RSI-Exam

name stringlengths 29 65	instruction stringlengths 2.05k 9.19k	description stringlengths 0 1.16k	keywords listlengths 0 0	schema_version stringclasses 3 values	config unknown	metadata unknown	files listlengths 11 1.07k	location stringlengths 15 52
autoresearch-arena/bbo_noisy_continuous	# Noisy Continuous Black-Box Optimizer Design Your task is to design a reproducible, noise-aware optimizer that minimizes 10-dimensional continuous black-box objectives with smooth periodic, multimodal structure. You inherit a weak uniform-random baseline, and every run uses the box `[-5, 5]^10` with a strict budget o...	Design a noise-aware optimizer for multimodal continuous functions under a fixed query budget and a comparable visible/sealed normalized anytime-final metric.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/bbo_noisy_continuous", "description": "Design a noise-aware optimizer for multimodal continuous functions under a fixed query budget and a comparable visible/sealed normalized anytime-final metric.", ...	{}	[ "environment/Dockerfile", "environment/bbo_harness.py", "environment/data/visible.json", "environment/methods/main/solver.py", "environment/selfcheck.py", "instruction.md", "task.toml", "tests/Dockerfile", "tests/child_optimizer.py", "tests/grade.py", "tests/heldout/bbo_harness.py", "tests/hel...	bbo_noisy_continuous
autoresearch-arena/bbo_simopt_inventory	# Query-efficient stochastic inventory policy search Design a reproducible optimizer for four normalized controls of a replenishment policy under stochastic demand and lead times. Minimize simulated operating cost within 48 queries over `[0, 1]^4`. You inherit a uniform-random starter and receive noisy objective feedb...	Design a query-efficient optimizer for a stochastic replenishment policy across varied cost and demand regimes.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/bbo_simopt_inventory", "description": "Design a query-efficient optimizer for a stochastic replenishment policy across varied cost and demand regimes.", "authors": [ { "name": "haoting...	{}	[ "environment/Dockerfile", "environment/bbo_harness.py", "environment/data/visible.json", "environment/data/visible_anchors.json", "environment/methods/main/solver.py", "environment/public_scoring.py", "environment/selfcheck.py", "instruction.md", "solution/reference_optimizer.py", "solution/solve....	bbo_simopt_inventory
autoresearch-arena/bigann_filtered_vector_search	# Filtered nearest-neighbour search at 10M scale A photo library holds 10,000,000 images. Each is a 192-dimensional `uint8` embedding plus a set of metadata tags (uploader, camera, place, year, and so on). A query arrives as an embedding plus a tag predicate: one or two tags. An image is a valid answer only if...	Build a filtered-ANN index/search solver over the real Big-ANN yfcc-10M library (10M x 192 uint8 vectors, 200,386 tags, 108M tag entries); graded by wall-clock throughput of one batch call at recall@10>=0.90 on a sealed hidden workload.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/bigann_filtered_vector_search", "description": "Build a filtered-ANN index/search solver over the real Big-ANN yfcc-10M library (10M x 192 uint8 vectors, 200,386 tags, 108M tag entries); graded by wall-...	{}	[ "environment/Dockerfile", "environment/methods/main/ann_utils.py", "environment/methods/main/solver.py", "environment/prep_data.py", "environment/selfcheck.py", "environment/visible_ids.npy", "instruction.md", "solution/methods/reference/ann_utils.py", "solution/methods/reference/solver.py", "solu...	bigann_filtered_vector_search
autoresearch-arena/citywide_signal_coordination	# City-Wide Signal Coordination You inherit a full-city traffic microsimulation — a real metropolitan road network with ~600 signalised junctions and a synthetic morning-peak demand — running on its original, un-optimised static signal timing. Improve the signal control so the whole city moves better under load, minim...	Coordinate signal timing across a full metropolitan road network (~600 signalised junctions) to minimise a weighted delay/teleport/stop cost under heavy demand; you submit a signal program / online controller, re-run on sealed hidden load and seed settings, lower is better.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/citywide_signal_coordination", "description": "Coordinate signal timing across a full metropolitan road network (~600 signalised junctions) to minimise a weighted delay/teleport/stop cost under heavy de...	{}	[ "environment/Dockerfile", "environment/methods/main/controller.py", "environment/methods/main/tls_info.json", "environment/prepare_data.py", "environment/selfcheck.py", "environment/sim/eval_core.py", "environment/sim/sim_rpc.py", "instruction.md", "task.toml", "tests/Dockerfile", "tests/anchors...	citywide_signal_coordination
