id stringlengths 36 36 | tier stringclasses 1
value | github_login stringclasses 179
values | created_at stringlengths 32 32 | status stringclasses 2
values | sha256 stringlengths 64 64 | source_bytes int64 341 260k | leaderboard_rank int64 | source large_stringlengths 341 260k | metadata_json large_stringlengths 645 734 | result_json large_stringlengths 5 4.8k |
|---|---|---|---|---|---|---|---|---|---|---|
137b6c75-6a85-48f7-a274-8e2c115a418a | easy | Bananafly | 2026-08-11 20:03:27.822341+00:00 | succeeded | d1dacc245b29e820f08d09ff06c74aef656dba16319d2cd9a7be7c156c33f878 | 5,111 | null | """Stock one-pass baseline with controlled small-weight initialization."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
assert_model_state,
)
D_MODEL = ... | {
"id": "137b6c75-6a85-48f7-a274-8e2c115a418a",
"created_at": "2026-08-11 20:03:27.822341+00:00",
"db_md5": "d64b671b9bc444ed8929c5d6708073ad",
"submitter": "Andre Kreidemann",
"github_login": "Bananafly",
"run_id": "0dcac44e-b796-4172-a0bf-8ad40356b86a",
"tier": "easy",
"dataset_id": "e5",
"status": ... | {
"score": {
"mean_loss": 2.3424338731952834,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.0016666666666666668
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.344159558214949,
"example_count": 60... |
1383dedd-8aae-4fba-8f97-33de8fd0cd2e | easy | oupadhyay | 2026-08-24 00:07:26.123888+00:00 | succeeded | 74e335aff210882666ef21fd02ded9051b546cf08082f6d6ef4241b5d0a7596f | 4,315 | null | """Conditional structured-loss screen: hard50."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import nn
from benchmark import ModelSpec,OptimizerSpec,OptimizerBundle,Submission,TokenLossBatch,assert_model_state
W,D,H=4,96,48
class C:
def __init__(s,vocab_size,max_... | {
"id": "1383dedd-8aae-4fba-8f97-33de8fd0cd2e",
"created_at": "2026-08-24 00:07:26.123888+00:00",
"db_md5": "c14f78d41fb84fe23cc4426349f3175f",
"submitter": "Ojasw Upadhyay",
"github_login": "oupadhyay",
"run_id": "ca0023cb-b660-40ef-8d11-da6aabbd608b",
"tier": "easy",
"dataset_id": "e5",
"status": "s... | {
"score": {
"mean_loss": 2.553661578718752,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.007916666679084302
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.3760102661201237,
"example_count": 600... |
13842023-3098-4a6c-b63b-b9db981f730c | easy | jaipise | 2026-08-15 15:47:53.617590+00:00 | succeeded | 7bf30e19d90e7f6cadb4ad67d0b2cd51634a4fd2d29a507be6f840a0dd78f8c1 | 3,628 | null | """Basic single-pass Transformer with PyTorch AdamW."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
assert_model_state,
)
D_MODEL = 128
NUM_HEADS = 4
... | {
"id": "13842023-3098-4a6c-b63b-b9db981f730c",
"created_at": "2026-08-15 15:47:53.617590+00:00",
"db_md5": "eaac69fe345f82973fcf392c504a40ff",
"submitter": "Jai Pise",
"github_login": "jaipise",
"run_id": "f5b27626-bf85-45ed-a866-7f4d4b86fa25",
"tier": "easy",
"dataset_id": "e1",
"status": "succeeded... | {
"score": {
"mean_loss": 3.0487178564071655,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.021666666492819786
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 4.027981281280518,
"example_count": 100... |
1387a7d3-c0c9-433e-a660-1a98095cfcbf | easy | alirezashirvani-jr | 2026-08-29 09:47:17.381453+00:00 | succeeded | 2b0223ff035c5c9c551319a2bfe8d9af881764d80db781b124aabaee8d426ac1 | 32,601 | null | from __future__ import annotations
import math
import time
from contextlib import nullcontext
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
TokenLossBatch,
assert_model_state,
)
MODEL_... | {
"id": "1387a7d3-c0c9-433e-a660-1a98095cfcbf",
"created_at": "2026-08-29 09:47:17.381453+00:00",
"db_md5": "050524d36ff9ab09e87d1ff83923e1e0",
"submitter": "alirezashirvani-jr",
"github_login": "alirezashirvani-jr",
"run_id": "8e856825-70d1-47b7-9239-c307808fe6bc",
"tier": "easy",
"dataset_id": "e9",
... | {
"score": {
"mean_loss": 1.1867033243179321,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.7666666805744171
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 0.6008268594741821,
"example_count": 90,
... |
1387f3cc-1084-4157-9f9c-efcdeb773570 | easy | Maxwells-Demons | 2026-08-07 18:35:26.365261+00:00 | succeeded | e02589fc28906fc867260c1b88cd94a039a77c663d39999b512dca9f2d9fdf7b | 10,654 | null | """v47: the transition as one wide MLP over the digit state and the modulus.
Two measurements drive this shape.
1. The tied attention block (v43-v46) tops out around mean 0.75-0.79 on e1 while
the older flat-MLP transition reached 0.87 on the same data. Attention mixes
slots only through a low-rank path; a dig... | {
"id": "1387f3cc-1084-4157-9f9c-efcdeb773570",
"created_at": "2026-08-07 18:35:26.365261+00:00",
"db_md5": "e5de631c984a9d94181b4fd4f1fc1af6",
"submitter": "Maxwells_Ayakashi",
"github_login": "Maxwells-Demons",
"run_id": "afcff625-0477-46d4-b4bb-aed0a52c46ec",
"tier": "easy",
"dataset_id": "e3",
"st... | {
"score": {
"mean_loss": 21.565858012949462,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.05187500000465661
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 21.391893692562245,
"example_count": 800... |
138c444c-1c93-4e92-8616-c10a9020c293 | easy | AFeinfield | 2026-08-25 01:29:18.793393+00:00 | succeeded | 6e6f2a1691d15d4d9b1629611908162c7c19d6c008900794deff30ab5e2cb280 | 7,714 | null | """Looped (weight-tied) bidirectional Transformer, d=256, 8 iterations.
023: 022 arch on e5 (variable N x variable T — most Hard-like easy set)
fit-first read on the joint-conditioning regime
to amortize the evaluator's ~200ms/iteration overhead into 8x more steps.
prior. Same parameter count order as a 1-block model;... | {
"id": "138c444c-1c93-4e92-8616-c10a9020c293",
"created_at": "2026-08-25 01:29:18.793393+00:00",
"db_md5": "6258a800f3d27ce92f882f5779c83472",
"submitter": "AFeinfield",
"github_login": "AFeinfield",
"run_id": "82c9257d-c458-4623-bacd-73c710feb3c2",
"tier": "easy",
"dataset_id": "e5",
"status": "succ... | {
"score": {
"mean_loss": 3.8872570486179647,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.005000000049670538
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.9707765686137795,
"example_count": 60... |
138cd638-eec5-43bb-9e0b-feb411638812 | easy | lpbb | 2026-08-13 12:03:28.635027+00:00 | failed | b590bea3f18f36ffa4885779cf2914a5db9af0563078980c413fb94d328846ce | 4,523 | null | """S1 — one tied cell looped 12*T times with prompt recall, plain final CE loss."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
assert_model_state,
)
CO... | {
"id": "138cd638-eec5-43bb-9e0b-feb411638812",
"created_at": "2026-08-13 12:03:28.635027+00:00",
"db_md5": "542ea03bfeeedd7e1e9976a8c65bdbf2",
"submitter": "Lpbb",
"github_login": "lpbb",
"run_id": "d0d0f015-47dd-4497-bd5b-d9a4d8211f73",
"tier": "easy",
"dataset_id": "e3",
"status": "failed",
"star... | null
|
1394f4fa-b892-45e1-b4f5-3362c74834b4 | easy | jagilley | 2026-08-21 23:43:48.989585+00:00 | succeeded | 1292624688cc0f831a21af251b965fba1ebf96b19d52e45b86bea6a942da462e | 21,757 | null | """One Layer Deeper submission — a tied operator applied `T` times, with an optional
label-free closure loss, ported from `experiments/one_layer_deeper/ballistic_depth/`.
Two arms, selected by `ARM` below. They share **every parameter of the model**; only the
loss differs, which is how the research node runs its arms ... | {
"id": "1394f4fa-b892-45e1-b4f5-3362c74834b4",
"created_at": "2026-08-21 23:43:48.989585+00:00",
"db_md5": "5a7cf7688ef9c2345cedf23a7bab54c0",
"submitter": "Jasper",
"github_login": "jagilley",
"run_id": "b959ea4a-5816-4316-9a14-adca79e97ca2",
"tier": "easy",
"dataset_id": "e5",
"status": "succeeded"... | {
"score": {
"mean_loss": 2.9735972679889895,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.006250000012417635
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.8548887994791894,
"example_count": 60... |
13985f00-083b-4e72-872d-119174371b1d | easy | shreyash-chonkie | 2026-08-24 20:35:40.467830+00:00 | succeeded | d64bbafa4a82d5d78982aec75998bcfec3e3d416bb3f8c44ff63efcf77aa93f5 | 15,459 | null | """Recurrent attention with improvement-weighted pool feedback."""
from __future__ import annotations
import math
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
TokenLossBatch,
assert_mo... | {
"id": "13985f00-083b-4e72-872d-119174371b1d",
"created_at": "2026-08-24 20:35:40.467830+00:00",
"db_md5": "1253f45c068549baf3c3f4905f581fdf",
"submitter": "Shreyash",
"github_login": "shreyash-chonkie",
"run_id": "c7d69e7b-3d0b-4a0e-9734-2a054f66ef62",
"tier": "easy",
"dataset_id": "e6",
"status": "... | {
"score": {
"mean_loss": 2.7396801114082336,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.05022522807121277
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.5259931087493896,
"example_count": 60,... |
13992870-82d9-44ca-8196-a5be3354f5df | easy | sapient-sapiens | 2026-08-29 22:07:36.169765+00:00 | succeeded | a2ddac84b6b4e687d6573aea21adf405908ab71bb3a35ac49418d7bc94908e58 | 15,077 | null | """Three-loop one-layer Transformer for iterated modular squaring.
The learned arithmetic operator is one tied pre-norm Transformer block. Every
square makes four Horner-digit calls and every digit makes nine micro-calls.
N and the complete operand are immutable during a square; scratch resets at a
digit boundary; an... | {
"id": "13992870-82d9-44ca-8196-a5be3354f5df",
"created_at": "2026-08-29 22:07:36.169765+00:00",
"db_md5": "22d2366cce6c3d58baa7f7b76f05f1e5",
"submitter": "Ertondy",
"github_login": "sapient-sapiens",
"run_id": "6975f1f0-93a8-46be-8fd1-c1d5b5295d03",
"tier": "easy",
"dataset_id": "e5",
"status": "su... | {
"score": {
"mean_loss": 2.2117171283435537,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.0141666666790843
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.214722651537224,
"example_count": 600,
... |
139f73ad-69da-492c-95f1-1068a09494dd | easy | ehonig | 2026-08-20 08:22:49.288919+00:00 | succeeded | 41797050fc5c1a4860698527f0ef3330c4044c0c8eabf5ddf2dfa89a813d678b | 23,251 | null | """A plain Transformer for repeated modular squaring, and nothing else yet.
This is a deliberate restart. The previous submission accumulated a field
parser, a carry scan, a learned reciprocal, periodic features, a digit
bottleneck, a depth selector and two cell types, most of them tuned around a
memorisation table th... | {
"id": "139f73ad-69da-492c-95f1-1068a09494dd",
"created_at": "2026-08-20 08:22:49.288919+00:00",
"db_md5": "7edca7f62fd625c6299ee50eefef2f89",
"submitter": "Edouardo Honig",
"github_login": "ehonig",
"run_id": "85e2d305-e05e-4a27-9c8a-7c18ad2b3430",
"tier": "easy",
"dataset_id": "e1",
"status": "succ... | {
"score": {
"mean_loss": 2.0970704555511475,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.07666666433215141
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.085104465484619,
"example_count": 100,... |
13a058b7-ec29-4ed1-8692-26d517ec9de8 | easy | AFeinfield | 2026-08-18 23:43:41.332042+00:00 | succeeded | 6bb044a895217f39cca056017c0554724f648c1c554b178429dd30e096c0f86b | 13,734 | null | """084: wide conv (kernel 9) + carry-GRU removed — the two best
fit-speed findings combined. 083 hit 100% train on e2; 081 gave +20%
steps/s. Generalization must be re-checked on m6.
Original 083 notes follow.
083: conv kernel 5 -> 9 in the multiply sublayer.
Direct test of the mechanism story: the conv gathers neighb... | {
"id": "13a058b7-ec29-4ed1-8692-26d517ec9de8",
"created_at": "2026-08-18 23:43:41.332042+00:00",
"db_md5": "4c146f4e4f1b8669095b715d12011ce4",
"submitter": "AFeinfield",
"github_login": "AFeinfield",
"run_id": "466bd830-fdad-4f9f-aaaf-323ab0acccdf",
"tier": "easy",
"dataset_id": "e2",
"status": "succ... | {
"score": {
"mean_loss": 16.627143383026123,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.013541667256504297
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 17.384193420410156,
"example_count": 30... |
13a21e76-5aee-4518-9e4e-e7c205b90e19 | easy | sapient-sapiens | 2026-08-18 11:22:01.204438+00:00 | succeeded | 85729495efc1f5aae59c5f5b8f16f41efb58a5c8dbb929b814f0c89a0e3c3276 | 26,969 | null | """v44 adaptive baseline: one encoder, four-layer T-free core, bounded halt.
Locked routing: field-local one-layer encoder; recurrent core sees only N/X;
T is pooled solely for the controller; decode once on a halt-weighted state
mix. Stopping support is 8 depths, not 64. T is never parsed in forward or
the loss.
