Tasks

Design an LLM agent memory architecture that generalizes

Design a distillation recipe for a 1.7B student using a fixed 8B teacher

Redesign a ReAct harness for long-horizon scientific discovery in unseen simulated worlds

Design a GRPO recipe for visual counting across unseen attribute combinations and renderers

Design a compiler-guided LLM workflow for unseen theorems under fixed inference and checking budgets

Select candidate trajectories and predict visibility under a four-switch budget on unseen videos

Train a control policy from fixed logged trajectories without simulator interaction

Fuse noisy multi-view observations into one point cloud

Optimize single-core top-10 vector retrieval under one- and two-tag predicates

Minimize leakage while satisfying timing, slew, and capacitance constraints on unseen netlists

Optimize multi-bit flip-flop grouping and legal placement across power, area, timing, and displacement costs

Reimplement a colour-code decoder without the usual libraries

Choose contraction pairs to minimize peak diagram size and runtime in exact equivalence checking

Accelerate packed variable-length attention with complete Q/K/V backward passes on unseen shapes

Accelerate gated causal long convolutions across independent streams and unseen shapes

Accelerate grouped-query attention over paged, variable-length KV caches and unseen shapes

Accelerate decode, prefill, and encoder attention across masks, head-sharing ratios, and unseen shapes

Jointly plan pickup routes and reusable macro recordings to minimize controller button presses

Predict leakage and conduction currents at destructive voltage-temperature conditions from safe-window measurements

Jointly infer customer phases, line impedances, and regulator taps on unseen feeders

Recover a spatial drag field from sparse, noisy tide-gauge amplitudes and phases on unseen basins

Maximize mean game score on unseen seeds under a pooled CPU budget

Minimize 10-dimensional multimodal objectives with 120 noisy evaluations on unseen instances

Minimize simulated inventory cost by tuning four replenishment controls within 48 noisy queries

Coordinate phases, green times, and offsets across ~600 junctions under unseen traffic demand

Minimize monochromatic edges on unseen hard graphs under a 45-second solve limit

Pack glass orders into minimum stock plates under staged cuts, defects, and stack-sequence constraints

Calibrate and rank frozen model scores to prioritize critical studies under a shared review budget

Remove batch effects while preserving cell-type structure in embeddings recomputed on unseen cells

Generate treated-cell image distributions from controls and condition embeddings without paired examples

Label donor-disjoint cells using five examples per class, an unlabeled pool, and a fixed reference model

Improve sampling, pose selection, and post-processing under physical-validity and 2 Å RMSD checks

Forecast growth, margins, reinvestment, WACC, and terminal growth for unseen companies from messy fundamentals

Answer regulatory and transactional questions across six matters and 1,000+ case documents

Predict daily forward-return rankings from a 360-factor temporal panel on a sealed later period