I am a Master's student at UPenn (Applied Math & Computational Science) and a graduate of Bowdoin College (Math & Economics), originally from Shanghai, China.

"The next gains won't come from bigger monoliths, but from populations of specialists — and from deciding who is allowed to keep knowing things when the world moves on."

I operationalize ecological dynamics (competitive exclusion, niche partitioning) to build learner populations that self-organize, retain dormant expertise that reward-chasing systems provably destroy, and stay calibrated to environments that return changed.

Emergent specialization and memory in learner populations under non-stationarity

One principle — what survives a bounded budget must be decided by structure, not by the reward/usage stream — instantiated across three substrates:

• GAUSE (weight space — learner populations) — competitive exclusion as a coordination-free assignment mechanism. Central finding: retention of dormant-regime knowledge tracks a single property — reward-independence of capacity assignment. Reward-driven allocators (capacity-bounded monoliths, learned MoE routers) are structurally blind to dormant tasks; emergent competition reaches the protective assignment with no gate, no schedule, no diversity objective — matching a hand-designed oracle it was never shown.

• RISP (decision space — trading strategy pools) — the financial successor: decision-focused (predict-then-optimize) strategy pools. Decomposes market non-stationarity into two independently-owned failure modes — between-regime forgetting (fixed by allocation) and within-regime drift (fixed by an episode-invariance objective) — with a measured super-additive interaction: neither fix works without the other. Ships its honest audits: a pre-registered real-data gate that returned null (shipped, not buried) and the measured boundary where the mechanism inverts.

• NicheMem (context space — long-horizon agent memory) — the same principle transported to agent memory: usage-driven policies (LRU paging, usage-weighted compression, learned utility) erase dormant-family skills as a class — what the score measures is irrelevant; that it is a function of the usage stream is decisive. Competitive memory ownership retains dormant runbooks exactly, recovering 97% of a privileged oracle with no labels, quotas, or trained router. Mechanism-level evidence today; the LLM tier is pre-registered (10 predictions, 4 refuted and reported).

• NicheMem's LLM tier — the decisive gate experiment its pre-registration commits to: capacity becomes context and memory, dormancy becomes stale expertise. The mechanism is validated; the substrate is the open question.
• RISP's equities flagship — CRSP S&P 500 constituents, 7 crisis episodes, structure gate pre-registered.

Key Results:

• 🚀 Efficiency: matches hand-built oracles without being shown them (GAUSE; RISP, p = 0.72; NicheMem, 97%) — all on laptop CPUs.
• 🧠 Emergence: specialization and ownership arise from local competition, never central design.
• 🔍 Honesty: every project pre-registers its predictions and ships its refutations — across the three repos, seven pre-registered predictions were refuted and reported as findings.

Key Notes:

• ❓ "Where does GAUSE most directly plug in?" A large MoE model whose rarely-triggered experts get overwritten because they emit no routing signal — GAUSE predicts the gate forgets them and competition-based ownership retains them.