Core16 · qualitative archive

64 final-Policy reruns

Sixteen Environments, four coding-agent/model lanes, and one original-environment rerun for every validation-selected checkpoint.

Acrobot

Swing-up control with a two-link underactuated arm.

ControlBest reported -84.69Result detail →

native env render

GPT-5.5

Codex

BEST

Held-out: -84.69
Checkpoint: #001
Rerun: 76 steps

Validation-selected checkpoint rerun in the original environment; case return -75.00.

native env render

Claude Opus 4.7

Claude Code

Held-out: -88.59
Checkpoint: #011
Rerun: 74 steps

Validation-selected checkpoint rerun in the original environment; case return -73.00.

native env render

MiniMax-M3

Claude Code

Held-out: -136.4
Checkpoint: #010
Rerun: 249 steps

Validation-selected checkpoint rerun in the original environment; case return -248.0.

native env render

DeepSeek-V4-Pro

Claude Code

Held-out: -91.16
Checkpoint: #007
Rerun: 130 steps

Validation-selected checkpoint rerun in the original environment; case return -129.0.

ContinuousCar

Continuous throttle control with delayed goal reward.

ControlBest reported 98.77Result detail →

native env render

GPT-5.5

Codex

Held-out: 95.22
Checkpoint: #005
Rerun: 88 steps

Validation-selected checkpoint rerun in the original environment; case return 95.69.

native env render

Claude Opus 4.7

Claude Code

BEST

Held-out: 98.77
Checkpoint: #010
Rerun: 344 steps

Validation-selected checkpoint rerun in the original environment; case return 98.89.

native env render

MiniMax-M3

Claude Code

Held-out: 91.88
Checkpoint: #010
Rerun: 143 steps

Validation-selected checkpoint rerun in the original environment; case return 91.61.

native env render

DeepSeek-V4-Pro

Claude Code

Held-out: 94.48
Checkpoint: #012
Rerun: 69 steps

Validation-selected checkpoint rerun in the original environment; case return 94.50.

Bipedal

Contact-rich locomotion across uneven terrain.

Box2DBest reported 248.9Result detail →

native env render

GPT-5.5

Codex

BEST

Held-out: 248.9
Checkpoint: #010
Rerun: 1600 steps

Validation-selected checkpoint rerun in the original environment; case return 273.3.

native env render

Claude Opus 4.7

Claude Code

Held-out: -15.84
Checkpoint: #027
Rerun: 1600 steps

Validation-selected checkpoint rerun in the original environment; case return -15.30.

native env render

MiniMax-M3

Claude Code

Held-out: -80.87
Checkpoint: #001
Rerun: 217 steps

Validation-selected checkpoint rerun in the original environment; case return -68.83.

native env render

DeepSeek-V4-Pro

Claude Code

Held-out: -97.48
Checkpoint: #004
Rerun: 113 steps

Validation-selected checkpoint rerun in the original environment; case return -97.29.

CarRacing

Pixel-observation driving on procedurally generated tracks.

Box2DBest reported 604.1Result detail →

native env render

GPT-5.5

Codex

BEST

Held-out: 604.1
Checkpoint: #009
Rerun: 1000 steps

Validation-selected checkpoint rerun in the original environment; case return 499.4.

native env render

Claude Opus 4.7

Claude Code

Held-out: 602.3
Checkpoint: #006
Rerun: 1000 steps

Validation-selected checkpoint rerun in the original environment; case return 499.4.

native env render

MiniMax-M3

Claude Code

Held-out: 233.4
Checkpoint: #008
Rerun: 1000 steps

Validation-selected checkpoint rerun in the original environment; case return 171.9.

native env render

DeepSeek-V4-Pro

Claude Code

Held-out: 25.20
Checkpoint: #002
Rerun: 1000 steps

Validation-selected checkpoint rerun in the original environment; case return -3.509.

Reacher

Continuous robotic-arm control toward a target point.

MuJoCoBest reported -3.473Result detail →

direct mujoco renderer

GPT-5.5

Codex

BEST

Held-out: -3.473
Checkpoint: #012
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -1.880.

direct mujoco renderer

Claude Opus 4.7

Claude Code

Held-out: -3.979
Checkpoint: #007
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -2.670.

direct mujoco renderer

MiniMax-M3

Claude Code

Held-out: -5.103
Checkpoint: #015
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -2.308.

direct mujoco renderer

DeepSeek-V4-Pro

Claude Code

Held-out: -6.506
Checkpoint: #019
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -4.613.