"""
... | {
"id": "13a21e76-5aee-4518-9e4e-e7c205b90e19",
"created_at": "2026-08-18 11:22:01.204438+00:00",
"db_md5": "435dabebe848b3a53725188abe315d21",
"submitter": "Ertondy",
"github_login": "sapient-sapiens",
"run_id": "8a248d5d-b1a4-4fab-9613-3dee6d4e6fcc",
"tier": "easy",
"dataset_id": "e9",
"status": "su... | {
"score": {
"mean_loss": 1.8596693922950878,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.03222222276031971
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 1.865815281867981,
"example_count": 90,
... |
13abb5ce-b8dc-4904-bb22-5bbf0b238100 | easy | rteehas | 2026-08-19 08:21:40.193039+00:00 | succeeded | 857818a3c205ebcc347bc054afc60dab80f6dc8e31540f215f193fbbe21c166a | 41,409 | null | """Anonymous marginal-program submission.
Mechanism
---------
The model is a differentiable probabilistic program. Its forward pass
computes, exactly, the answer-token predictions of a small structured latent
hypothesis space:
(prompt segmentation = which token types head the three fields)
x (field-role permutat... | {
"id": "13abb5ce-b8dc-4904-bb22-5bbf0b238100",
"created_at": "2026-08-19 08:21:40.193039+00:00",
"db_md5": "d0b4777362f88dadc43578b2f35d8934",
"submitter": "Ryan Teehan",
"github_login": "rteehas",
"run_id": "1b20a72b-c68a-4069-9096-783d4c8004a9",
"tier": "easy",
"dataset_id": "e4",
"status": "succee... | {
"score": {
"mean_loss": 6.922630187631873,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.006354166679084301
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 6.95901619253553,
"example_count": 1200,... |
13ad0858-5e67-4c2a-a477-4f4837d55859 | easy | karanganesan | 2026-08-07 09:20:47.892439+00:00 | succeeded | 78d822b7d1f9bb5b46f977e12ceb542f196ddd50f2f7a90f51b4f9012f406971 | 25,547 | null | """Parametric looped-transformer family (P1).
One weight-tied transformer block applied k times in latent space. Config
flags cover four P1 families with one file:
looped recall=0 gated=0 tfilm=0 plain weight-tied loop
looped-recall recall=1 re-inject the input embedding each
... | {
"id": "13ad0858-5e67-4c2a-a477-4f4837d55859",
"created_at": "2026-08-07 09:20:47.892439+00:00",
"db_md5": "c6174b599773d987b1c6e74c1c2f0956",
"submitter": "Karan Ganesan",
"github_login": "karanganesan",
"run_id": "5c3442da-62fe-4ee0-8bfb-5c40dea4c751",
"tier": "easy",
"dataset_id": "e4",
"status": ... | {
"score": {
"mean_loss": 6.732511758804321,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.00437499990221113
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 7.163880825042725,
"example_count": 1200,... |
13b24e2a-31dc-49c6-a3f8-73c0d18d25d5 | easy | paulpister | 2026-08-04 17:42:00.847856+00:00 | succeeded | 7432f246240eb29b39847ce1acc85ccc897d45dbf8196795f7f5e31b11340c3e | 15,935 | null | """Loop-T tied recurrence v8: v6 + T-curriculum loss + depth-aware core init.
Design:
- Structural features computed from input_ids (field segment, digit place
value from the little end, reverse position from sequence end). Sinusoidal
place/revpos features so longer OOD-N numbers extrapolate smoothly.
- T is parse... | {
"id": "13b24e2a-31dc-49c6-a3f8-73c0d18d25d5",
"created_at": "2026-08-04 17:42:00.847856+00:00",
"db_md5": "b623331304f5c7b6d9e459c1329035ca",
"submitter": "paulv",
"github_login": "paulpister",
"run_id": "235fd83f-1b8e-4052-afe4-2bd6a16c0d6b",
"tier": "easy",
"dataset_id": "e1",
"status": "succeeded... | {
"score": {
"mean_loss": 2.8979979753494263,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.08500000089406967
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.6213784217834473,
"example_count": 100... |
13b6e5e8-9ff5-433c-be00-e1c888296e55 | easy | gauravmishra | 2026-08-08 01:43:53.195145+00:00 | succeeded | af789f16bda135f12701f5e4f1539519eff075de301671f92b85f3074e767415 | 3,917 | null | """Four-pass Transformer with four independent blocks and 10% residual dropout."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
assert_model_state,
)
D_... | {
"id": "13b6e5e8-9ff5-433c-be00-e1c888296e55",
"created_at": "2026-08-08 01:43:53.195145+00:00",
"db_md5": "4c747e1c7fa73b9eb545b7c37099b2dc",
"submitter": "Gaurav Mishra",
"github_login": "gauravmishra",
"run_id": "f0d576ce-b9b6-4843-be91-e9e3c26a25ee",
"tier": "easy",
"dataset_id": "e5",
"status": ... | {
"score": {
"mean_loss": 3.9466909012271767,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.007083333333333334
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 4.266020751320193,
"example_count": 600... |
13b98735-64cc-4e8c-8efa-d0ac2c5e6a15 | easy | viridale | 2026-08-15 06:53:31.147926+00:00 | succeeded | d8b603ef9c1a7a8a091546b8aaccbcf4c2bbbbed6438bfd1b557b82c5ea54e06 | 15,882 | null | """Raw additive i+j measure for the learned translation-shared scan.
hypothesis: v633 erases the mass of coefficient convolution by sqrt-count scaling and
per-slot RMS normalization; preserving the raw scatter-add measure should expose a
length-extensible interaction even before its known support defect is rep... | {
"id": "13b98735-64cc-4e8c-8efa-d0ac2c5e6a15",
"created_at": "2026-08-15 06:53:31.147926+00:00",
"db_md5": "90f060978ff75d1aae0841e87f452187",
"submitter": "priormancer",
"github_login": "viridale",
"run_id": "01313ae8-f553-487a-b8e1-8400f8510d43",
"tier": "easy",
"dataset_id": "e6",
"status": "succe... | {
"score": {
"mean_loss": 2.3138102889060974,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.32500001788139343
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 1.8493751287460327,
"example_count": 60,... |
13ba1e76-97d2-41f8-8d4f-2340ee4c4e93 | easy | erdavis0 | 2026-08-29 01:46:13.485416+00:00 | succeeded | f077398b7175491463614d7d5f9797f7e906aed49fee3a5dc6a084b27c7de37e | 33,177 | null | """Typed-prefix carrier product-VQ B0 with separate X and N codebooks.
Public delimiter syntax routes opaque N, X, and T fields into three strict
functional roles. X initializes only an evolving auxiliary tape, N initializes
only an invariant environment lane, and T reaches only a learned monotone
prefix controller. ... | {
"id": "13ba1e76-97d2-41f8-8d4f-2340ee4c4e93",
"created_at": "2026-08-29 01:46:13.485416+00:00",
"db_md5": "c87995e7bae5b04ce68954761d511271",
"submitter": "Ethan",
"github_login": "erdavis0",
"run_id": "6e626d64-c449-4285-ab73-4304d2d3d0dc",
"tier": "easy",
"dataset_id": "e7",
"status": "succeeded",... | {
"score": {
"mean_loss": 3.7868927717208862,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.0441622119396925
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 4.466894626617432,
"example_count": 85,
... |
13c4dafa-2fcc-4f8b-947e-14b486060415 | easy | sapient-sapiens | 2026-08-17 02:45:00.798249+00:00 | succeeded | 96d137d90e52e00d4bfe7084329085f05efee0981345708cdf647be0670dae02 | 15,120 | null | """Fixed-4 recurrence candidate: v70 source without adaptive halt.
Parent: act64_e6_w1024_dh4_L6_wd1_v70 (width 1024, d_h=4, L=6, WD=1.0).
Primary variable: adaptive ACT over 1--64 (typically hard-halt 2) versus a
fixed four applications of the shared core. The halt controller is removed.
T remains an ordinary prompt ... | {
"id": "13c4dafa-2fcc-4f8b-947e-14b486060415",
"created_at": "2026-08-17 02:45:00.798249+00:00",
"db_md5": "a360270b1bd62e5c960d2cc497021af4",
"submitter": "Ertondy",
"github_login": "sapient-sapiens",
"run_id": "15e60f0c-8356-4e7c-8e5b-a15ca060406c",
"tier": "easy",
"dataset_id": "e4",
"status": "su... | {
"score": {
"mean_loss": 2.0808408865332497,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.006354166666666667
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.082646403968423,
"example_count": 120... |
13c62c67-a993-4892-a8c7-96718a86d5e2 | easy | epicosy | 2026-08-22 23:13:21.008721+00:00 | succeeded | ace2f1895b8e10d235e6e5d2b879cf42f08ac8becaa36d40dd1aefb58b3d8e14 | 5,430 | null | """Latent-state Universal Transformer baseline for the Easy tier.
The prompt stream is encoded once and kept fixed. A single shared block
recurrently updates learned latent/query tokens through cross-attention to the
prompt. The final logits remain aligned with the original input sequence, as
required by the public su... | {
"id": "13c62c67-a993-4892-a8c7-96718a86d5e2",
"created_at": "2026-08-22 23:13:21.008721+00:00",
"db_md5": "8678e29543bec659c082834b004aacdb",
"submitter": "Eduard Pinconschi",
"github_login": "epicosy",
"run_id": "9590dc98-bb88-4f36-ada7-e2103c23170b",
"tier": "easy",
"dataset_id": "e1",
"status": "... | {
"score": {
"mean_loss": 2.282889485359192,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.05833333358168602
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.312285900115967,
"example_count": 100,
... |
13d9b335-3cd2-4a86-be49-21a8ff639651 | easy | khushidahi | 2026-08-02 00:54:55.244754+00:00 | succeeded | 4e6cb25ffcba265c2644563ca078e62b406821b912bd14bdfe1969f9c9e0e092 | 7,576 | null | """Prelude -> recalled shared core -> coda model for One Layer Deeper.
R2 architecture reset:
- use a real one-block prelude and one-block coda;
- apply one shared recurrent Transformer core four times;
- inject immutable prelude context at every recurrent update;
- use sandwich RMS normalization for stable repeated a... | {
"id": "13d9b335-3cd2-4a86-be49-21a8ff639651",
"created_at": "2026-08-02 00:54:55.244754+00:00",
"db_md5": "ea30d4b1d5be7defe698bd3308a745f1",
"submitter": "khushidahi",
"github_login": "khushidahi",
"run_id": "d3ad24a6-e2bb-46c7-96c7-bc407fb5f38c",
"tier": "easy",
"dataset_id": "e5",
"status": "succ... | {
"score": {
"mean_loss": 2.1780117750167847,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.0054166666232049465
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.1852827072143555,
"example_count": 6... |
13e2f870-ee06-4e9c-bdf9-2e71e665b7ab | easy | yashkant | 2026-08-25 11:09:19.803708+00:00 | succeeded | 900ac22b1d10790e5083e8b0dc3041e050db83ddf980bcaac92fa32d24639825 | 24,709 | null | """One Layer Deeper HARD submission v7: prime-agnostic repeated-squaring digit machine.
The entire model is three tied digit cells (~86K params):
mul: (digit, digit, carry 0..8) -> digit, carry [schoolbook partial products]
add: (digit, digit, carry 0/1) -> digit, carry [partial accumulation, N-doubling]
... | {
"id": "13e2f870-ee06-4e9c-bdf9-2e71e665b7ab",
"created_at": "2026-08-25 11:09:19.803708+00:00",
"db_md5": "8b11c17bf3d832af322c69dd402a3028",
"submitter": "Yash Kant",
"github_login": "yashkant",
"run_id": "83968aae-a285-4737-92df-33eb61a0f347",
"tier": "easy",
"dataset_id": "e5",
"status": "succeed... | {
"score": {
"mean_loss": 2.2778506048426435,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.07041666663562259
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.3462400436401367,
"example_count": 600... |
13e56cd2-3b70-4681-b959-74849f04c654 | easy | SunstarKK | 2026-08-18 10:34:21.529259+00:00 | succeeded | 09e74b4254454452c9b67140bdbe2f731da57a60f24e6ddcbab486a2752f5dbb | 15,164 | null | """Scalar Fourier neural operator for One Layer Deeper."""
from __future__ import annotations
import math
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
TokenLossBatch,
assert_model_stat... | {
"id": "13e56cd2-3b70-4681-b959-74849f04c654",
"created_at": "2026-08-18 10:34:21.529259+00:00",
"db_md5": "48296e1f92d6a1dcd2b6335d8682c7a8",
"submitter": "SunstarK",
"github_login": "SunstarKK",
"run_id": "8753e7c3-17f8-48db-9fec-8ca693587ff2",
"tier": "easy",
"dataset_id": "e1",
"status": "succeed... | {
"score": {
"mean_loss": 1.9215131998062134,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.6866666674613953
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.1789982318878174,
"example_count": 100,... |
13e79b45-323e-49ad-8786-9245751ac4cb | easy | TZZheng | 2026-08-07 05:07:39.181055+00:00 | succeeded | 0c1662ea1dd9ae423b733ac49356d0d714783379a76fd72e75e5310804cdc96f | 12,528 | null | """Learned recurrent modular-transition model.
The mutable digit state is initialized from x. A learned square-oriented
self-mixing network and an N-conditioned reduction network form one transition,
and exactly the same transition parameters are reused T times. No arithmetic
operation is hard-coded; only the public p... | {
"id": "13e79b45-323e-49ad-8786-9245751ac4cb",
"created_at": "2026-08-07 05:07:39.181055+00:00",
"db_md5": "c1ee577be7abbc21a923ce694dc67829",
"submitter": "TZZheng",
"github_login": "TZZheng",
"run_id": "5ea8d72c-78db-4f49-a3e1-aace7d565271",
"tier": "easy",
"dataset_id": "e1",
"status": "succeeded"... | {
"score": {
"mean_loss": 8.405210256576538,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.04333333298563957
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 10.485795021057129,
"example_count": 100,... |
13eb1816-209b-4579-977b-100958b2373f | easy | AFeinfield | 2026-08-22 14:48:18.940476+00:00 | succeeded | b44d3a600d938bd970958a912e9d8a0bf15ff5aee691b5b4bfd7a597fce3bff0 | 22,371 | null | """103: RARE-LOW-T ARENA (filtered) + group-balanced loss
See the RARE-LOW-T ARENA note above token_training_loss.
