Malthe Have Musaeus

Malthe Have Musaeus

I work on language models and neural networks, mostly the mechanics of training and evaluation: data, post training, model compression, and the infrastructure behind the experiments.

At Paper Instruments, I worked on post-training for agentic tasks, long-horizon RL infrastructure, and evals. I’m now in San Francisco participating in The Bridge by Entrepreneurs First.

Selected work

01DiligenceBench: An Equity-Research Agent EvaluationMalthe Have Musaeus, Faisal Sayed, Mersad Abbasi, Daanish Khazi & Karina NguyenA benchmark for evaluating agents on long-form equity research. It tests whether agents can find the right filings, calculate from evidence, and make risk-aware judgments across 150 tasks and ten sectors.02Kos-1 Experimental: Env-Free RL on a 1T Parameter Agentic PriorQi Fang, Daanish Khazi, Gavin Bains, Joey Besgen, Malthe Have Musaeus & Rohit RastogiEarly results from post-training Kimi K2.5 on medical reinforcement learning data, exploring how much domain expertise can be added while preserving frontier tool-calling behavior.03Iterative Structured Knowledge Distillation: Optimizing Language Models Through Layer-by-Layer DistillationMalthe Have Musaeus & Rob van der GootA layer-by-layer distillation method that replaces large transformer blocks with smaller counterparts to make language models more computationally efficient.

Things I’m thinking about

  1. 01

    Can reasoning tokens be compressed into fewer learned tokens?

    I’m testing whether a set of learned embedding vectors can carry forward the useful information from a long reasoning trace. Similar to neural KV-cache compaction and Cartridges, but directly in the embedding space instead.

  2. 02

    Can agents rewrite their own history?

    I want to benchmark compaction methods that let agents rewrite their history in place with the same text-editing tools already available in their harness.

  3. 03

    Can models backtrack out of hallucinations?

    I was thinking whether we can detect hallucinations during autoregressive generation and return a model to its last non-hallucinated state.

  4. 04

    Do LLMs exhibit human personality traits?

    Good work exists on personality traits in LLMs, but what does that actually mean and is it fair to say that a model "has a personality" in the same sense a human has? And what is the implication of these traits on the model's behavior and decision-making? Is a model with highly measured extraversion better at selling? I think it would be fun tofine-tune a model towards a specific personality trait and measure its downstream performance.

More from the notebook