autoresearch-arena/clevr_cogent_grpo_qwen2vl	# Train Qwen2-VL-2B with GRPO so it generalizes off the training distribution You inherit a deliberately weak reinforcement-learning (GRPO) training project for the vision-language model `Qwen2-VL-2B-Instruct`: the starting point in `/app/methods/main/` trains on the provided counting-VQA split with a trivial reward a...	Train Qwen2-VL-2B-Instruct with GRPO on CLEVR-CoGenT-A so it generalizes to hidden CLEVR-CoGenT-B and SuperCLEVR; graded by mean keyword-matching counting accuracy.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/submission", "/app/methods" ], "task": { "name": "autoresearch-arena/clevr_cogent_grpo_qwen2vl", "description": "Train Qwen2-VL-2B-Instruct with GRPO on CLEVR-CoGenT-A so it generalizes to hidden CLEVR-CoGenT-B and SuperCLEVR; graded by mean keywor...	{}	[ "environment/Dockerfile", "environment/docker-compose.yaml", "environment/eval_common.py", "environment/methods/main/README.md", "environment/methods/main/train.sh", "environment/prepare_data.py", "environment/selfcheck.py", "instruction.md", "task.toml", "tests/Dockerfile", "tests/anchors.json"...	clevr_cogent_grpo_qwen2vl
autoresearch-arena/cxr_ood_triage_policy	# Chest X-ray triage under cross-site distribution shift You inherit a radiology triage component from a teammate. A fixed, frozen chest-radiograph classifier already emits per-finding probabilities for every study; the teammate's stop-gap refers a study to a radiologist whenever the largest critical-finding probabili...	Improve a weak chest-X-ray triage policy over frozen classifier scores. You submit triage(cases, resources, seed); a sealed verifier refers the top-risk studies under a fixed shared budget across an in-domain and an unseen site, and scores mean sensitivity minus a Brier (calibration) penalty minus a worst-site gap. Raw...	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/cxr_ood_triage_policy", "description": "Improve a weak chest-X-ray triage policy over frozen classifier scores. You submit triage(cases, resources, seed); a sealed verifier refers the top-risk studies u...	{}	[ "environment/.dockerignore", "environment/Dockerfile", "environment/data/calibration.jsonl", "environment/data/grade.jsonl", "environment/data/score_schema.yaml", "environment/methods/main/solver.py", "environment/selfcheck.py", "instruction.md", "solution/methods/README.md", "solution/methods/ref...	cxr_ood_triage_policy
autoresearch-arena/device_iv_regime_extrapolation_v1	# Commit terminal-current numbers for a diode lot beyond the safe test window You are an AI agent assisting the device engineer who must qualify a new power-diode lot at a semiconductor supplier. The probe station covers only a safe bias/temperature window — 290–325 K, reverse bias to −3.6 V — but the qualification si...	Power-device qualification sign-off: submit a prediction method that turns safe-window I-V/temperature measurements of a diode lot into terminal-current commitments at mission-profile extremes (high temperature, deep reverse bias); a sealed evaluation re-runs it on an unseen lot drawn from the same population as the pr...	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/device_iv_regime_extrapolation_v1", "description": "Power-device qualification sign-off: submit a prediction method that turns safe-window I-V/temperature measurements of a diode lot into terminal-curre...	{}	[ "environment/.dockerignore", "environment/Dockerfile", "environment/data/practice/DATA_CARD.md", "environment/data/practice/instances.json", "environment/data/qualification/records.json", "environment/examples/devsim_diode_demo.py", "environment/methods/main/solver.py", "environment/run_solver.py", ...	device_iv_regime_extrapolation
autoresearch-arena/discoveryworld_agent_harness_low2	# Improve a Scientific Discovery Agent You inherit a weak but working ReAct harness that drives a fixed language model through open-ended scientific discovery scenarios in a simulated world. Improve its reasoning, experimentation, memory, action selection, and the explanatory conclusions it states; the submitted harne...	Improve a ReAct harness over a fixed language model on low-baseline Space Sick and Chemistry Challenge tasks.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/discoveryworld_agent_harness_low2", "description": "Improve a ReAct harness over a fixed language model on low-baseline Space Sick and Chemistry Challenge tasks.", "authors": [ { "name...	{}	[ "environment/Dockerfile", "environment/knowledge_eval.py", "environment/methods/main/agent.py", "environment/runner.py", "environment/selfcheck.py", "instruction.md", "solution/methods/sota/agent.py", "solution/solve.sh", "task.toml", "tests/Dockerfile", "tests/anchors.json", "tests/grade.py",...	discoveryworld_agent_harness_low2
