Research lab

Language models,
looked at closely.

Independent research in cross-lingual NLP and LLM interpretability, with open-source ESP32 hardware built alongside it.

Read the papers The engineering side
3Papers
Aug 2026Latest
NLPFocus
3Hardware builds
MITLicense
Research

What the work is about

My research sits at the intersection of natural language processing, multilingual alignment and model interpretability. I want to know what small language models learn about language structure across typologically different languages, and whether that learning is genuinely cross-lingual or a surface-level statistical artefact. Lately that has extended to a more adversarial question: whether the internal signals we use to evaluate models, like linear probes for benchmark contamination, hold up under the same scrutiny.

arXiv Zenodo ResearchGate GitHub

Interests

  • LLM evaluation integrityDetecting benchmark contamination and other confounds that make reported model capabilities look better than they are.
  • Cross-lingual NLPHow models transfer knowledge across languages and what internal representations make this possible.
  • LLM interpretabilityProbing internal model representations to understand what language models actually learn, and what they don't.
  • Multilingual alignmentAligning semantic spaces across typologically diverse languages without losing cultural and pragmatic nuance.
  • On-device AIEfficient model compression, quantisation and deployment strategies for real-world edge constraints.
  • Behavioural signal processingUsing passive sensing and physiological data to model human wellbeing and detect behavioural patterns.

Papers

August 2026 Published Latest

Excess Separability: Nuisance-Controlled Residual-Stream Probing for Benchmark Contamination Detection

  • Nuisance-controlled protocol for contamination probing: zero-sum depth-profile contrast against a level-matched placebo, tested with a label-permutation null at 2× reference-to-suspect sizing
  • Shows naive linear-probe contamination detection fails: false-positive rate ranges 0.03–0.99 under a true null depending on control-set size; a flat-baseline contrast rejects a true null up to 72% of the time
  • 6-audit empirical study across transformer/dataset pairs — all 4 well-matched Pile arms return null, the protocol abstains on the confounded temporal split rather than reporting a verdict; implementation, tests, and audits released
Read the abstract

Benchmark contamination is diagnosed today with n-gram overlap, with likelihood-based membership inference, or with canary strings, and each needs something usually unavailable: the training corpus, a well-chosen test statistic, or foresight at dataset release. A recent alternative reads contamination off a linear probe on internal activations — this work shows the natural way to do this does not work, and specifies one that survives measurement. The protocol reports a zero-sum contrast on the depth profile of probe accuracy, recentred on a level-matched placebo baseline, tested against a label-permutation null, with the reference set twice the size of the suspect set. Reporting the level of excess separability rather than its shape makes the false-positive rate track the size of the analyst's own control set, from 0.03 to 0.99 under a true null. On real transformers, baseline depth profiles are measurably not flat, spanning up to 29.1 accuracy points on a temporal split, and their non-flatness tracks the surface difference between item sets (correlation 0.87 over 6 audits) — so the correction is largest exactly where it's needed.

NLPLLM EvaluationProbingBenchmark ContaminationStatistics
arXiv Zenodo Code
April 2026 Published

Interpretable Cross-Lingual Alignment in Small Language Models: Probing Cultural and Pragmatic Reasoning in Japanese-English Bilingual LLMs

  • J-PragEval-v0 — minimal-pair benchmark for four core Japanese pragmatic phenomena
  • Probing study of TinySwallow-1.5B: honorific register strongly encoded in residual stream; in-group/out-group and zero anaphora resolved dynamically at generation time
  • Pragmatic Representation Steering (PRS) — parameter-free inference-time steering via residual-stream activation editing
Read the abstract

Investigates how small Japanese language models represent and handle pragmatic meaning, focusing on areas where general-purpose LLMs often fail Japanese users. Addresses two major gaps: the lack of fine-grained evaluation methods for Japanese small language models beyond broad benchmarks such as JGLUE, and the limited understanding of culturally and pragmatically sensitive phenomena — honorifics, in-group vs. out-group reference, zero anaphora, and indirect refusal. Introduces J-PragEval-v0, a minimal-pair benchmark isolating four core pragmatic phenomena from surface fluency. Using linear probing and teacher-forced log-probability analysis, the study examines how these distinctions are encoded inside TinySwallow-1.5B. Proposes Pragmatic Representation Steering (PRS), a parameter-free inference-time method for steering model behaviour by editing residual-stream activations.

NLPJapanesePragmaticsProbingRepresentation SteeringLLMs
ResearchGate
October 2025 Published

The Foundational Role of Data Structures and Algorithms in Artificial Intelligence Systems

  • Systematic mapping of classical algorithms (A*, dynamic programming, graph search) to modern ML training and inference
  • Historical analysis tracing symbolic AI through to transformer architectures via algorithmic lineage
  • Concrete examples of how data structure choices affect memory consumption, processing speed, and scalability in real-world AI systems
Read the abstract

Artificial Intelligence has transformed from a theoretical discipline to a cornerstone of modern technology, powering applications from autonomous vehicles to personalised medicine. At the heart of every AI system lies a carefully orchestrated interplay between data structures and algorithms that determines efficiency, scalability, and capability. Through theoretical analysis and practical case studies, this thesis demonstrates that the selection of appropriate data structures and algorithmic approaches directly impacts AI performance across search, optimisation, machine learning, and knowledge representation. Traces the historical evolution from early symbolic AI relying on tree structures and graph search to modern deep learning architectures utilising tensors and backpropagation — and argues that continued innovation in algorithms remains essential for addressing current AI challenges.

AlgorithmsData StructuresAI SystemsML InfrastructureThesis
ResearchGate

Toolkit

PythonPyTorchHugging Face Transformersscikit-learnLangChainClaude APILinear AlgebraLinear ProbingActivation SteeringPermutation TestingLaTeX
About

One lab, two ways of understanding things.

Mabushi Lab is where Florian Braun publishes independent research and open-source engineering. The research side studies how language models represent meaning across languages and whether the tools used to evaluate them can be trusted.

The engineering side builds hardware from scratch: RF sniffers that decode 433 MHz bursts in real time, paired motion sensors that stream over WebSocket, and a small wall display. Both sides share the same habit of measuring carefully and releasing the code.

Python / PyTorchResearch code
Hugging Face TransformersModels and probing
C++ / Arduino FrameworkFirmware language
PlatformIOBuild system
ESP32 / ESP32-S3Microcontroller
WebSocket / BLE / WiFiConnectivity