{ "experiment": "recurrent_ctx_interaction", "question": "Does recurrent depth amplify the benefit of longer context at tiny scale?", "timestamp": "2026-10-09T22:44:59Z", "config": { "d_model": 256, "n_heads": 4, "vocab": 256, "batch": 32, "steps": 2000, "lr": 0.0003, "device": "cuda" }, "results": { "recurrent_ctx64": { "final_train_loss": 1.853339433670044, "avg_train_loss_last_100": 1.8315880990028381, "best_val_loss": 1.80549016436752, "final_val_loss": 1.80549016436752, "final_val_ppl": 6.08295236438963, "val_curve": [ [ 200, 2.67759696902061 ], [ 400, 2.4178282270626146 ], [ 600, 2.2530713908526363 ], [ 800, 2.111119897024972 ], [ 1000, 2.0204621748048432 ], [ 1200, 1.9230032847852123 ], [ 1400, 1.8634966042577004 ], [ 1600, 1.827961488645904 ], [ 1800, 1.8110538823263986 ], [ 2000, 1.80549016436752 ] ], "params": 1397504, "train_time_s": 10.449508428573608 }, "recurrent_ctx256": { "final_train_loss": 2.2531955242156982, "avg_train_loss_last_100": 2.245353183746338, "best_val_loss": 2.235594819516552, "final_val_loss": 2.235594819516552, "final_val_ppl": 9.352042975781874, "val_curve": [ [ 200, 2.5640992836076384 ], [ 400, 2.4392180549855134 ], [ 600, 2.3669187321954843 ], [ 800, 2.323085509514322 ], [ 1000, 2.2937028310736833 ], [ 1200, 2.2744714941297257 ], [ 1400, 2.2546324447709685 ], [ 1600, 2.245326196904085 ], [ 1800, 2.2375385070333675 ], [ 2000, 2.235594819516552 ] ], "params": 1446656, "train_time_s": 25.86048126220703 }, "stacked_ctx64": { "final_train_loss": 1.8972289562225342, "avg_train_loss_last_100": 1.9378939366340637, "best_val_loss": 1.9216734706139078, "final_val_loss": 1.9216734706139078, "final_val_ppl": 6.832382699238212, "val_curve": [ [ 200, 2.790465888198541 ], [ 400, 2.4131838263297567 ], [ 600, 2.278608127516143 ], [ 800, 2.14184259103269 ], [ 1000, 2.0658977236066547 ], [ 1200, 2.006630532595576 ], [ 1400, 1.9629625875122694 ], [ 1600, 1.9364350644909605 ], [ 1800, 1.9249796546235376 ], [ 2000, 1.9216734706139078 ] ], "params": 1402880, "train_time_s": 7.927839279174805 }, "stacked_ctx256": { "final_train_loss": 2.2880020141601562, "avg_train_loss_last_100": 2.284539570808411, "best_val_loss": 2.2725286454570535, "final_val_loss": 2.2725286454570535, "final_val_ppl": 9.703907558344023, "val_curve": [ [ 200, 2.7255869145296057 ], [ 400, 2.4116121185069184 ], [ 600, 2.3523950440543038 ], [ 800, 2.3312676225389755 ], [ 1000, 2.3142523259532695 ], [ 1200, 2.298062850991074 ], [ 1400, 2.284544699532645 ], [ 1600, 2.278283355673965 ], [ 1800, 2.2737569964661892 ], [ 2000, 2.2725286454570535 ] ], "params": 1452032, "train_time_s": 13.692972898483276 } }, "analysis": { "recurrent_ctx_gain": -3.2690906113922438, "stacked_ctx_gain": -2.8715248591058105, "recurrent_advantage_ctx64": 0.749430334848582, "recurrent_advantage_ctx256": 0.3518645825621487 } }