autoresearch-arena/eda_gate_sizing	# Optimize logic gate sizing under timing and design-rule penalties ## Background A chip is built from millions of tiny logic gates, and the same gate comes in several sizes that do the identical job. A bigger one switches faster, but burns more power and is heavier for the gate feeding it to push. Picking a size for...	Choose a library cell per instance to minimize power under timing/DRV constraints on placed netlists (CircuitOps IR tables); submit an algorithm, graded on sealed hidden designs, lower is better.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/eda_gate_sizing", "description": "Choose a library cell per instance to minimize power under timing/DRV constraints on placed netlists (CircuitOps IR tables); submit an algorithm, graded on sealed hidde...	{}	[ "environment/Dockerfile", "environment/eval/OpenROAD_helper.py", "environment/eval/OpenROAD_helper.tcl", "environment/eval/check_validity.py", "environment/eval/evaluate_case.py", "environment/eval/evaluation.py", "environment/methods/main/solver.py", "environment/prepare_data.py", "environment/self...	eda_gate_sizing
autoresearch-arena/feeder_phase_impedance_inversion_v1	# Repair a feeder's phase and impedance records from smart-meter archives You are an AI agent assisting a distribution planning engineer at a mid-size utility with a solar interconnection queue. Before any hosting-capacity or voltage-drop study can be signed off, the feeder model has to be calibrated: the GIS says whi...	Distribution-feeder model calibration before hosting-capacity studies: submit a method that jointly repairs per-customer phase labels, per-segment conductor impedances and the regulator tap from magnitude-only AMI archives plus head SCADA; a sealed evaluation re-runs it on a larger fleet of unseen feeders drawn from th...	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/feeder_phase_impedance_inversion_v1", "description": "Distribution-feeder model calibration before hosting-capacity studies: submit a method that jointly repairs per-customer phase labels, per-segment c...	{}	[ "environment/.dockerignore", "environment/Dockerfile", "environment/data/assessment/records.json", "environment/data/practice/DATA_CARD.md", "environment/data/practice/instances.json", "environment/methods/main/dss_forward.py", "environment/methods/main/solver.py", "environment/run_solver.py", "envi...	feeder_phase_impedance_inversion
autoresearch-arena/finscope_dcf_valuation	# FinScope coverage desk — forward value-driver model You have inherited a buy-side equity-research desk's intrinsic-value model. The desk values companies with a fixed, trusted discounted-cash-flow (DCF) engine that turns a set of forward value-drivers into a per-share value; the open problem is **forecasting...	Forecast forward value-drivers for a coverage universe from a messy fundamentals panel; graded by median absolute DCF valuation error on a sealed held-out tier.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/finscope_dcf_valuation", "description": "Forecast forward value-drivers for a coverage universe from a messy fundamentals panel; graded by median absolute DCF valuation error on a sealed held-out tier."...	{}	[ "environment/.dockerignore", "environment/Dockerfile", "environment/README.md", "environment/TODO.md", "environment/data/calibration_truth.csv", "environment/data/coverage.csv", "environment/data/data_dictionary.md", "environment/data/fundamentals.csv", "environment/fit_calibration.py", "environme...	finscope_dcf_valuation
autoresearch-arena/flashattention_varlen_feature_full_vjp_speedup	# Noncausal Global Extreme-GQA Ragged Attention Full VJP Optimize an inherited eager implementation of packed ragged GQA cross-attention with extreme query-head sharing, bidirectionally unequal Q/K lengths, per-sequence ALiBi, softcapping, non-dense packed views, and complete Q/K/V vector-Jacobian products. Maximize r...	Optimize six preregistered D160 noncausal-global ragged attention strata per panel with 48–96:1 GQA, ALiBi, softcap, non-dense packed views, and complete Q/K/V VJP.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/flashattention_varlen_feature_full_vjp_speedup", "description": "Optimize six preregistered D160 noncausal-global ragged attention strata per panel with 48–96:1 GQA, ALiBi, softcap, non-dense packed vie...	{}	[ "environment/.dockerignore", "environment/Dockerfile", "environment/docker-compose.yaml", "environment/methods/experiment_log.md", "environment/methods/main/solver.py", "environment/methods/versions/v0/solver.py", "environment/problems/visible_baseline_manifest.json", "environment/problems/visible_cas...	flashattention_varlen_feature_full_vjp_speedup