Original notes follow.
092: INSTRUMENTED BASE — two measurement fixes, no architecture change.
(1) Step-indexed LR schedule with a fixed screening step budget. The old
schedule decayed on WALL CLOCK... | {
"id": "13eb1816-209b-4579-977b-100958b2373f",
"created_at": "2026-08-22 14:48:18.940476+00:00",
"db_md5": "10290afd08e78337d25a230e15118389",
"submitter": "AFeinfield",
"github_login": "AFeinfield",
"run_id": "6485ee43-2d07-424b-bade-19401050203d",
"tier": "easy",
"dataset_id": "e10",
"status": "suc... | {
"score": {
"mean_loss": 2.04192852973938,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.01705050515010953
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.042482376098633,
"example_count": 125,
... |
13ebf511-5009-4ae0-8376-e62f4ac08b8a | easy | khushidahi | 2026-08-06 02:51:57.191775+00:00 | succeeded | 4f93e0988825c339ce923c26aa72c55cf9848dd31fa4f0d76be8de6e1a9743fe | 14,389 | null | """R6 structured decimal-workspace model for One Layer Deeper.
The model does not implement multiplication, modular reduction, or the public
recurrence. It uses the public tokenizer structure to build learned, right-
aligned decimal tapes for N, X, and T, then applies a shared neural transition
to a mutable work tape.... | {
"id": "13ebf511-5009-4ae0-8376-e62f4ac08b8a",
"created_at": "2026-08-06 02:51:57.191775+00:00",
"db_md5": "53941de636b0d60f765aa6eb9e1b1476",
"submitter": "khushidahi",
"github_login": "khushidahi",
"run_id": "0627e622-b8f3-47b3-9851-d560a4f22069",
"tier": "easy",
"dataset_id": "e2",
"status": "succ... | {
"score": {
"mean_loss": 2.235159993171692,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.014791666995733976
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.2334556579589844,
"example_count": 300... |
13ef27d5-8c6e-4284-a526-2318f205af66 | easy | oupadhyay | 2026-08-29 17:31:53.720327+00:00 | succeeded | ac990a8c6c23bd6f297e4dc30a65d4e18a68ae08f66a480ca8f70184600df301 | 3,991 | null | """Conditional optimizer screen: lr15."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import nn
from benchmark import ModelSpec,OptimizerSpec,OptimizerBundle,Submission,assert_model_state
W,D,H=4,64,32
class C:
def __init__(s,vocab_size,max_seq_len):s.vocab_size,s... | {
"id": "13ef27d5-8c6e-4284-a526-2318f205af66",
"created_at": "2026-08-29 17:31:53.720327+00:00",
"db_md5": "967aba6008b829b65f37ee4449c22fea",
"submitter": "Ojasw Upadhyay",
"github_login": "oupadhyay",
"run_id": "37583342-d5f4-4a2a-8720-75b7ebbf876b",
"tier": "easy",
"dataset_id": "e6",
"status": "s... | {
"score": {
"mean_loss": 1.602417767047882,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.4286036342382431
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 1.1916471719741821,
"example_count": 60,
... |
13f07a74-1945-4fd6-8dbe-c661a9ba8aad | easy | im-a-king-but-im-far-from-a-saint | 2026-08-11 06:50:34.595981+00:00 | succeeded | f2b26833964109079a78cd6274e9d941e13874e018f66de36fc742351956446f | 8,930 | null | """Weight-tied recurrent (Universal-Transformer-style) baseline.
Applies ONE shared transformer block NUM_LOOPS times instead of stacking
distinct layers. Re-injecting the original embedding at every iteration
keeps the recurrence anchored to the problem statement, which is what lets
this kind of network extrapol... | {
"id": "13f07a74-1945-4fd6-8dbe-c661a9ba8aad",
"created_at": "2026-08-11 06:50:34.595981+00:00",
"db_md5": "28bc9e1256cb9f6556a42c01b570a00f",
"submitter": "im-a-king-but-im-far-from-a-saint",
"github_login": "im-a-king-but-im-far-from-a-saint",
"run_id": "08426e87-29c0-4617-b9eb-ad7b183e0752",
"tier": "... | {
"score": {
"mean_loss": 2.2278494834899902,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.026666666381061077
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.2293386459350586,
"example_count": 10... |
13f14514-6462-4e23-86fd-bd8760fa9502 | easy | erdavis0 | 2026-08-26 04:01:30.199940+00:00 | succeeded | f323ea27a95b25c17ef240c64fd37d3915d7f3e1d86c503b6dc262e4235c4217 | 30,352 | null | """Latent-role continuous REV C1: selector-oracle distillation.
Public delimiter syntax only routes opaque token categories into distinct,
right-aligned N, X, and T tapes. Learned local cells compile separate
interaction, carry, and modulus-conditioned workspaces. N is isolated from
interaction and carry and first e... | {
"id": "13f14514-6462-4e23-86fd-bd8760fa9502",
"created_at": "2026-08-26 04:01:30.199940+00:00",
"db_md5": "1658fcec88828e061d8640031c475500",
"submitter": "Ethan",
"github_login": "erdavis0",
"run_id": "78147d5c-74d1-4261-936c-abb86fcb93f3",
"tier": "easy",
"dataset_id": "e2",
"status": "succeeded",... | {
"score": {
"mean_loss": 3.584113597869873,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.032500000670552254
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.5345613956451416,
"example_count": 300... |
13f40f1e-7dcf-4578-8a95-228eb9b870bd | easy | jordanrubin | 2026-08-10 20:28:53.022051+00:00 | succeeded | d66d3596d565e364e36b2e65e3e2d2e8d5bb226368b36c4f9dd3d714e86dfa54 | 19,332 | null | """Weight-tied MLP-cell loop over soft digit states (mlploop).
Cell = MLP over [expected digits, pairwise digit products, RNS residue
simplices (fixed differentiable mixing over Z_p), N digits]. The cell that
learns one-step modular squaring from direct pairs (52% unseen at 14k rows,
day-20 screen) inside the exact-T ... | {
"id": "13f40f1e-7dcf-4578-8a95-228eb9b870bd",
"created_at": "2026-08-10 20:28:53.022051+00:00",
"db_md5": "881e1109fb87d3faf946c8d57e92a5d3",
"submitter": "Jordan Rubin",
"github_login": "jordanrubin",
"run_id": "67019bd8-877c-4ea2-bb64-96e60b36c251",
"tier": "easy",
"dataset_id": "e5",
"status": "s... | {
"score": {
"mean_loss": 2.5895534655976893,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.007499999832361937
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.1691552769442843,
"example_count": 60... |
13f472d3-886b-4607-8118-e8499f4ddb95 | easy | sapient-sapiens | 2026-08-30 21:25:20.267517+00:00 | succeeded | 55861fb9677d8bad7758eb7eda678868891b47b52451d0b83f501ec65b4c9af6 | 16,158 | null | """Eligible direct-Horner Transformer with parameter-free A-commit RMS matching.
Each outer square has four fixed Horner positions. S resets at each position,
then one tied call commits latent S before a second tied call commits A. The
full model therefore makes 8*4*2=64 calls. Each accepted A write is RMS-matched
... | {
"id": "13f472d3-886b-4607-8118-e8499f4ddb95",
"created_at": "2026-08-30 21:25:20.267517+00:00",
"db_md5": "ad0b23963f5a5710954e010e11bfbc93",
"submitter": "Ertondy",
"github_login": "sapient-sapiens",
"run_id": "807704a1-4cdd-4a51-91c8-8ca632af0744",
"tier": "easy",
"dataset_id": "e5",
"status": "su... | {
"score": {
"mean_loss": 2.6138517055092354,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.006250000012417635
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.6503473996046947,
"example_count": 60... |
13f48276-8763-40e6-8e05-308e523453fc | easy | yashkant | 2026-08-09 02:33:37.601675+00:00 | succeeded | da39b642da1afd6e1de460cec068bfa4cf1d92197848ffb327a0255903cfa95d | 22,836 | null | """Round 211 common TRM-style x/y/z shared-transition template."""
from __future__ import annotations
import math
import torch
import torch.nn.functional as F
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
TokenLossBatch,
assert_model_state,
)
from torch import... | {
"id": "13f48276-8763-40e6-8e05-308e523453fc",
"created_at": "2026-08-09 02:33:37.601675+00:00",
"db_md5": "d191e6ac3e82ab0f77439af3060169fc",
"submitter": "Yash Kant",
"github_login": "yashkant",
"run_id": "c41fc849-1901-4830-830b-ef0202ec478e",
"tier": "easy",
"dataset_id": "e1",
"status": "succeed... | {
"score": {
"mean_loss": 4.42952299118042,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.046666666865348816
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 4.741913795471191,
"example_count": 100,
... |
13f54401-2754-4ed9-9392-6d3c49507237 | easy | khushidahi | 2026-08-01 23:45:55.227580+00:00 | succeeded | 41c54cc52a7d0f9507ffe227ca6f5a3e95b795ef708f65e6fc9a20f5c8ad0168 | 6,940 | null | """Prelude -> recalled shared core -> coda model for One Layer Deeper.
R2 architecture reset:
- use a real one-block prelude and one-block coda;
- apply one shared recurrent Transformer core four times;
- inject immutable prelude context at every recurrent update;
- use sandwich RMS normalization for stable repeated a... | {
"id": "13f54401-2754-4ed9-9392-6d3c49507237",
"created_at": "2026-08-01 23:45:55.227580+00:00",
"db_md5": "3d9a77124b3463e9101b55a7e0aacc9a",
"submitter": "khushidahi",
"github_login": "khushidahi",
"run_id": "a99d0c94-387a-4d97-8d0a-d054ddf1ced2",
"tier": "easy",
"dataset_id": "e5",
"status": "succ... | {
"score": {
"mean_loss": 2.2019288539886475,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.006666666828095913
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.20788836479187,
"example_count": 600,... |
13fdb019-2506-41c3-aba2-16d48ff36a7c | easy | khushidahi | 2026-08-06 00:12:33.680956+00:00 | succeeded | 5bf864f34bf3f92c66e40dc8633aafd14d99a7497d43edbe2f24ad4877d3d7d7 | 20,869 | null | """R21 hierarchical value-and-scratch model for One Layer Deeper.
The model does not implement multiplication, modular reduction, or the public
recurrence. It uses the public tokenizer structure to build learned, right-
aligned decimal tapes for N, X, and T, then applies a shared neural transition
to a mutable work ta... | {
"id": "13fdb019-2506-41c3-aba2-16d48ff36a7c",
"created_at": "2026-08-06 00:12:33.680956+00:00",
"db_md5": "f770df6f815f55cb48c53b021d3f4961",
"submitter": "khushidahi",
"github_login": "khushidahi",
"run_id": "23686ab6-7ad6-435f-9df9-4b91e58b070a",
"tier": "easy",
"dataset_id": "e2",
"status": "succ... | {
"score": {
"mean_loss": 2.2327709197998047,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.013958333991467953
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.2321996688842773,
"example_count": 30... |
140c3419-b264-49eb-8958-374137e8b137 | easy | neel04 | 2026-08-11 11:44:50.140191+00:00 | succeeded | 7677b37847b3013b38510409e4d143bc1233043f841ea62987bab9eea09dc6db | 8,289 | null | """Segment-size-one online training with normalized attractor dynamics."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
BatchReuseContext,
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
TokenLossBa... | {
"id": "140c3419-b264-49eb-8958-374137e8b137",
"created_at": "2026-08-11 11:44:50.140191+00:00",
"db_md5": "5e0ae6058663bd268846a39914fdb950",
"submitter": "Neel Gupta",
"github_login": "neel04",
"run_id": "346caaf3-3305-4e03-ad60-56e44bf532b0",
"tier": "easy",
"dataset_id": "e1",
"status": "succeede... | {
"score": {
"mean_loss": 1.9501488208770752,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.01666666753590107
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 1.7150871753692627,
"example_count": 100... |
140d3582-b7f0-422a-a24c-efd6c5d2e163 | easy | saketirl | 2026-08-04 03:51:37.563437+00:00 | succeeded | 328477cc33c93c405430a4da6c62b34cd06e7f2e122051f7b31c6cc10af4bf46 | 6,936 | null | """S4D (diagonal structured state space) model trained with Adam.
Stack of bidirectional S4D blocks (S4D-Lin initialization, Vandermonde
kernel computed in closed form, applied via FFT convolution) with gated
MLPs between them. SSM kernel parameters (dt, A, C) follow the S4 recipe:
capped learning rate and no weight d... | {
"id": "140d3582-b7f0-422a-a24c-efd6c5d2e163",
"created_at": "2026-08-04 03:51:37.563437+00:00",
"db_md5": "862a5a71e460e8c55cba0012a9b64cf8",
"submitter": "saketirl",
"github_login": "saketirl",
"run_id": "47237c56-bced-4d34-a650-339478dc56ea",
"tier": "easy",
"dataset_id": "e1",
"status": "succeede... | {
"score": {
"mean_loss": 3.4214309453964233,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.0833333320915699
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.482327699661255,
"example_count": 100,
... |
14129844-791a-4dbc-a8f5-f263cfc41c6a | easy | khushidahi | 2026-08-07 22:56:49.281124+00:00 | succeeded | 76fd6951f818ebbec073e34e757cce06a252768085b10f092e4e34dae68ddfc0 | 8,514 | null | """H0 conservative R3 Hard anchor.
This is the existing R3 architecture and AdamW recipe evaluated after exactly
3,000 optimizer steps. It provides the matched control for the boundary-aware
sequence-loss experiment.