HalfCheetah

High-dimensional planar locomotion with dense feedback.

MuJoCoBest reported 606.2Result detail →

direct mujoco renderer

GPT-5.5

Codex

Held-out: 601.8
Checkpoint: #003
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 211.2.

direct mujoco renderer

Claude Opus 4.7

Claude Code

Held-out: 452.0
Checkpoint: #038
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 119.7.

direct mujoco renderer

MiniMax-M3

Claude Code

BEST

Held-out: 606.2
Checkpoint: #040
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 128.0.

direct mujoco renderer

DeepSeek-V4-Pro

Claude Code

Held-out: -0.468
Checkpoint: #017
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return -0.893.

Ant

Multi-legged locomotion with eight continuous actions.

MuJoCoBest reported 990.1Result detail →

direct mujoco renderer

GPT-5.5

Codex

Held-out: 989.6
Checkpoint: #006
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 221.1.

direct mujoco renderer

Claude Opus 4.7

Claude Code

BEST

Held-out: 990.1
Checkpoint: #002
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 204.6.

direct mujoco renderer

MiniMax-M3

Claude Code

Held-out: 983.4
Checkpoint: #007
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 215.7.

direct mujoco renderer

DeepSeek-V4-Pro

Claude Code

Held-out: 898.0
Checkpoint: #019
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 192.5.

Pusher

Contact manipulation that pushes an object to a goal.

MuJoCoBest reported -37.11Result detail →

direct mujoco renderer

GPT-5.5

Codex

BEST

Held-out: -37.11
Checkpoint: #005
Rerun: 100 steps

Validation-selected checkpoint rerun in the original environment; case return -38.36.

direct mujoco renderer

Claude Opus 4.7

Claude Code

Held-out: -38.52
Checkpoint: #014
Rerun: 100 steps

Validation-selected checkpoint rerun in the original environment; case return -37.53.

direct mujoco renderer

MiniMax-M3

Claude Code

Held-out: -39.25
Checkpoint: #010
Rerun: 100 steps

Validation-selected checkpoint rerun in the original environment; case return -41.99.

direct mujoco renderer

DeepSeek-V4-Pro

Claude Code

Held-out: -54.18
Checkpoint: #004
Rerun: 100 steps

Validation-selected checkpoint rerun in the original environment; case return -55.45.

DoorKey

Find a key, unlock a door, and reach the goal.

MiniGridBest reported 0.986Result detail →

native env render

GPT-5.5

Codex

BEST

Held-out: 0.986
Checkpoint: #007
Rerun: 43 steps

Validation-selected checkpoint rerun in the original environment; case return 0.985.

native env render

Claude Opus 4.7

Claude Code

Held-out: 0.983
Checkpoint: #005
Rerun: 66 steps

Validation-selected checkpoint rerun in the original environment; case return 0.977.

native env render

MiniMax-M3

Claude Code

Held-out: 0.000
Checkpoint: #015
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 0.000.

native env render

DeepSeek-V4-Pro

Claude Code

Held-out: 0.000
Checkpoint: #025
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 0.000.

KeyCorridor

Multi-room search with object interaction and memory.

MiniGridBest reported 0.930Result detail →

native env render

GPT-5.5

Codex

Held-out: 0.918
Checkpoint: #005
Rerun: 25 steps

Validation-selected checkpoint rerun in the original environment; case return 0.953.

native env render

Claude Opus 4.7

Claude Code

BEST

Held-out: 0.930
Checkpoint: #003
Rerun: 26 steps

Validation-selected checkpoint rerun in the original environment; case return 0.951.

native env render

MiniMax-M3

Claude Code

Held-out: 0.000
Checkpoint: #012
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 0.000.

native env render

DeepSeek-V4-Pro

Claude Code

Held-out: 0.000
Checkpoint: #023
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 0.000.

FourRooms

Partial-observation navigation through bottlenecks.