autoresearch-arena/flashfftconv_multistream_gated_forward_speedup	# Multi-stream Gated FlashFFTConv Forward Optimize multiple independent gated causal convolutions across the published long-convolution range. The final implementation is re-run on sealed stream counts, shapes, gate distributions, kernel scales, lengths, and seeds. ## Hard Constraints - Edit only `/app/methods/main/...	Optimize independent multi-stream gated causal convolutions in the published FP16 D=768, length-2048 regime across multiple batch/stream/head factorizations on NVIDIA H100.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/flashfftconv_multistream_gated_forward_speedup", "description": "Optimize independent multi-stream gated causal convolutions in the published FP16 D=768, length-2048 regime across multiple batch/stream/...	{}	[ "environment/Dockerfile", "environment/docker-compose.yaml", "environment/methods/experiment_log.md", "environment/methods/main/solver.py", "environment/problems/visible_baseline_manifest.json", "environment/problems/visible_cases.json", "environment/protocol.py", "environment/selfcheck.py", "enviro...	flashfftconv_multistream_gated_forward_speedup
autoresearch-arena/game2048_policy_search	# 2048 Strategy You inherit a weak Python policy that plays deterministic seeded 2048 games. Improve the policy using repeated experiments on the public seed suite; your submitted policy is replayed from scratch on sealed seeds from the same generator for scoring. ## Hard Constraints - Edit `methods/main/policy.py`;...	Improve a Python policy for seeded 2048 games, using public games for feedback and sealed same-distribution seeds for final scoring.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/game2048_policy_search", "description": "Improve a Python policy for seeded 2048 games, using public games for feedback and sealed same-distribution seeds for final scoring.", "authors": [ { ...	{}	[ "environment/.dockerignore", "environment/Dockerfile", "environment/docker-compose.yaml", "environment/evaluate.py", "environment/game2048.py", "environment/methods/main/policy.py", "environment/selfcheck.py", "environment/visible_seeds.json", "instruction.md", "task.toml", "tests/.dockerignore"...	game2048_policy_search
autoresearch-arena/johnson1991_leighton_graph_coloring	# Drive coloring conflicts to zero at a fixed color budget You inherit a TabuCol graph-coloring baseline. For every graph you are given a fixed color budget `target_k`. Assign each vertex a color in `{0, 1, ..., target_k-1}` so as to minimize the number of monochromatic edges (edges whose two endpoints share a...	Minimize coloring conflicts at a fixed color budget on real, hard DIMACS/COLOR02 graphs (Leighton le450, Johnson DSJC, DSJR; n=450..1000). target_k is each graph's PUBLISHED best-known chromatic number, so a proper 0-conflict coloring provably exists but sits at the coloring frontier. The verifier reruns solve(data) on...	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/johnson1991_leighton_graph_coloring", "description": "Minimize coloring conflicts at a fixed color budget on real, hard DIMACS/COLOR02 graphs (Leighton le450, Johnson DSJC, DSJR; n=450..1000). target_k ...	{}	[ "environment/Dockerfile", "environment/data/visible.json", "environment/methods/main/gc_lib.py", "environment/methods/main/solver.py", "environment/selfcheck.py", "instruction.md", "task.toml", "tests/Dockerfile", "tests/child_solve.py", "tests/grade.py", "tests/heldout/hidden_anchors.json", "...	johnson1991_leighton_graph_coloring
autoresearch-arena/lean_formal_proof_workflow_design	# Design a verifier-guided Lean 4 proof workflow You inherit a small compiler-guided proof workflow over a pinned language model. Improve the workflow in `/app/methods/main/` so it completes canonical Lean 4 theorem statements reliably under a tight inference and compiler-feedback budget. The verifier copies your fina...	Design one verifier-guided workflow that proves fixed formal theorem statements with a pinned language model under per-problem inference, compiler-check, and wall-clock budgets.	[]	1.3	{ "schema_version": "1.3", "artifacts": [ "/app/methods" ], "task": { "name": "autoresearch-arena/lean_formal_proof_workflow_design", "description": "Design one verifier-guided workflow that proves fixed formal theorem statements with a pinned language model under per-problem inference, compiler-che...	{}	[ "environment/.dockerignore", "environment/Dockerfile", "environment/child_runner.py", "environment/data/visible.jsonl", "environment/evaluator.py", "environment/lean_compiler.py", "environment/methods/main/prompt_utils.py", "environment/methods/main/solver.py", "environment/model_proxy.py", "envir...	lean_formal_proof_workflow_design