"""
from __future__ import annotations
import random
import torch
import torch.nn.functional as F
f... | {
"id": "14129844-791a-4dbc-a8f5-f263cfc41c6a",
"created_at": "2026-08-07 22:56:49.281124+00:00",
"db_md5": "6f95b0b4617103842f3f3d1674be973d",
"submitter": "khushidahi",
"github_login": "khushidahi",
"run_id": "1261d1f8-0620-4bab-bede-150738a0c1e0",
"tier": "easy",
"dataset_id": "e4",
"status": "succ... | {
"score": {
"mean_loss": 2.083592414855957,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.007083333330228925
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.0998880863189697,
"example_count": 120... |
1414bb32-4d01-4530-b26e-eca22e8c4d2e | easy | YuMingQian1234 | 2026-08-07 09:36:56.322027+00:00 | failed | 4fad565c482eefdb50a56cfde4963ed6432efc914a1e97ca0676640221ec6ab7 | 7,194 | null | """Simplified LoopFormer for one-layer-deeper submission."""
from __future__ import annotations
import math
from dataclasses import dataclass
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor
from benchmark import ModelSpec, OptimizerBundle, OptimizerSpec, Submission, asser... | {
"id": "1414bb32-4d01-4530-b26e-eca22e8c4d2e",
"created_at": "2026-08-07 09:36:56.322027+00:00",
"db_md5": "afd13effdb331a70b5b6eb729d394278",
"submitter": "Mingqian Yu",
"github_login": "YuMingQian1234",
"run_id": "156968e9-3dc8-47b0-bc9e-94b62076e191",
"tier": "easy",
"dataset_id": "e5",
"status": ... | null
|
14259c74-dd92-468c-80d2-961cecf17745 | easy | gauravmishra | 2026-08-10 19:05:28.286517+00:00 | succeeded | 014d6d3661b14388011da5050cad84c15ccb81125f3b10e1924873c106bbf5bb | 14,523 | null | """Capability reset: end-relative output queries with optional recurrent tape.
The only line changed between hosted Easy variants is ``VARIANT``. Every
variant is generic: it uses only input embeddings, positions, and the supplied
attention mask. It does not inspect token IDs, parse fields, or implement an
arithmeti... | {
"id": "14259c74-dd92-468c-80d2-961cecf17745",
"created_at": "2026-08-10 19:05:28.286517+00:00",
"db_md5": "6eef63956abf2473ec2a1f1cefc2c5ac",
"submitter": "Gaurav Mishra",
"github_login": "gauravmishra",
"run_id": "0b8d4693-b7bb-4b6e-9959-bc50489581cd",
"tier": "easy",
"dataset_id": "e1",
"status": ... | {
"score": {
"mean_loss": 1.9237627387046814,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.06833333149552345
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 1.847787857055664,
"example_count": 100,... |
14270dc9-e5d2-4ac9-97db-0dbeed01d92c | easy | erdavis0 | 2026-08-29 01:17:52.910858+00:00 | succeeded | c30171170d2368108bbd1f8de018dfbd1d387e0e28dec2d543ae91ea35b7ad59 | 20,136 | null | """Curvature table with a GPU-resident cached modulus-relative residual."""
from __future__ import annotations
import math
import time
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
TokenLos... | {
"id": "14270dc9-e5d2-4ac9-97db-0dbeed01d92c",
"created_at": "2026-08-29 01:17:52.910858+00:00",
"db_md5": "efb4da441ead29fd5826581673cba669",
"submitter": "Ethan",
"github_login": "erdavis0",
"run_id": "1e2ca73e-2cf5-458a-9414-fe274f9fd0e2",
"tier": "easy",
"dataset_id": "e3",
"status": "succeeded",... | {
"score": {
"mean_loss": 2.859876850863035,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.021875000013969838
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.9351959433132326,
"example_count": 800... |
142d2aad-c2f2-4605-ac90-d0275744d6ca | easy | sapient-sapiens | 2026-08-26 10:54:18.499352+00:00 | failed | 0bdbde8aa9be5239bf45b7d57e2d82ce6083d573927359367c4f7b15c1583436 | 45,618 | null | """Matched hierarchy, operator, N-routing, and optimizer research scaffold.
This is a legitimate learned model. It embeds categorical prompt tokens and
structural field positions; it never parses N, X, or T as numeric values. All
recurrence counts are fixed by ``CampaignConfig`` (or randomized independently
of the i... | {
"id": "142d2aad-c2f2-4605-ac90-d0275744d6ca",
"created_at": "2026-08-26 10:54:18.499352+00:00",
"db_md5": "f4c786991d8dc4849e5966eb882f7d4b",
"submitter": "Ertondy",
"github_login": "sapient-sapiens",
"run_id": "6222e669-07ae-4f82-a663-0881551933eb",
"tier": "easy",
"dataset_id": "e5",
"status": "fa... | null
|
14327886-bdcd-4280-a63e-3cea6295eb98 | easy | DDanlov | 2026-08-11 19:31:15.923405+00:00 | failed | 12ec03f89baf98bd28ddb40b43dcf4b4527ebaceea6517f483abb92960eac0f6 | 18,994 | null | from __future__ import annotations
import sys
import os
import math
import time
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim.optimizer import Optimizer
VOCAB_SIZE = 17
DIGIT_OFFSET = 7
class RohanShampoo(Optimizer):
"""Self-contained RohanShampoo optimizer with bias-correcte... | {
"id": "14327886-bdcd-4280-a63e-3cea6295eb98",
"created_at": "2026-08-11 19:31:15.923405+00:00",
"db_md5": "133698370ebc70c15a7f51bc9b23c992",
"submitter": "DDanlov",
"github_login": "DDanlov",
"run_id": "74f7313e-cce2-4810-aadb-ece81ae00ea9",
"tier": "easy",
"dataset_id": "e1",
"status": "failed",
... | null
|
14334cf6-5da6-44f0-8422-6d0f20387d79 | easy | DDanlov | 2026-08-26 19:31:27.220404+00:00 | succeeded | c5284981926d3da0269416e78d9fa2616dc0ca14306e547ae37faa06706b96ce | 31,774 | null | import math
import time
from dataclasses import dataclass
from typing import Optional, Tuple, Dict, Any, List, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim import Optimizer
try:
from onelayerdeeper.api import Submission, OptimizerBundle, ModelSpec, OptimizerSpec
except ... | {
"id": "14334cf6-5da6-44f0-8422-6d0f20387d79",
"created_at": "2026-08-26 19:31:27.220404+00:00",
"db_md5": "a77b79cfbeb9b3b72894c03c48cf3dfb",
"submitter": "DDanlov",
"github_login": "DDanlov",
"run_id": "decbec64-84e0-496b-912c-3a5057e500d1",
"tier": "easy",
"dataset_id": "e5",
"status": "succeeded"... | {
"score": {
"mean_loss": 2.4265682007684752,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.004166666691501935
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.4435885070150745,
"example_count": 60... |
1437f040-9208-4e5d-aba0-84cc677259cd | easy | ehonig | 2026-08-20 03:59:17.688774+00:00 | succeeded | cb0fb17d159da4c5f0448c0d63eff479e376838537c61912e16ea15794489b4b | 21,896 | null | """A plain Transformer for repeated modular squaring, and nothing else yet.
This is a deliberate restart. The previous submission accumulated a field
parser, a carry scan, a learned reciprocal, periodic features, a digit
bottleneck, a depth selector and two cell types, most of them tuned around a
memorisation table th... | {
"id": "1437f040-9208-4e5d-aba0-84cc677259cd",
"created_at": "2026-08-20 03:59:17.688774+00:00",
"db_md5": "22c3366c330f09351548094dabf8261d",
"submitter": "Edouardo Honig",
"github_login": "ehonig",
"run_id": "4d9a70f2-d6ca-4844-9cbf-ccaa398d15a3",
"tier": "easy",
"dataset_id": "e1",
"status": "succ... | {
"score": {
"mean_loss": 1.9168425798416138,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.06499999761581421
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 1.9152061939239502,
"example_count": 100... |
143ae158-30d5-4203-a1eb-4027c5a2ff96 | easy | yashkant | 2026-08-14 15:37:07.481294+00:00 | succeeded | 4e0f5354b5d70d43ec0d5519a064f476778d1179feddc04debfc2d63bc13e331 | 48,492 | null | """Round1868 learned shuffle-exchange circuit on the SAM carrier.
The exact Round1829 SAM/Compositional-Muon Dykstra model stays intact. Eight
small rails repeatedly call one shared learned pair transition four times per
parsed-T step, with the immutable carrier anchor reinjected into every call.
The candidate cycles... | {
"id": "143ae158-30d5-4203-a1eb-4027c5a2ff96",
"created_at": "2026-08-14 15:37:07.481294+00:00",
"db_md5": "262e66011d7c898f095571a0c145c90f",
"submitter": "Yash Kant",
"github_login": "yashkant",
"run_id": "0596cfac-3bf9-4d1a-ac4c-0bf7d63be190",
"tier": "easy",
"dataset_id": "e2",
"status": "succeed... | {
"score": {
"mean_loss": 2.220536470413208,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.015625000931322575
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.237366199493408,
"example_count": 300,... |
14407bb9-d06e-4d44-915f-1c342755d800 | easy | ehonig | 2026-08-18 13:02:45.611057+00:00 | succeeded | eedc11bb7437c9f834e0c512cd2e08b0c79c24802c0195db6ba28d2414eac966 | 13,504 | null | """A plain Transformer for repeated modular squaring, and nothing else yet.
This is a deliberate restart. The previous submission accumulated a field
parser, a carry scan, a learned reciprocal, periodic features, a digit
bottleneck, a depth selector and two cell types, most of them tuned around a
memorisation table th... | {
"id": "14407bb9-d06e-4d44-915f-1c342755d800",
"created_at": "2026-08-18 13:02:45.611057+00:00",
"db_md5": "3319ebf686b078036467bd6a65cf6971",
"submitter": "Edouardo Honig",
"github_login": "ehonig",
"run_id": "eec7de2d-b847-4184-9307-6fa1ce4d5e22",
"tier": "easy",
"dataset_id": "e5",
"status": "succ... | {
"score": {
"mean_loss": 2.1549670953337134,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.0050000000807146235
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.1655468940734863,
"example_count": 6... |
1443ee9c-2103-4f82-808e-9076511476e9 | easy | liam-gb | 2026-08-13 01:51:26.970303+00:00 | succeeded | a922a24f6d897ff79c7df36202bf5adaf215adc204bf4e9750ac5097b36dc04f | 17,291 | null | """rns_pm_gather_ct_widet_rank: prime-level state with a CRT rank channel.
Base:
The wide readout is what makes long answers reachable; the trim is what pays
for it, by not running loop iterations whose update is multiplied by zero.
Base docstring:
Single change vs rns_pm_gather_ct: the readout carries one
distance... | {
"id": "1443ee9c-2103-4f82-808e-9076511476e9",
"created_at": "2026-08-13 01:51:26.970303+00:00",
"db_md5": "38bc09c3be37222b2dcdfacc053d0c51",
"submitter": "liam-gb",
"github_login": "liam-gb",
"run_id": "7e0ce758-1ada-492f-907f-9c16a843836b",
"tier": "easy",
"dataset_id": "e3",
"status": "succeeded"... | {
"score": {
"mean_loss": 3.5295393466949463,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.02937499899417162
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.9673264026641846,
"example_count": 800... |
1447e43e-b4ab-42ff-9a2d-839d15b7cd61 | easy | YangWang92 | 2026-08-17 05:34:38.418421+00:00 | succeeded | 44fc930d809c31a6480ef79b59b1620f4f36943377541193bab7b764615787de | 46,820 | null | """Limb (bignum) circuit submission.
The h1 metrics pinned the hard tier at ~60-64 bit moduli (the training-loss
floor of 100-250 is exactly the fp64 quantisation error of 18-19 digit
integers), so every value here is a little-endian limb vector, base 10^4.
No intermediate ever exceeds ~1.4e9, so a single float64 hold... | {
"id": "1447e43e-b4ab-42ff-9a2d-839d15b7cd61",
"created_at": "2026-08-17 05:34:38.418421+00:00",
"db_md5": "7c68cdea41972b37f3eb5b23a332c615",
"submitter": "Yang Wang ",
"github_login": "YangWang92",
"run_id": "72aa7879-e4f1-45c0-b1d1-b9b8dbede54e",
"tier": "easy",
"dataset_id": "e5",
"status": "succ... | {
"score": {
"mean_loss": 0.0,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 1.0
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 0.0,
"example_count": 600,
"exact_accuracy": 1.0,
"... |
14508e3c-7885-4d8b-b05e-a5b7b20479b2 | easy | CodeReclaimers | 2026-08-15 05:32:14.522895+00:00 | succeeded | acaae3e694638e70e9600d2f2a9476aaee199d3cdd4ad6d8df1a49005e3dd588 | 57,228 | null | """v31: merge of the prime-agent low-rank universal-CRT design (hosted M4:
100%, Max T=64, OOD-N Max T=64) with the v30 streamed large-N readout.
From the low-rank design: per-prime residue embeddings E (sum m, DIM) with a
shared readout w (DIM, WMAX) — per-prime cost m*DIM, so the bank covers all
primes 2..4095 (~17M... | {
"id": "14508e3c-7885-4d8b-b05e-a5b7b20479b2",
"created_at": "2026-08-15 05:32:14.522895+00:00",
"db_md5": "fc08ccee7087a70aeefaaa03c48367d2",
"submitter": "CodeReclaimers",
"github_login": "CodeReclaimers",
"run_id": "8d79b25b-3482-4ff7-be18-33fe2d614c9e",
"tier": "easy",
"dataset_id": "e10",
"statu... | {
"score": {
"mean_loss": 0.0,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.9999999854564667
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 0.0,
"example_count": 125,
"exact_accuracy": 0... |
145464fa-f9ed-44fa-809a-eb353e9d2f1f | easy | AdityaVG13 | 2026-08-15 08:24:02.727667+00:00 | succeeded | cfb992f4a1d5cb6ac4cceae8b4ac9cf57d9609c88f3b7d2570f96d1e4e3ff2be | 110,237 | null | """AEGIS V8 v8_join_lookahead: frontier JOIN with low-overhead slow-weight interpolation.