MiniGridBest reported 0.701Result detail →

native env render

GPT-5.5

Codex

Held-out: 0.664
Checkpoint: #013
Rerun: 38 steps

Validation-selected checkpoint rerun in the original environment; case return 0.658.

native env render

Claude Opus 4.7

Claude Code

BEST

Held-out: 0.701
Checkpoint: #009
Rerun: 32 steps

Validation-selected checkpoint rerun in the original environment; case return 0.712.

native env render

MiniMax-M3

Claude Code

Held-out: 0.403
Checkpoint: #009
Rerun: 100 steps

Validation-selected checkpoint rerun in the original environment; case return 0.000.

native env render

DeepSeek-V4-Pro

Claude Code

Held-out: 0.079
Checkpoint: #036
Rerun: 100 steps

Validation-selected checkpoint rerun in the original environment; case return 0.000.

ObstructedMaze

Maze navigation with blockers and interaction sequences.

MiniGridBest reported 0.911Result detail →

native env render

GPT-5.5

Codex

Held-out: 0.904
Checkpoint: #021
Rerun: 90 steps

Validation-selected checkpoint rerun in the original environment; case return 0.887.

native env render

Claude Opus 4.7

Claude Code

BEST

Held-out: 0.911
Checkpoint: #012
Rerun: 48 steps

Validation-selected checkpoint rerun in the original environment; case return 0.940.

native env render

MiniMax-M3

Claude Code

Held-out: 0.000
Checkpoint: #009
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 0.000.

native env render

DeepSeek-V4-Pro

Claude Code

Held-out: 0.000
Checkpoint: #034
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return 0.000.

Parking

Continuous vehicle control toward a target parking pose.

DrivingBest reported -30.24Result detail →

highway state renderer

GPT-5.5

Codex

BEST

Held-out: -30.24
Checkpoint: #012
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return -48.79.

highway state renderer

Claude Opus 4.7

Claude Code

Held-out: -38.18
Checkpoint: #010
Rerun: 116 steps

Validation-selected checkpoint rerun in the original environment; case return -46.70.

highway state renderer

MiniMax-M3

Claude Code

Held-out: -32.71
Checkpoint: #013
Rerun: 220 steps

Validation-selected checkpoint rerun in the original environment; case return -45.59.

highway state renderer

DeepSeek-V4-Pro

Claude Code

Held-out: -53.38
Checkpoint: #003
Rerun: 54 steps

Validation-selected checkpoint rerun in the original environment; case return -24.69.

Roundabout

Structured traffic control through a roundabout.

DrivingBest reported 10.20Result detail →

highway state renderer

GPT-5.5

Codex

Held-out: 9.818
Checkpoint: #012
Rerun: 11 steps

Validation-selected checkpoint rerun in the original environment; case return 10.25.

highway state renderer

Claude Opus 4.7

Claude Code

Held-out: 9.331
Checkpoint: #004
Rerun: 11 steps

Validation-selected checkpoint rerun in the original environment; case return 10.08.

highway state renderer

MiniMax-M3

Claude Code

Held-out: 9.521
Checkpoint: #007
Rerun: 8 steps

Validation-selected checkpoint rerun in the original environment; case return 6.500.

highway state renderer

DeepSeek-V4-Pro

Claude Code

BEST

Held-out: 10.20
Checkpoint: #009
Rerun: 11 steps

Validation-selected checkpoint rerun in the original environment; case return 10.17.

FetchPush

Goal-conditioned robotic pushing with a gripper.

RoboticsBest reported -18.16Result detail →

direct mujoco renderer

GPT-5.5

Codex

BEST

Held-out: -18.16
Checkpoint: #003
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -17.00.

direct mujoco renderer

Claude Opus 4.7

Claude Code

Held-out: -19.69
Checkpoint: #008
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -11.00.

direct mujoco renderer

MiniMax-M3

Claude Code

Held-out: -25.88
Checkpoint: #018
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -14.00.

direct mujoco renderer

DeepSeek-V4-Pro

Claude Code

Held-out: -27.66
Checkpoint: #011
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -24.00.

FetchPickAndPlace

Grasp, lift, and place an object at a target.

RoboticsBest reported -11.81Result detail →

direct mujoco renderer

GPT-5.5

Codex

BEST

Held-out: -11.81
Checkpoint: #007
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -12.00.

direct mujoco renderer

Claude Opus 4.7

Claude Code

Held-out: -13.34
Checkpoint: #008
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -13.00.

direct mujoco renderer

MiniMax-M3

Claude Code

Held-out: -13.22
Checkpoint: #006
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -13.00.

direct mujoco renderer

DeepSeek-V4-Pro

Claude Code

Held-out: -20.56
Checkpoint: #011
Rerun: 50 steps

Validation-selected checkpoint rerun in the original environment; case return -35.00.