windowed-minority-guidance — Best Poster Award, EEML 2026

Tested whether minority guidance (Um, Lee & Ye, ICLR 2024) produces timestep-localized effects in diffusion denoising. Split the chain into 3 equal windows; mid-phase guidance recovered 45.6% of full-chain loss reduction across 250 iterations / 50 seeds on LSUN Bedroom. Wilcoxon signed-rank p < 0.001 for mid, early, and full guidance vs. baseline.

PrefectHQ/prefect · #21707 — count_flow_runs re-evaluated one correlated EXISTS subquery per filter per candidate row: O(k · N · log M) → O(N + ΣMᵢ) for both dialects.

Explicit JOINs (N:1 foreign keys only, task runs excluded) · SQLite: 51–66× at 100k rows · PostgreSQL: up to 18× at 100k rows · 21 regression tests

PrefectHQ/prefect · #21754 — read_block_schemas performed a linear next() scan of flat results for every nested reference lookup: O(N²) → O(N). for bulk reads.

Implemented a checksum_index dictionary with first-wins semantics to achieve O(1) lookups · ~125× speedup at 1,000 schemas (55.9ms → 0.39ms p50) · 12 new tests covering index threading and collision parity

NousResearch/hermes-agent · #62082 — the curator's LLM review fork built its AIAgent without enabled_toolsets, shipping the full default catalog plus lcm_* schema on a loop whose prompt names four tools.

Restricted the fork to enabled_toolsets=["skills", "terminal"] · behavior-neutral, mirrors the background-review fork · 21.5 KB → 6.6 KB advertised tool schema (69%, ~3.7K tokens) on every call, 50-100 calls per pass

SWE-agent/mini-swe-agent · #821 — FormatError discarded the raw model response on parse failure; trajectory logs showed the error marker but nothing to inspect

try/except FormatError in all 7 model query() wrappers · repr() fallback when model_dump() fails · 16 regression tests

PrefectHQ/prefect · #21004 — prefect worker start had no way to skip pool creation; operators hitting pre-existing pools got silent creation side-effects.

Fix introduced --no-create-pool-if-not-found flag, eliminating the side-effect entirely · orchestrated end-to-end via multi-agent workflow

NVIDIA/cutlass · #3242 — fast_math.h had PTX-accelerated specializations for half_t but none for bfloat16_t; every BF16 activation fell through to a 3-instruction float round-trip.

Added fast_exp and fast_tanh scalar + array specializations for bfloat16_t · tanh.approx.bf16x2 PTX on SM90+/CUDA 12+ · h2exp(__nv_bfloat162) on SM80+/CUDA 11+ · 6x instruction reduction per element (array path, Hopper/Blackwell) · 11 test cases

Lightricks/LTX-2 · #215 — tiled_encode_video invoked the VAE encoder sequentially per spatial tile: O(Tiles) → O(Shape Groups) via mini-batching.

Implemented two-phase collect/batch logic with shape-grouping for edge tiles · Hoisted loop-invariant linspace allocations · 3.75× to 15× reduction in VAE calls (1920p to 4K) · 12 parity and regression tests

sktime/sktime-mcp · #126 + #124 — 5 bugs in RegistryInterface (2 race conditions, 3 correctness errors) → open, awaiting review

Double-checked locking, 100% branch coverage, 7.8× throughput improvement (239ms → 31ms, cold-cache registry lookup)

fastapi/fastapi · #15476 — jsonable_encoder built set(obj.keys()) unconditionally on every dict branch, causing pure overhead when include and exclude filters were None

Skipped set allocation and membership tests when filters are absent · 8.2% wall-time reduction on small dicts (5.37 µs → 4.93 µs) · 6.0% reduction on large nested dicts · 7 regression tests added

MEGA Hackathon 2026 — Silver Medal (900+ participants)

Built Sestara in-contest: AI study roadmap platform (personalized paths, flashcards, quizzes, study assistant) addressing SDGs 4 · 10 · 11 · 16. Stack: React · TypeScript · Vite · Supabase · Google Gemini · shadcn/ui · Framer Motion · Wolfram Alpha LLM.