TTT-Discover
An experiment recovery tool for scenarios in which test-time training and search are coupled. It explicitly pairs model checkpoints with search archives and adopts a fail-closed strategy to ensure state consistency.
Timeline (last 90 days)
On the Packing 26 problem, the TTT-Discover recipe reached a certified radius sum of 2.635983 within 50 steps, exactly the same as the 2.635983 reported by TTT-Discover, with a target of 2.636, where higher is better.
In each search step, TTT-Discover first does eight groups of 64 rollouts each, then performs one rank-32 LoRA update.
TTT-Discover keeps only one trajectory per task, so there is no variance estimate across random seeds.
On the Erdős minimum overlap problem, the TTT-Discover recipe of "eight groups × 64 rollouts plus one LoRA update" reached a certified C₅ upper bound of 0.38094 within 25 steps.
The target value for the Erdős minimum overlap problem is 0.38080, where lower is better; the value reported by TTT-Discover is 0.38093, and this run's result of 0.38094 is almost identical to the reported value but still has a tiny gap from the target.
The TTT-Discover recipe cited by Reddit discussants uses Qwen3-8B with thinking mode enabled, running on two B200s.
Reef Infra 的 TTT-Discover 示例给出了一种显式配对的做法：把 post-step 的 PUCT 搜索档案与已提交（committed）的模型版本绑定在一起。
TTT-Discover 的具体机制是：搜索步骤完成后，先写入一个 pending archive；只有当与之对应的模型提交也建立起来后，才把这个档案标记为 committed。
TTT-Discover 在重启时使用配对状态——档案和模型版本必须互相匹配；如果某一步缺失或不匹配，则 fail closed，不会悄悄用一份更新的检查点去启动 PUCT 搜索。
The weight recipes loop includes SAO, OpenClaw-RL, TTT-Discover, and Guidance-TTT, and requires GPUs and a full training stack.
All Facts (10)
On the Packing 26 problem, the TTT-Discover recipe reached a certified radius sum of 2.635983 within 50 steps, exactly the same as the 2.635983 reported by TTT-Discover, with a target of 2.636, where higher is better.
50%UnverifiedIn each search step, TTT-Discover first does eight groups of 64 rollouts each, then performs one rank-32 LoRA update.
50%UnverifiedTTT-Discover keeps only one trajectory per task, so there is no variance estimate across random seeds.
50%UnverifiedOn the Erdős minimum overlap problem, the TTT-Discover recipe of "eight groups × 64 rollouts plus one LoRA update" reached a certified C₅ upper bound of 0.38094 within 25 steps.
50%UnverifiedThe target value for the Erdős minimum overlap problem is 0.38080, where lower is better; the value reported by TTT-Discover is 0.38093, and this run's result of 0.38094 is almost identical to the reported value but still has a tiny gap from the target.
50%UnverifiedThe TTT-Discover recipe cited by Reddit discussants uses Qwen3-8B with thinking mode enabled, running on two B200s.
50%UnverifiedReef Infra 的 TTT-Discover 示例给出了一种显式配对的做法：把 post-step 的 PUCT 搜索档案与已提交（committed）的模型版本绑定在一起。
50%UnverifiedTTT-Discover 的具体机制是：搜索步骤完成后，先写入一个 pending archive；只有当与之对应的模型提交也建立起来后，才把这个档案标记为 committed。
50%UnverifiedTTT-Discover 在重启时使用配对状态——档案和模型版本必须互相匹配；如果某一步缺失或不匹配，则 fail closed，不会悄悄用一份更新的检查点去启动 PUCT 搜索。
50%UnverifiedThe weight recipes loop includes SAO, OpenClaw-RL, TTT-Discover, and Guidance-TTT, and requires GPUs and a full training stack.
50%