One Layer Deeper submission candidate. All task transitions remain learned
end-to-end. The deterministic shell only identifies marker-delimited roles,
constructs a finite semantic tape, schedules learned interval programs, and
ap... | {
"id": "145464fa-f9ed-44fa-809a-eb353e9d2f1f",
"created_at": "2026-08-15 08:24:02.727667+00:00",
"db_md5": "9d821f9c0aa1b82a02d3061a26710d5b",
"submitter": "AdityaG",
"github_login": "AdityaVG13",
"run_id": "c8eb0be1-f8ae-4057-bd19-e0440e98905a",
"tier": "easy",
"dataset_id": "e1",
"status": "succeed... | {
"score": {
"mean_loss": 2.576925754547119,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.5116666555404663
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 1.7776596546173096,
"example_count": 100,
... |
145df4f9-f0ce-45ba-acc5-ca2ff3242ea7 | easy | repst | 2026-08-23 19:10:54.283766+00:00 | failed | 2c7861bb2c63cf2d98e214e0ceba7501c6a7945257f530b6fb91cbf177cc268f | 27,550 | null | # R529: streaming-MDL grammar-frontier recurrent machine.
#
# The R525 factor graph and exact recurrent interpreter are retained, including
# their small dense associative memory for direct credit. Expression syntax is
# now streamed in minimum-operation order and stopped at a semantic compute
# budget, rather than ma... | {
"id": "145df4f9-f0ce-45ba-acc5-ca2ff3242ea7",
"created_at": "2026-08-23 19:10:54.283766+00:00",
"db_md5": "c128bfd6cace8a3bb540d49c5a1f15ad",
"submitter": "Ryan Epstein",
"github_login": "repst",
"run_id": "2d201024-ca3b-44f5-84cf-3ce1e298efe5",
"tier": "easy",
"dataset_id": "e1",
"status": "failed"... | null
|
1461e06c-bea6-4173-95bd-94e4d313e89e | easy | yashkant | 2026-08-31 23:24:37.991900+00:00 | succeeded | 699817a2e38633e98908c594e5f7145e716652e8f8c8f6ff9ba2128f3d4fd7e0 | 107,115 | null | """One Layer Deeper submission -- a learned register machine.
Architecture (all outputs produced by learned parameters; no task arithmetic on the answering path):
a bank of R=6 digit registers (W=16 positions each) carried as DISCRETE tokens between micro-steps: five
data registers and one COUNTDOWN register holding t... | {
"id": "1461e06c-bea6-4173-95bd-94e4d313e89e",
"created_at": "2026-08-31 23:24:37.991900+00:00",
"db_md5": "133da8fba65619712c05954617d2f04d",
"submitter": "Yash Kant",
"github_login": "yashkant",
"run_id": "3b2e2c55-7aec-4e56-9e87-4f556c455dc5",
"tier": "easy",
"dataset_id": "e1",
"status": "succeed... | {
"score": {
"mean_loss": 4.0559388399124146,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.006666666828095913
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 4.175271987915039,
"example_count": 100... |
146614a5-1ab1-4865-a482-f08f8f5e3d7f | easy | ArkinDharawat | 2026-08-13 17:16:14.849688+00:00 | succeeded | 5b10be95608ba91c9a2bf1cc6fd3ac2ca82b2e302249625b44a53482740f98f6 | 7,630 | null | """Universal-Transformer-style looped block, v2: step embed + pos offset + LR schedule + 6 loops.
Extends ``looped_ut_padded`` with four ideas drawn from
Graves 2016 (ACT), Dehghani et al. 2018 (Universal Transformer), and
Merrill & Sabharwal 2025 (log-depth transformers):
1. **Learned per-step (timestep) embedding**... | {
"id": "146614a5-1ab1-4865-a482-f08f8f5e3d7f",
"created_at": "2026-08-13 17:16:14.849688+00:00",
"db_md5": "68d62056da13113b443270bf4fff3c62",
"submitter": "Arkin Dharawat",
"github_login": "ArkinDharawat",
"run_id": "d6989177-2f48-4fdc-9bb9-b94160ed5231",
"tier": "easy",
"dataset_id": "e3",
"status"... | {
"score": {
"mean_loss": 2.2682138251131247,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.016875000037252903
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.304916852955864,
"example_count": 800... |
14668c27-b9ee-4ecc-9128-1005113fcfc9 | easy | viridale | 2026-08-03 09:33:08.832307+00:00 | succeeded | 2ab7cba2961297bfd1d3fd25e786d3b53a173f72d2647e0b47600e40a5f5622b | 134,838 | null | """Prompt-depth I/O wrappers over the exact global source posterior.
hypothesis: the hidden source edit may use prompt T in initialization or the
returned observation, loci still absent after the T-conditioned clock and
transition closure despite being explicitly elevated by the API hint.
axis: arch
target: re... | {
"id": "14668c27-b9ee-4ecc-9128-1005113fcfc9",
"created_at": "2026-08-03 09:33:08.832307+00:00",
"db_md5": "317a01daef78471923437720d63f2031",
"submitter": "priormancer",
"github_login": "viridale",
"run_id": "754b0429-a367-46bd-b96e-8d9ce4478d89",
"tier": "easy",
"dataset_id": "e3",
"status": "succe... | {
"score": {
"mean_loss": 0.0030146650951345506,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 1.0
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 0.0030146650919047383,
"example_count": 800,
... |
146f781d-b3fc-4ccb-9b26-0766300f9a38 | easy | karanganesan | 2026-08-04 01:50:32.048805+00:00 | succeeded | a2a6a52febad15170cc09f132592c0b881e3662c6d1acc0808786b9f15e5e47f | 11,993 | null | """Parametric looped-transformer family (P1).
One weight-tied transformer block applied k times in latent space. Config
flags cover four P1 families with one file:
looped recall=0 gated=0 tfilm=0 plain weight-tied loop
looped-recall recall=1 re-inject the input embedding each
... | {
"id": "146f781d-b3fc-4ccb-9b26-0766300f9a38",
"created_at": "2026-08-04 01:50:32.048805+00:00",
"db_md5": "59a0513e316bcd4bf0f0a6b66c911b28",
"submitter": "Karan Ganesan",
"github_login": "karanganesan",
"run_id": "a5d9a2e0-e1a2-4d13-80ad-1cf07e0497f4",
"tier": "easy",
"dataset_id": "e1",
"status": ... | {
"score": {
"mean_loss": 2.2287914752960205,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.061666667461395264
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.0691158771514893,
"example_count": 10... |
147033c9-e1be-4c88-bfd5-6202c8d54703 | easy | jordanrubin | 2026-08-08 10:22:27.328948+00:00 | succeeded | 19e434c7e9fc9d9d601292a0f5c0a4083813bc7790caf87c0582fc1bbcc2cad8 | 16,909 | null | """Autonomous digit-state recurrent Transformer for repeated modular squaring.
The model is deliberately organized around one learned transition:
p_0 = right_aligned_decimal_digits(x)
p_{k+1} = F_theta(p_k, decimal_digits(N))
The same two-block Transformer cell is applied exactly T times. T is used only
as ... | {
"id": "147033c9-e1be-4c88-bfd5-6202c8d54703",
"created_at": "2026-08-08 10:22:27.328948+00:00",
"db_md5": "671c15c2db39c821999bae96a7defa51",
"submitter": "Jordan Rubin",
"github_login": "jordanrubin",
"run_id": "14948586-562e-454c-bf3d-29429463097e",
"tier": "easy",
"dataset_id": "e5",
"status": "s... | {
"score": {
"mean_loss": 4.179352021027447,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.009166666384165486
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 4.1744710662707085,
"example_count": 600... |
14840b73-182d-46d3-aff6-40ef32b01907 | easy | TZZheng | 2026-08-06 21:11:53.754481+00:00 | succeeded | 8ed119744976af18c314e728fd72c5ed9655225dec68493b943a85ac099fa0a6 | 7,300 | null | """Self-contained standard Transformer scaling-sweep submission."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
assert_model_state,
)
VARIANT = {'block... | {
"id": "14840b73-182d-46d3-aff6-40ef32b01907",
"created_at": "2026-08-06 21:11:53.754481+00:00",
"db_md5": "136b65e97351ed5257a1af84a0ed649c",
"submitter": "TZZheng",
"github_login": "TZZheng",
"run_id": "05a2f536-e501-4e29-bcb5-e947b3da7a6d",
"tier": "easy",
"dataset_id": "e1",
"status": "succeeded"... | {
"score": {
"mean_loss": 2.590088188648224,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.06666666455566883
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.272279739379883,
"example_count": 100,
... |
1488051f-261e-4a96-9e47-35d2187acc16 | easy | shirvani-jr | 2026-08-31 01:45:23.187487+00:00 | succeeded | f36b045c50800cbdceef864d677daf420a793727133eaedab87b1505981fd664 | 21,222 | null | from __future__ import annotations
import math
import time
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
TokenLossBatch,
assert_model_state,
)
PERIOD_MINIMUM = 3
PERIOD_MAXIMUM = 100
EX... | {
"id": "1488051f-261e-4a96-9e47-35d2187acc16",
"created_at": "2026-08-31 01:45:23.187487+00:00",
"db_md5": "52d288f37eb9ac99c594e70a5f2357e8",
"submitter": "Ali",
"github_login": "shirvani-jr",
"run_id": "84bd6c24-b89c-49cd-8f3b-e5c2b7308717",
"tier": "easy",
"dataset_id": "e6",
"status": "succeeded"... | {
"score": {
"mean_loss": 1.2563760578632355,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.7038288712501526
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 1.8861925601959229,
"example_count": 60,
... |
148b60c3-b751-4d82-bfb9-629480a4b81d | easy | repst | 2026-08-23 19:03:42.808949+00:00 | failed | 0bf626f616ae5a260342167a04ac963eba78f576e1653c0963768abec86be063 | 90,978 | null | # R547: exposure-aware recurrent program-space machine.
#
# ARCHITECTURE/VOCABULARY (analogous to an LM): generic modular register VM;
# bases {0,1,-1,x,r0,r1}; ADD/SUB/MUL expression grammar and semantic quotient;
# exact recurrent execution; decimal tokenization; MDL/dataflow-generated search
# views; and a balanced ... | {
"id": "148b60c3-b751-4d82-bfb9-629480a4b81d",
"created_at": "2026-08-23 19:03:42.808949+00:00",
"db_md5": "b2f10aaf928c4887737a03aeccdf37e6",
"submitter": "Ryan Epstein",
"github_login": "repst",
"run_id": "b1cfc3c2-ce7b-4f67-ba0f-fcb82ec17346",
"tier": "easy",
"dataset_id": "e1",
"status": "failed"... | null
|
148d06fd-5bc3-4770-9e6d-304a6dcb57d7 | easy | EyimofeA | 2026-08-29 19:34:10.302576+00:00 | succeeded | df552728b3013033ad6fab41fa797afced1f076f74eadce9ab492449467be7e8 | 9,471 | null | """Card 2: T=1 Assassin — one-step specialist with weak multi-T fallback.
Learned pairwise x-digit interactions cross-attend to N; iterative refinement;
T-conditioned router between T=1 specialist and generic head. No explicit
digit-product arithmetic.
Source design: GPT-5 Pro (2026-08-07 five-card suite).
"""
from _... | {
"id": "148d06fd-5bc3-4770-9e6d-304a6dcb57d7",
"created_at": "2026-08-29 19:34:10.302576+00:00",
"db_md5": "da8faead62936dc6cfe9331be671ecdb",
"submitter": "mof",
"github_login": "EyimofeA",
"run_id": "3add3f3c-aea6-4261-90e3-28595ec012cf",
"tier": "easy",
"dataset_id": "e1",
"status": "succeeded",
... | {
"score": {
"mean_loss": 1.8284724354743958,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.021666666492819786
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 1.7611820697784424,
"example_count": 10... |
149060a9-4d14-46d7-8b69-dd775d2e770b | easy | khushidahi | 2026-08-01 23:46:22.433512+00:00 | succeeded | f37d8a71736fca58ee44b5b0cb9a9262ffaf5e535914112844c2173c6a2c45a7 | 6,940 | null | """Prelude -> recalled shared core -> coda model for One Layer Deeper.
R2 architecture reset:
- use a real one-block prelude and one-block coda;
- apply one shared recurrent Transformer core four times;
- inject immutable prelude context at every recurrent update;
- use sandwich RMS normalization for stable repeated a... | {
"id": "149060a9-4d14-46d7-8b69-dd775d2e770b",
"created_at": "2026-08-01 23:46:22.433512+00:00",
"db_md5": "87939af61289f52f848d2257ee604293",
"submitter": "khushidahi",
"github_login": "khushidahi",
"run_id": "1beef5ac-c354-425d-b513-98424135cc32",
"tier": "easy",
"dataset_id": "e5",
"status": "succ... | {
"score": {
"mean_loss": 2.20361864566803,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.006666666828095913
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.217721700668335,
"example_count": 600,
... |
14915d60-7b63-4ff2-a3f0-c99945a27517 | easy | kushasareen | 2026-08-12 02:18:56.806874+00:00 | succeeded | e94d94f9b8a10681922a18d7d8de0ba0095e71c814067fe99fd1a1bd40e75e5f | 6,914 | null | """v2 capacity descent: d_model 32, lr 1.6e-2 (1/width scaling).
