

Main Lobe Labs is an independent research lab in Melbourne, Australia, run by practitioners who build and govern production AI systems by day and take models apart by night.

We think the interesting problems in applied LLM research are measurement problems, and that a small lab with a narrow beam can move faster on them than a large one with a wide one.