End of preview. Expand in Data Studio

Benchmarking Recursive Self-Improvement through Executable Research

Can frontier AI agents achieve Recursive Self-Improvement by turning a weak method into one that performs better on hidden data?

Overview

RSI-Exam is a benchmark of 88 executable research tasks for evaluating whether frontier AI agents can achieve recursive self-improvement. Instead of measuring one-shot problem solving, each task hands the agent a working but weak method and lets it run its own experiments for up to 12 hours against a visible split.

Whatever it submits is then replayed once, from scratch, on a sealed split it never touched, and that single number is the score. An improvement that does not transfer earns nothing. We publicly release 35 tasks with their full grading containers.

Task Taxonomy

The full benchmark holds 88 active tasks across six domains.

88 active tasks across six domains

This repository holds the 35 public tasks. The other 53 are held back.

Domain	What it covers	Public	Private	Total
AI Models & Agents	Agent harnesses, LLM memory, post-training and sampler recipes, offline RL	8	10	18
Systems & Hardware	GPU and TPU kernels, EDA placement, vector search, QEC decoders	9	4	13
Physical Sciences & Engineering	Inference and extrapolation from measurement data, control policies, physical design	4	15	19
Optimization, Planning & Control	Routing, scheduling, cutting and black-box search under a hard budget	6	12	18
Life Sciences & Medicine	Single-cell genomics, molecular structure, clinical policy under shift	5	7	12
Finance, Law & Business	Valuation, alpha factors, long-horizon professional casework	3	5	8
Total		35	53	88

Task Structure & Evaluation

<task>/
  task.toml         metric, timeouts, network policy, artifact declarations
  instruction.md    the brief the agent is given
  environment/      the agent's container: starting method, visible data, tools
  tests/            the grading container: grade.py, anchors, the sealed split
  solution/         frontier-calibrated reference (where one exists)

environment/ and tests/ build into two separate images, and that split is the sealed boundary. The agent edits, measures and repeats inside its own container; the only thing that ever crosses is the artifact it submits. A fresh container then reruns that artifact against hidden data with a fixed verifier, and the number it returns is the score.

The evaluation pipeline: the agent iterates on the visible set, one artifact crosses the sealed boundary, a fresh container reruns it on hidden data with a fixed verifier

Leaderboard

Mean hidden-set score. Every task is normalised the same way: a shipped baseline scores 0, 0.6 is a frontier-calibrated reference, and a proven upper bound scores 1. No agent reaches the reference.

Mean hidden-set score over all 88 tasks, from Opus 5 at 0.464 down to Gemini 3.7 Flash at 0.309