T-conditioned recurrent Transformer with the measured stability fixes:
- 2 untied pre-blocks (pre-norm), 1 untied post-block, tied loop block
- prompt re-injection every loop iteration (gradient + information path)
- UT-style double post-norm inside the ... | {
"id": "14915d60-7b63-4ff2-a3f0-c99945a27517",
"created_at": "2026-08-12 02:18:56.806874+00:00",
"db_md5": "0261554935df0d7bbec00d47291cb76d",
"submitter": "Kusha Sareen",
"github_login": "kushasareen",
"run_id": "77c426b2-0de7-4068-82ac-39e023e8bbc7",
"tier": "easy",
"dataset_id": "e5",
"status": "s... | {
"score": {
"mean_loss": 3.2420193981697576,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.005416666666666667
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.3886644177372682,
"example_count": 60... |
1496e7eb-ee92-4fde-b25a-71a7f66439fa | easy | karanganesan | 2026-08-05 11:40:34.291388+00:00 | succeeded | 3fa4bad9f7d931cc6838498350afa0475b98b2b74760fd0bd1d13915b10472e8 | 22,081 | null | """Parametric looped-transformer family (P1).
One weight-tied transformer block applied k times in latent space. Config
flags cover four P1 families with one file:
looped recall=0 gated=0 tfilm=0 plain weight-tied loop
looped-recall recall=1 re-inject the input embedding each
... | {
"id": "1496e7eb-ee92-4fde-b25a-71a7f66439fa",
"created_at": "2026-08-05 11:40:34.291388+00:00",
"db_md5": "ab114f8387eea2da80ade4a7ece2fb4c",
"submitter": "Karan Ganesan",
"github_login": "karanganesan",
"run_id": "2633cc74-6726-431c-b042-d1d7366c306c",
"tier": "easy",
"dataset_id": "e1",
"status": ... | {
"score": {
"mean_loss": 3.247635006904602,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.014999999664723873
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 4.204451560974121,
"example_count": 100,... |
149e8323-cf0f-4b11-99b5-aaaa9aa01e8a | easy | jordanrubin | 2026-08-02 18:26:07.932945+00:00 | succeeded | 16a52d22982dfef736863488be48d497f494793eb9aba7fdf8b93962885bc8fc | 13,994 | null | """Bidirectional transformer, D=256 x 3 blocks, grokking-tuned AdamW.
Design notes:
- Inputs are padded to config.max_seq_len inside forward and logits sliced
back, so every training batch presents one static shape (compile/cudagraph
friendly, and uniform kernel shapes even in eager).
- Attention mask gets an iden... | {
"id": "149e8323-cf0f-4b11-99b5-aaaa9aa01e8a",
"created_at": "2026-08-02 18:26:07.932945+00:00",
"db_md5": "1c62e1fd584dfb5f0181cb2641199911",
"submitter": "Jordan Rubin",
"github_login": "jordanrubin",
"run_id": "7bea065d-8be4-43f3-b4db-42d2f540b5f9",
"tier": "easy",
"dataset_id": "e1",
"status": "s... | {
"score": {
"mean_loss": 5.161844237005502,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.22666666004806757
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 7.919300227022883,
"example_count": 100,
... |
149fb118-23ee-4696-a0f6-1b7c1dbc39c7 | easy | yashkant | 2026-08-13 00:00:47.397110+00:00 | succeeded | 462ef777aa8bbb91c05eed81cbbbfa97179749c7b94c4b7182ddf587e9a50974 | 40,955 | null | """Round1786 evolutionary crossover of two controlled survivors.
The Round1780 W8/D2 carrier keeps its winning contrastive, exactly
antisymmetric learned set codes. The candidate additionally removes each
initial and recurrent Dykstra correction residual's coordinate mean before
the unchanged tanh bound, combining Ro... | {
"id": "149fb118-23ee-4696-a0f6-1b7c1dbc39c7",
"created_at": "2026-08-13 00:00:47.397110+00:00",
"db_md5": "d690bb5204860cdff9357dd5287ebd6b",
"submitter": "Yash Kant",
"github_login": "yashkant",
"run_id": "2eebaa77-4639-4fba-a024-f50ee7056507",
"tier": "easy",
"dataset_id": "e3",
"status": "succeed... | {
"score": {
"mean_loss": 2.128150036339842,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.017500000121071932
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.129232571312596,
"example_count": 800,... |
149fe2c4-b685-448f-8ce4-8a84661ecdc9 | easy | isaac0804 | 2026-08-07 16:03:12.752844+00:00 | succeeded | 53a5aacca12b9146c04eacdbb2ef159d59b3428a65aa78b86bd1060d52753a4d | 9,428 | null | """Clean variable-depth loop (no damped gate, no step-embedding) optimized
with Muon. The gate and step-emb are individually proven (local, this
session) to block single-squaring learning; the bare loop fits the primitive
at plain-block quality. The earlier Muon test (looped_transformer_muon) ran
on the DAMPED architec... | {
"id": "149fe2c4-b685-448f-8ce4-8a84661ecdc9",
"created_at": "2026-08-07 16:03:12.752844+00:00",
"db_md5": "4c39625379ef4c0c168ecc34bd6e1175",
"submitter": "Isaac Yong",
"github_login": "isaac0804",
"run_id": "5a2abbf5-dad7-446e-a211-3daf7f16314c",
"tier": "easy",
"dataset_id": "e3",
"status": "succe... | {
"score": {
"mean_loss": 8.854278316760071,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.005625000009313226
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 9.014243718501545,
"example_count": 800,... |
14a73324-e781-4af7-9514-d163e2102d0b | easy | khushidahi | 2026-08-15 13:24:30.520418+00:00 | succeeded | 532376dab19964727c45bb2ecbfe96a228979f725217ef5ccab9dc1f88ffb335 | 25,323 | null | """R6 structured decimal-workspace model for One Layer Deeper.
The model does not implement multiplication, modular reduction, or the public
recurrence. It uses the public tokenizer structure to build learned, right-
aligned decimal tapes for N, X, and T, then applies a shared neural transition
to a mutable work tape.... | {
"id": "14a73324-e781-4af7-9514-d163e2102d0b",
"created_at": "2026-08-15 13:24:30.520418+00:00",
"db_md5": "7526afc0b23fabeca3118e95189b2688",
"submitter": "khushidahi",
"github_login": "khushidahi",
"run_id": "2210ef41-b44d-40a5-b78b-dd0f80eaa1be",
"tier": "easy",
"dataset_id": "e5",
"status": "succ... | {
"score": {
"mean_loss": 2.1457722187042236,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.01000000024214387
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.1476519107818604,
"example_count": 600... |
14afe792-60c1-4ce6-b8fd-ad0f090b4ead | easy | oupadhyay | 2026-08-15 22:39:31.334965+00:00 | succeeded | 6aabe7f902eafd72bca8518a0bdf10a6b3c7ebf090db2afed5eb920fed202c16 | 4,273 | null | """Conditional structured-loss screen: focal20."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import nn
from benchmark import ModelSpec,OptimizerSpec,OptimizerBundle,Submission,TokenLossBatch,assert_model_state
W,D,H=4,64,32
class C:
def __init__(s,vocab_size,max... | {
"id": "14afe792-60c1-4ce6-b8fd-ad0f090b4ead",
"created_at": "2026-08-15 22:39:31.334965+00:00",
"db_md5": "255b8b3d7089ab4181b16c174feeaf6a",
"submitter": "Ojasw Upadhyay",
"github_login": "oupadhyay",
"run_id": "5bc4e529-147e-4164-b15c-8d61a2d84aa3",
"tier": "easy",
"dataset_id": "e5",
"status": "s... | {
"score": {
"mean_loss": 2.427054116622427,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.006666666679084301
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.3475711185301367,
"example_count": 600... |
14b5f969-5b0c-4abe-b294-2bd73ed814bd | easy | chuk-uzowihe | 2026-08-12 18:59:06.684827+00:00 | succeeded | 0cfd94bb13ad2a0f5776a30ae5df2681493fb7e0b6c155c8e0f22867b221094c | 26,028 | null | """v10: dense fixed-loop recurrence with deep supervision.
Lineage: FPRM structure (alpha/beta-tied residual scaling, pre-rms_norm,
input injection every iteration) + the v8 2D digit lattice / RoPE task
encoding. Three deliberate departures from the v9 MoEUT line:
1. **Full BPTT.** The loop runs a fixed N_ITER iterat... | {
"id": "14b5f969-5b0c-4abe-b294-2bd73ed814bd",
"created_at": "2026-08-12 18:59:06.684827+00:00",
"db_md5": "d74d6555582e0ca3ed49086d931cd84f",
"submitter": "Chuk Uzowihe",
"github_login": "chuk-uzowihe",
"run_id": "1f007e21-37ad-4339-a5c4-9f8df1e1a93a",
"tier": "easy",
"dataset_id": "e1",
"status": "... | {
"score": {
"mean_loss": 26.526833534240723,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.07999999821186066
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 22.726125717163086,
"example_count": 100... |
14c30ee4-2a4b-4a1d-8cb1-48b850a5ec03 | easy | pculaf | 2026-08-15 22:36:22.745375+00:00 | succeeded | 0e55d3eb6c6b39cabb13b2f1213bc31ee5d8680d140f8edd872bbabcc07e887e | 10,442 | null | """Muon-optimized recurrent Transformer for the E6 activation/depth sweep."""
from __future__ import annotations
from collections.abc import Iterable
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submissio... | {
"id": "14c30ee4-2a4b-4a1d-8cb1-48b850a5ec03",
"created_at": "2026-08-15 22:36:22.745375+00:00",
"db_md5": "6b3ed990f9ed8b7d974ae9aa53b71007",
"submitter": "Pavle Culafic",
"github_login": "pculaf",
"run_id": "fb2725b7-4878-4e3e-9d0b-e3c8b7d312e8",
"tier": "easy",
"dataset_id": "e6",
"status": "succe... | {
"score": {
"mean_loss": 2.0458202362060547,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.011711712228134274
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.0590827465057373,
"example_count": 60... |
14c3dc91-3452-4f5a-a97f-9200f0d86eaf | easy | richardcepka | 2026-08-17 20:15:06.432224+00:00 | succeeded | 6b2ef54701284dd2a170d566b9e09b5270006d4581bdd08333468fe0ff1d3c82 | 44,425 | null | """Recurrent digit-register model for One Layer Deeper.
Given a modulus N, a starting value x, and a step count T, predict the residue after squaring modulo N exactly T times.
repo/README.md
A looped transformer whose recurrent state is a decimal digit register. The loop is nested: an
outer step is one squaring, an... | {
"id": "14c3dc91-3452-4f5a-a97f-9200f0d86eaf",
"created_at": "2026-08-17 20:15:06.432224+00:00",
"db_md5": "edf5092a761357998957395ef3002719",
"submitter": "Richard Cepka",
"github_login": "richardcepka",
"run_id": "cb93d4c1-d1b8-4aa3-99d2-84e139ae22b2",
"tier": "easy",
"dataset_id": "e5",
"status": ... | {
"score": {
"mean_loss": 2.1900546550750732,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.007500000298023224
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.190497636795044,
"example_count": 600... |
14c93f32-f009-4d15-874d-66fe16529900 | easy | oupadhyay | 2026-08-07 10:39:43.763314+00:00 | succeeded | ba1cd921e940c748344be95afb9bef409a728389c0c460a8d4bd9022381844ce | 9,400 | null | """Baseline Transformer plus four diagnostic axial-cell calls."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
assert_model_state,
)
PAD, N, X, T, ANS, ... | {
"id": "14c93f32-f009-4d15-874d-66fe16529900",
"created_at": "2026-08-07 10:39:43.763314+00:00",
"db_md5": "d190908fbb432ad5046ed45cd33f2083",
"submitter": "Ojasw Upadhyay",
"github_login": "oupadhyay",
"run_id": "cf8c97c8-73c4-448b-83b4-ce64cfd4d802",
"tier": "easy",
"dataset_id": "e5",
"status": "s... | {
"score": {
"mean_loss": 74.31122951336025,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.0004166666666666667
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 75.17941961160155,
"example_count": 600... |
14cc1fe1-59a7-47bf-9f9b-d747a2393fcf | easy | pepinu | 2026-08-26 15:15:04.995310+00:00 | failed | e1bb48bacd4d7f87cbf4a88103bc16ab551f9b9fcaad4a15bbb6862cf5eb9f93 | 45,960 | null | """Squaring mod N as a learned permutation on CRT components.
WHY THIS SHAPE. (Z_N, x) is a commutative monoid, so every finite-dimensional
linear representation splits into multiplicative characters and squaring is
diagonal there -- the component basis is the only one in which this operation is
shallow. Two consequen... | {
"id": "14cc1fe1-59a7-47bf-9f9b-d747a2393fcf",
"created_at": "2026-08-26 15:15:04.995310+00:00",
"db_md5": "a526b54eb57b21c5a06feb52f6352071",
"submitter": "pepinu",
"github_login": "pepinu",
"run_id": "4e5a8db1-0a28-4b15-91ea-31b4923ad849",
"tier": "easy",
"dataset_id": "e1",
"status": "failed",
"... | null
|
14d4166f-11b0-4b92-a3bd-8895d697df8e | easy | yashkant | 2026-08-23 21:45:08.632899+00:00 | failed | f5a6ccbe53a5d783c16f9ace2e9e00eb600648aca1a876338ba2de3170c842e5 | 11,557 | null | """One Layer Deeper submission: CRT residue-register machine with a learned squaring
transition and a Fourier "clock" bilinear reconstruction readout.
Idea (end-to-end learned; nothing about the task is hard-coded):
* State = per-modulus residue registers over a fixed bank of small primes. The input
integer x is... | {
"id": "14d4166f-11b0-4b92-a3bd-8895d697df8e",
"created_at": "2026-08-23 21:45:08.632899+00:00",
"db_md5": "3b4fcc872bbd87ef7c8d539383458d95",
"submitter": "Yash Kant",
"github_login": "yashkant",
"run_id": "037ace99-e132-4702-a586-66d97a71b5ce",
"tier": "easy",
"dataset_id": "e1",
"status": "failed"... | null
|
14dd93f8-5d4d-408d-81b8-f2f1ee77373f | easy | jordanrubin | 2026-08-02 17:49:50.522661+00:00 | succeeded | ed50e303345b0710236c6ad682f1592f44b2b7ac84c4b93577f612e4da2e5879 | 13,994 | null | """Bidirectional transformer, D=256 x 3 blocks, grokking-tuned AdamW.
