Researcher
直接向此求職者傳送職缺邀請
Research Interests: World models; open-ended learning; active scientific discovery; sequential Bayesian experimental design; model-based reinforcement learning.
Researcher - OpenQFT-Bench: Active World-Model Induction in Hidden Particle Universes at Independent research (2026-07 – Present)
Designed an interactive benchmark in which agents choose collision experiments, infer latent particles and typed interaction mechanisms, maintain uncertainty, and revise hypotheses in procedurally generated hidden universes. Built a JAX stack spanning a typed symbolic DSL, oracle-verified simulator, functional experiment environment, predictive and structural evaluation, and an SMC learner with expected-information-gain experiment selection; the current CPU suite has 1,899 passing tests and 5 skips. Used preregistered audits to separate proposal reachability from evidence-driven recovery.
Later matched-budget controls recovered exact topology across 90/90 cells in a frozen compound family; held-out generalisation and the proposed open-ended co-evolution track remain open questions.
Co-first author - CLIMB: Open-Ended Continual Learning with DreamerV3 at UCL group research (2026-02 – 2026-04)
Co-developed three interventions for Craftax-symbolic: soft feasibility masking in real and imagined rollouts, episodic spatial/crafting novelty rewards, and replay partitioned by novelty, learnability, and recency. Evaluated 19 configurations over three seeds. At 10M steps, the full system reached mean achievement rate 0.200 ± 0.008 versus 0.099 ± 0.009 for DreamerV3's 50:50 replay baseline, and 13.4 versus 6.6 achievements per episode.
Isolated a super-additive interaction: masking alone improved performance, whereas novelty and structured replay helped only when masking opened deeper technology-tree tiers. Also documented failures of hard masking and exploration-only Plan2Explore in this symbolic setting.
Researcher - RoPE-Aligned Rotations for Dynamic 4-Bit LLM Quantisation at Independent research (2026-05 – Present)
Proved that orthogonal Q/K rotations commuting with RoPE form a band-diagonal SO(2)dh/2 centraliser and derived a training-free closed-form angle that is locally optimal for a paired-channel variance objective. Conducted multi-model, multi-seed, short- and long-context experiments under dynamic INT4 quantisation. The locally optimal rotation did not beat generic Hadamard mixing: layering was equivalent within the registered margin and replacement degraded perplexity.
Traced the null result to quantiser adaptivity through static-versus-dynamic controls, corrected a RoPE-scaling error, and withdrew an earlier apparent gain after identifying it as a seed-specific artefact.
MSc dissertation - Temporal Value Estimation for Software-Engineering Agents at UCL (2026-04 – 2026-08)
Supervisor: Dr Ilija Bogunovic. Studied whether partial execution trajectories reveal when a low-cost coding agent should continue or hand control to a stronger model; developed a compact 149M-parameter trajectory encoder, approximately 1/47 the size of the 7B reference value model. Reframed routing around the counterfactual net benefit of switching and ran preregistered multi-seed, bootstrap, fixed-budget, and cross-domain audits on SWE-bench Verified and SWE-Smith.
Found that several apparent gains were sensitive to seed, model pair, normalisation, or domain shift; documented the resulting limits on model selection and on claims of a universally deployable routing policy.
Fourth author - SWE-Router: Cost-Aware Routing for Multi-turn Coding Agents at Collaborative research (2026-04 – 2026-07)
Built the repository-disjoint data and evaluation pipeline: mini-SWE-agent trajectory collection, official SWE-bench/SWE-Smith grading in Docker containers, per-step cost annotation, and adapters to the paper's routing harness. The paper lets a lower-cost model explore before a value function decides whether to continue or switch; on SWE-bench Verified, its Route-AUC exceeded the strongest prompting baseline by 12 percentage points.
MSc in Machine Learning – University College London (UCL) (2025-09 – 2026-12)
BSc (Hons) in Mathematics and Physics – University of Manchester (2022-09 – 2025-06)