Design notes:
- Inputs are padded to config.max_seq_len inside forward and logits sliced
back, so every training batch presents one static shape (compile/cudagraph
friendly, and uniform kernel shapes even in eager).
- Attention mask gets an iden... | {
"id": "14dd93f8-5d4d-408d-81b8-f2f1ee77373f",
"created_at": "2026-08-02 17:49:50.522661+00:00",
"db_md5": "cf4268bd0d1c13f90d014c8a284e9b94",
"submitter": "Jordan Rubin",
"github_login": "jordanrubin",
"run_id": "a1345a1e-235c-43e5-b86a-1855676d442e",
"tier": "easy",
"dataset_id": "e1",
"status": "s... | {
"score": {
"mean_loss": 3.016563943199386,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.019999998932083447
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.4067912101745605,
"example_count": 100... |
14ef3006-6c2a-4317-8659-3502f21a0351 | easy | khushidahi | 2026-08-17 22:30:24.605894+00:00 | succeeded | 7b478978d5350f86d61fa31ef4be9f80a1130763dca5badab36008cee0163283 | 36,459 | null | """R234: Abacus-aligned relational cross-attention reducer.
This diagnostic preserves R216's learned low-to-high product/carry front end
and exact outer recurrence. It changes only the high-to-low reducer. At every
product event, every evolving work place interacts with every immutable
categorical modulus place thro... | {
"id": "14ef3006-6c2a-4317-8659-3502f21a0351",
"created_at": "2026-08-17 22:30:24.605894+00:00",
"db_md5": "a88cb3eded54c0348fb9d8d88920ac33",
"submitter": "khushidahi",
"github_login": "khushidahi",
"run_id": "b21000d5-5495-44c8-af4f-a6cdd859fea2",
"tier": "easy",
"dataset_id": "e6",
"status": "succ... | {
"score": {
"mean_loss": 0.9271394312381744,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.8013513684272766
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 0.5376716256141663,
"example_count": 60,
... |
14f6a66f-359c-45db-8c76-77c15e2ace3a | easy | Yalyenea | 2026-08-12 05:47:50.780843+00:00 | succeeded | 19e285547e4b51779826b8853039604b1cca0f0be9cdde0e2a4e81469a889dc5 | 11,690 | null | """Adaptive learned bit-state recurrence with a lightweight causal decoder."""
from __future__ import annotations
import math
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
assert_model_stat... | {
"id": "14f6a66f-359c-45db-8c76-77c15e2ace3a",
"created_at": "2026-08-12 05:47:50.780843+00:00",
"db_md5": "377d2bf02c3b75d001a154f3772c2ccd",
"submitter": "yfff",
"github_login": "Yalyenea",
"run_id": "6f4fa7f3-4411-46fd-979d-ad67a8c1aa60",
"tier": "easy",
"dataset_id": "e5",
"status": "succeeded",
... | {
"score": {
"mean_loss": 6.549317665330999,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.013333333935588599
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 6.60391092300415,
"example_count": 600,
... |
14f9ef2d-ba1d-4919-9f33-4b915b8cef3f | easy | AFeinfield | 2026-08-28 04:12:24.967672+00:00 | succeeded | ebc2e492aa5c558230e4172d0a761b926fb7db1491c70dacf6698415f71a5dd0 | 9,266 | null | """189: DIGIT-TUPLE GROKMAP — the h1-scale fusion.
F maps per-place digit distributions to per-place digit logits through
one dense quadratic trunk, conditioned on N via per-place digit
embeddings. The output IS the composable state (per-place CE lands on
it directly — no head to shortcut through, the 181 lesson), com... | {
"id": "14f9ef2d-ba1d-4919-9f33-4b915b8cef3f",
"created_at": "2026-08-28 04:12:24.967672+00:00",
"db_md5": "e6c0ba375db62da06f001e62befe1ea9",
"submitter": "AFeinfield",
"github_login": "AFeinfield",
"run_id": "5c1226e3-0721-4306-981a-590ed6ddc0a3",
"tier": "easy",
"dataset_id": "e1",
"status": "succ... | {
"score": {
"mean_loss": 2.378935396671295,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.6150000095367432
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 1.4376839399337769,
"example_count": 100,
... |
14fe68cf-ed32-4bd7-bae0-907602b5b2d4 | easy | newjordan | 2026-08-15 23:29:07.911585+00:00 | succeeded | 6da6e937508b55c0269fa90d648775dd5beecc4c8d7f4934fa089f933655e8a8 | 178,654 | null | """Helix-Crawler — a recursive-weight-shared "crawler" ALU for modular squaring.
The task is r_{i+1} = r_i^2 mod n, chained T times: a *recurrence* over a single
operator. This build embodies the crawler / Frugendorff lineage (spark:~/sota_crawler)
mapped onto that structure. n never enters as a table row (unseen modu... | {
"id": "14fe68cf-ed32-4bd7-bae0-907602b5b2d4",
"created_at": "2026-08-15 23:29:07.911585+00:00",
"db_md5": "e6c32df21410d4316bc488cf172a8410",
"submitter": "Frosty40",
"github_login": "newjordan",
"run_id": "44b4bfbf-13c6-440f-9d5e-6e94ba9d9d7f",
"tier": "easy",
"dataset_id": "e2",
"status": "succeed... | {
"score": {
"mean_loss": 0.00039827566070016474,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 1.0
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 0.00040032563265413046,
"example_count": 300,
... |
1503f9e8-094e-408e-aeda-dc8a9f5c3756 | easy | KaustubhKumar05 | 2026-08-25 16:03:28.429898+00:00 | succeeded | b6d72aabc7da752cbf89d97df87ad90c9e9600f805ec3541ee7c84fab6ab2bb3 | 17,454 | null | """conv_sat: final_v1 + the DNGPU saturation cost (Freivalds & Liepins 2017).
Our hard gate clamp(1.2*sigmoid-0.1,0,1) has EXACTLY ZERO gradient outside
|x|>=2.398. The paper that introduced this gate says an explicit saturation
penalty is mandatory: without it units leave the linear band early and can
never return ("... | {
"id": "1503f9e8-094e-408e-aeda-dc8a9f5c3756",
"created_at": "2026-08-25 16:03:28.429898+00:00",
"db_md5": "5c8b9f4ebca3a2775b1bf2c2ddbd787e",
"submitter": "koz",
"github_login": "KaustubhKumar05",
"run_id": "7cd85a9a-a017-4d95-89b2-acbd937e65e2",
"tier": "easy",
"dataset_id": "e6",
"status": "succee... | {
"score": {
"mean_loss": 5.790616512298584,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.20202703028917313
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 7.067209720611572,
"example_count": 60,
... |
15080118-1b48-4d99-90d9-2ed514360572 | easy | jordanrubin | 2026-08-11 00:32:41.594585+00:00 | succeeded | 5627289ebca0e4853183d79604695f32d270f428a44ccf506384676ba86c6797 | 20,108 | null | """Weight-tied MLP-cell loop over soft digit states (mlploop).
Cell = MLP over [expected digits, pairwise digit products, RNS residue
simplices (fixed differentiable mixing over Z_p), N digits]. The cell that
learns one-step modular squaring from direct pairs (52% unseen at 14k rows,
day-20 screen) inside the exact-T ... | {
"id": "15080118-1b48-4d99-90d9-2ed514360572",
"created_at": "2026-08-11 00:32:41.594585+00:00",
"db_md5": "bdc145351ed7c8a97c4c6df2f8d43449",
"submitter": "Jordan Rubin",
"github_login": "jordanrubin",
"run_id": "16f40a57-e987-4e11-a7ef-6366a7fbad96",
"tier": "easy",
"dataset_id": "e5",
"status": "s... | {
"score": {
"mean_loss": 2.745816451003101,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.004999999888241291
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.1693024184939036,
"example_count": 600... |
150e8422-d4b5-45dc-a0ab-ab0f57419d2b | easy | thehimalayanleo | 2026-08-04 04:02:18.041647+00:00 | succeeded | 63c3dfc69bd486886b224a3a7b9422d7f9fc61a02e67443e68b757fc616a7c9a | 12,263 | null | """T-driven tensor-product recurrent network for repeated modular squaring.
The model is still fully learned: no arithmetic result is computed in Python.
Its inductive bias is a bank of shared digit interactions. A recurrent cell
mixes local digit pairs multiplicatively, communicates globally with attention,
and stor... | {
"id": "150e8422-d4b5-45dc-a0ab-ab0f57419d2b",
"created_at": "2026-08-04 04:02:18.041647+00:00",
"db_md5": "b0dc0cf466582b5bbbdcbf394085be6c",
"submitter": "Ajinkya K Mulay",
"github_login": "thehimalayanleo",
"run_id": "05b0aad1-7177-4a6b-a9a5-6c93da022e91",
"tier": "easy",
"dataset_id": "e1",
"stat... | {
"score": {
"mean_loss": 3.8053699731826782,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.03999999910593033
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.922940492630005,
"example_count": 100,... |
152d9acc-e1c3-4932-9950-82b0ff01f05d | easy | sapient-sapiens | 2026-09-01 00:17:18.786796+00:00 | succeeded | b3d82a142e5a0d205841d6c8d27e0d02f7821ae3447f57c56ec388eb8b9e8a0b | 20,399 | null | """One-block nested-Horner E5 model with an associative character carrier."""
from __future__ import annotations
import math
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
assert_model_state... | {
"id": "152d9acc-e1c3-4932-9950-82b0ff01f05d",
"created_at": "2026-09-01 00:17:18.786796+00:00",
"db_md5": "de920466d37befc1b2a3d8851538a4a5",
"submitter": "Ertondy",
"github_login": "sapient-sapiens",
"run_id": "87ace4d7-406f-4239-a225-15c713c05af9",
"tier": "easy",
"dataset_id": "e5",
"status": "su... | {
"score": {
"mean_loss": 2.141658423108102,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.004166666666666667
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.144833762014928,
"example_count": 600,... |
153c3667-9a9c-440f-aa32-407dafdf119b | easy | jordanrubin | 2026-08-08 10:19:53.599714+00:00 | succeeded | 19e434c7e9fc9d9d601292a0f5c0a4083813bc7790caf87c0582fc1bbcc2cad8 | 16,909 | null | """Autonomous digit-state recurrent Transformer for repeated modular squaring.
The model is deliberately organized around one learned transition:
p_0 = right_aligned_decimal_digits(x)
p_{k+1} = F_theta(p_k, decimal_digits(N))
The same two-block Transformer cell is applied exactly T times. T is used only
as ... | {
"id": "153c3667-9a9c-440f-aa32-407dafdf119b",
"created_at": "2026-08-08 10:19:53.599714+00:00",
"db_md5": "671c15c2db39c821999bae96a7defa51",
"submitter": "Jordan Rubin",
"github_login": "jordanrubin",
"run_id": "6d824d87-0014-4074-817d-f9b10ed637be",
"tier": "easy",
"dataset_id": "e5",
"status": "s... | {
"score": {
"mean_loss": 4.497614974637543,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.007916666489715379
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 4.498259844667708,
"example_count": 600,... |
1541ac74-b26f-47ee-96c5-66c2b6b42acc | easy | Yalyenea | 2026-08-12 12:49:39.124216+00:00 | succeeded | 64e396554dfe9df0975d29b313bc164f27c2b738e8ca7463cef0a37a8ac9a72b | 11,550 | null | """Gated decimal recurrence with wall-clock WSD learning rate."""
from __future__ import annotations
import math
import time
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
TokenLossBatch,
... | {
"id": "1541ac74-b26f-47ee-96c5-66c2b6b42acc",
"created_at": "2026-08-12 12:49:39.124216+00:00",
"db_md5": "ce4fd9088891918c370ae690bae01c9d",
"submitter": "yfff",
"github_login": "Yalyenea",
"run_id": "34e55241-3359-4568-b98e-4c52011328c1",
"tier": "easy",
"dataset_id": "e5",
"status": "succeeded",
... | {
"score": {
"mean_loss": 3.188721065513589,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.011666666679084302
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.7293391398784825,
"example_count": 600... |
15423aee-ea03-4bc1-8eae-7480c0994f8d | easy | oupadhyay | 2026-08-25 07:01:25.087105+00:00 | succeeded | 73090853935408b46f5c5e80173d153040a35a9e4c64a237b5386129ba37a60d | 4,080 | null | """Dynamic-width shared conditional quadratic recurrence."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import nn
from benchmark import OptimizerBundle,Submission,TokenLossBatch,assert_model_state
D,H=32,16
class C:
def __init__(s,vocab_size,max_seq_len):s.vocab_size,s.m... | {
"id": "15423aee-ea03-4bc1-8eae-7480c0994f8d",
"created_at": "2026-08-25 07:01:25.087105+00:00",
"db_md5": "418fe174c1189b2c2f8fdfe5edcaf7c6",
"submitter": "Ojasw Upadhyay",
"github_login": "oupadhyay",
"run_id": "0a480bb7-89dc-4da8-8240-cd258ec7758f",
"tier": "easy",
"dataset_id": "e5",
"status": "s... | {
"score": {
"mean_loss": 2.2841174908995807,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.0025000000124176342
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.15836570829554,
"example_count": 600... |
15519377-bc96-4f85-ad41-fa7c647f6148 | easy | TZZheng | 2026-08-07 04:45:49.381009+00:00 | succeeded | 1ed456f0d3844b1cae2c6af816fe8746785d40240b04b323eaedb333207b0df8 | 12,532 | null | """Learned recurrent modular-transition model.
The mutable digit state is initialized from x. A learned square-oriented
self-mixing network and an N-conditioned reduction network form one transition,
and exactly the same transition parameters are reused T times. No arithmetic
operation is hard-coded; only the public p... | {
"id": "15519377-bc96-4f85-ad41-fa7c647f6148",
"created_at": "2026-08-07 04:45:49.381009+00:00",
"db_md5": "0f7a2b2eb023528dd16ce9c174f85d6d",
"submitter": "TZZheng",
"github_login": "TZZheng",
"run_id": "a78b804c-61c4-4741-a105-45dab1cf2f9a",
"tier": "easy",
"dataset_id": "e1",
"status": "succeeded"... | {
"score": {
"mean_loss": 8.584530591964722,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.036666667088866234
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 10.598949432373047,
"example_count": 100... |
15621731-efd8-4537-b631-27c37f6c4328 | easy | RevellFTW | 2026-08-26 17:07:42.558122+00:00 | failed | 7bf30e19d90e7f6cadb4ad67d0b2cd51634a4fd2d29a507be6f840a0dd78f8c1 | 3,628 | null | """Basic single-pass Transformer with PyTorch AdamW."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
assert_model_state,
)
D_MODEL = 128
NUM_HEADS = 4
... | {
"id": "15621731-efd8-4537-b631-27c37f6c4328",
"created_at": "2026-08-26 17:07:42.558122+00:00",
"db_md5": "eaac69fe345f82973fcf392c504a40ff",
"submitter": "brtlnkrisztian",
"github_login": "RevellFTW",
"run_id": "f1c10485-b68a-4178-aa63-28d15d0d8e4c",
"tier": "easy",
"dataset_id": "e1",
"status": "f... | null
|
15694219-9a77-4eba-a497-faffc8b430ec | easy | christopherhwood | 2026-08-06 14:47:46.935628+00:00 | succeeded | 99a47737e0c2aa17db26d59b40583783851ccd44116a24aa06b6c7f19aa3adc0 | 21,687 | null | """Global learned binary-relation particles for repeated modular squaring.
Every arithmetic coefficient, branch polarity, global program weight, and
decimal decoder weight is randomly initialized and trained from evaluator
endpoints. Fixed architecture consists only of observable binary digits,
two-state carry, gener... | {
"id": "15694219-9a77-4eba-a497-faffc8b430ec",
"created_at": "2026-08-06 14:47:46.935628+00:00",
"db_md5": "0dc99d1ed8a0b715a29d1198d48f6b7f",
"submitter": "Chris Wood",
"github_login": "christopherhwood",
"run_id": "5a2ae400-2631-4720-a0cd-26983282badf",
"tier": "easy",
"dataset_id": "e1",
"status":... | {
"score": {
"mean_loss": 14.714484214782715,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.011666666716337204
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 14.985076904296875,
"example_count": 10... |
156ad0ef-37d7-44b9-8245-43fa0ccb6968 | easy | khushidahi | 2026-08-07 18:43:41.017416+00:00 | succeeded | 56dedd8c2e6e83495648967ada6356f078a22d091dd620ef8fc5dac255056c34 | 7,006 | null | """H0 conservative R3 Hard anchor.
This is the existing R3 architecture and AdamW recipe evaluated after exactly
3,000 optimizer steps. It provides the matched control for the boundary-aware
sequence-loss experiment.
"""
from __future__ import annotations
import random
import torch
import torch.nn.functional as F
f... | {
"id": "156ad0ef-37d7-44b9-8245-43fa0ccb6968",
"created_at": "2026-08-07 18:43:41.017416+00:00",
"db_md5": "6e4f5fd98bb6797a8e2c0b7c7f2b147a",
"submitter": "khushidahi",
"github_login": "khushidahi",
"run_id": "9a2d610b-3e69-44fa-9602-9be8ef200950",
"tier": "easy",
"dataset_id": "e5",
"status": "succ... | {
"score": {
"mean_loss": 2.1583372354507446,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.006666666828095913
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.1648740768432617,
"example_count": 60... |
156fdf51-d9a5-4c48-b855-7bc215ff9f05 | easy | DDanlov | 2026-08-29 12:16:41.105496+00:00 | failed | 659ca8a4bb5903a0cd25b8ec628ae63832238348b430b6ad482e244d6362ec50 | 16,188 | null | """
submission_eqprop_option_a.py
Option A: Modern Continuous Hopfield Energy Deliberator (eqprop_deep)
with 2 coupled physical layers (L=2), Mimetic Gram Attention Initialization,
Conformal Symplectic Velocity Verlet solver, and pure AdamW (lr=1e-3, weight_decay=1e-4)
as specified in ALGORITHM_GUIDE.md Section 7.1.
""... | {
"id": "156fdf51-d9a5-4c48-b855-7bc215ff9f05",
"created_at": "2026-08-29 12:16:41.105496+00:00",
"db_md5": "e3f4d3bbca01919a0d9c16d1593063fe",
"submitter": "DDanlov",
"github_login": "DDanlov",
"run_id": "9a5197cb-42c4-4415-a20d-d32d0ee13470",
"tier": "easy",
"dataset_id": "e5",
"status": "failed",
... | null
|
1570ad6a-3203-4c51-b57e-18ea0e7c1ff5 | easy | lzy54 | 2026-08-29 09:55:59.476882+00:00 | succeeded | 48b9b4aae7452f55f9fe1ada4450c7e32476444560da2bb60e5b5171a14a8a3f | 6,778 | null | """R5: long looped-tf run with post-pilot cosine learning-rate decay.
The recurrent state starts at zero. Each of four loops re-injects the encoded
problem, then applies the same three-block d=256 core. The original repository
uses d=256, 8 heads, three shared blocks, batch 64, and AdamW at 1e-4 for its
addition set... | {
"id": "1570ad6a-3203-4c51-b57e-18ea0e7c1ff5",
"created_at": "2026-08-29 09:55:59.476882+00:00",
"db_md5": "f0ab19609849e5dd5f23d3f4c49e4c54",
"submitter": "lzy54",
"github_login": "lzy54",
"run_id": "466204ce-c1dd-4a44-84cc-774e1efb3ef0",
"tier": "easy",
"dataset_id": "e5",
"status": "succeeded",
... | {
"score": {
"mean_loss": 2.2037573861420174,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.0054166666790843015
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.221217353247741,
"example_count": 60... |
1571633c-d2ef-4fad-b2e0-30417bfaa555 | easy | yashkant | 2026-08-09 01:50:34.602976+00:00 | succeeded | 89874e575e4ac80f9a89b1722969ae2412b6b1d26e5de54e8d315f13fa8c992f | 17,720 | null | """Round 143 with PyTorch add-zero attention sinks and raw EMA."""
from __future__ import annotations
import math
import torch
import torch.nn.functional as F
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
TokenLossBatch,
assert_model_state,
)
from torch import... | {
"id": "1571633c-d2ef-4fad-b2e0-30417bfaa555",
"created_at": "2026-08-09 01:50:34.602976+00:00",
"db_md5": "f537c7575db2a1b9242d6051d756f210",
"submitter": "Yash Kant",
"github_login": "yashkant",
"run_id": "51f8b86f-62f8-4c46-ae1c-cdc37c285d21",
"tier": "easy",
"dataset_id": "e5",
"status": "succeed... | {
"score": {
"mean_loss": 2.1403509285708067,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.006250000012417635
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 2.1418309254496624,
"example_count": 60... |
1578c365-a2fb-4518-adc2-07f4eef4662f | easy | briheet | 2026-08-07 04:50:34.384689+00:00 | succeeded | b42e29756f693b0e4170123dea2eb55a8a3938f03093b72493da8b64a18bd8b5 | 6,984 | null | """Small Universal Transformer with shared refinement passes."""
from __future__ import annotations
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from benchmark import (
ModelSpec,
OptimizerBundle,
OptimizerSpec,
Submission,
assert_model_state,
)
D_MODEL = 32
NUM_HEA... | {
"id": "1578c365-a2fb-4518-adc2-07f4eef4662f",
"created_at": "2026-08-07 04:50:34.384689+00:00",
"db_md5": "6b9c1899f2e0f948cfc2aad618dd9e2a",
"submitter": "Aizenn",
"github_login": "briheet",
"run_id": "eb9e97b2-24d5-4d04-84f7-a266f5997a1f",
"tier": "easy",
"dataset_id": "e1",
"status": "succeeded",... | {
"score": {
"mean_loss": 2.541015148162842,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.04333333298563957
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 3.099623441696167,
"example_count": 100,
... |
157c7e93-6c0e-4ad1-a1ea-12862db91f4c | easy | rudygt | 2026-08-04 17:41:40.041777+00:00 | succeeded | 7e9f98841a702ba896bfc2e6c9d5b2c9171faef445966679b5ff85ab155d4710 | 20,195 | null | """OL-2 — quantized tied-step recurrence in the contract's digit I/O ("contract
wrapper", LOG.md 2026-08-04 T-series recipe).
Pipeline: digit-token prompt -> 1-block bidirectional encoder -> attention-pooled
h0 -> tied step block (2x RMSNorm-MLP residual + zero-init complex-square path),
unrolled T times (T parsed fro... | {
"id": "157c7e93-6c0e-4ad1-a1ea-12862db91f4c",
"created_at": "2026-08-04 17:41:40.041777+00:00",
"db_md5": "e1008728bb06a4dbe924f75014d31319",
"submitter": "Rudy Alvarez",
"github_login": "rudygt",
"run_id": "5ba4df85-30be-46b5-a6b0-56edc7adb1d0",
"tier": "easy",
"dataset_id": "e1",
"status": "succee... | {
"score": {
"mean_loss": 1.9060572385787964,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.0
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 1.8416697978973389,
"example_count": 100,
"exa... |
15834463-9a82-496b-97a2-b3ad6b4c5813 | easy | DDanlov | 2026-08-26 18:53:33.831965+00:00 | failed | 79ad120240e8b184f8de8eadbf8d8ae04621646312f4cc84a0b3db219b4068f4 | 29,714 | null | import math
import time
from typing import Optional, Tuple, Dict, Any
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim import Optimizer
try:
from onelayerdeeper.api import Submission, OptimizerBundle, ModelSpec, OptimizerSpec
except ImportError:
Submission, OptimizerBundle, M... | {
"id": "15834463-9a82-496b-97a2-b3ad6b4c5813",
"created_at": "2026-08-26 18:53:33.831965+00:00",
"db_md5": "7df05019081fa949ba26637c34973db2",
"submitter": "DDanlov",
"github_login": "DDanlov",
"run_id": "5ac6af4f-d9d5-4afc-a574-d7e77e44a79e",
"tier": "easy",
"dataset_id": "e5",
"status": "failed",
... | null
|
1588f27c-456c-433e-b7dc-3f577369fe57 | easy | DDanlov | 2026-08-21 07:22:23.126809+00:00 | failed | a3ef1886e59ab484b87b3bc2ec185eac39ceaf29ae01b9397d01cdedf2828c59 | 25,755 | null | from __future__ import annotations
import math
import time
from typing import Any, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim import Optimizer
try:
from benchmark.api import ModelSpec, OptimizerBundle, OptimizerSpec, Submission, assert_model_state
except Imp... | {
"id": "1588f27c-456c-433e-b7dc-3f577369fe57",
"created_at": "2026-08-21 07:22:23.126809+00:00",
"db_md5": "a6cd921133563dd28e4761243b1ff3a2",
"submitter": "DDanlov",
"github_login": "DDanlov",
"run_id": "fe7252d2-4717-44a0-b3a7-39f3a4d0b4fc",
"tier": "easy",
"dataset_id": "e5",
"status": "failed",
... | null
|
15947cf3-ea68-441c-ab7f-00c7ce8e8d4f | easy | YangWang92 | 2026-08-17 05:30:11.310334+00:00 | succeeded | 44fc930d809c31a6480ef79b59b1620f4f36943377541193bab7b764615787de | 46,820 | null | """Limb (bignum) circuit submission.
The h1 metrics pinned the hard tier at ~60-64 bit moduli (the training-loss
floor of 100-250 is exactly the fp64 quantisation error of 18-19 digit
integers), so every value here is a little-endian limb vector, base 10^4.
No intermediate ever exceeds ~1.4e9, so a single float64 hold... | {
"id": "15947cf3-ea68-441c-ab7f-00c7ce8e8d4f",
"created_at": "2026-08-17 05:30:11.310334+00:00",
"db_md5": "7c68cdea41972b37f3eb5b23a332c615",
"submitter": "Yang Wang ",
"github_login": "YangWang92",
"run_id": "870335e2-4ebd-4b9d-a9bb-d61699071f2d",
"tier": "easy",
"dataset_id": "e5",
"status": "succ... | {
"score": {
"mean_loss": 0.0,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 1.0
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 0.0,
"example_count": 600,
"exact_accuracy": 1.0,
"... |
15a3e52a-3816-4749-b798-baeee5c9cf2e | easy | datathesisai | 2026-08-05 07:24:24.902381+00:00 | succeeded | b56a373960bbfdb2a52a526f83fadd676bf070c6f7fe9a98f9d2318175f3eeda | 10,468 | null | """Fast task-agnostic recurrent encoder-processor-decoder.
This candidate prioritizes learning the base transition before spending compute
on very deep latent unrolls. It uses one immutable encoded prompt, one shared
input-recalled processor, and an independent small-initialized output head.
"""
from __future__ impo... | {
"id": "15a3e52a-3816-4749-b798-baeee5c9cf2e",
"created_at": "2026-08-05 07:24:24.902381+00:00",
"db_md5": "9c064525f98d2f181354502a66b271cf",
"submitter": "datathesisai",
"github_login": "datathesisai",
"run_id": "5a04c5dd-3634-4ca8-b264-2e7c63de93b0",
"tier": "easy",
"dataset_id": "e5",
"status": "... | {
"score": {
"mean_loss": 8.149816751480103,
"primary_metric": "mean_exact_accuracy",
"num_measurements": 2,
"mean_exact_accuracy": 0.011666666716337204
},
"seeds": [
{
"seed": 74,
"evaluation": {
"ood": {
"loss": 7.958191394805908,
"example_count": 600,... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.