MovieLens-1M · Amazon · KuaiRand-1K · Amazon-Qwen · MIND-large · seeds 42, 43 · four RTX cards

How does the notepad compare to the field?

MovieLens-1M
Amazon
KuaiRand-1K
MIND-large
frozen Qwen
2 seeds
test set
19 models

Every model in this project is our own rewrite of someone else's, so a comparison is only worth anything if all of them get the same deal. We gave nineteen models one identical training budget, ran each twice with different random starts, and scored them the same way on three very different datasets: movie ratings, Amazon game purchases, and short-video watch logs. Nothing was tuned for our own method. The point was not to win — the notepad's case rests on being cheap and inspectable — but to find out honestly where it sits among models people actually publish.

Takeaway

On ID vectors the notepad is mid-table everywhere. On shared frozen content vectors its published configuration leads Amazon outright — but a configuration ladder shows where that comes from: given the identical one-variable swap the baselines received, the notepad scores 0.085, below even its own ID row; the content anchor lifts it to 0.130; the tuned write constants more than double it to 0.284. The lead is real but belongs to the tuned configuration, and the baselines were not given an equivalent tuning pass. On MIND-large the content-anchored notepad sits upper-mid-table. The one claim that survives every column untouched: rank-16 matches full memory everywhere.

Part 1 · Learned ID vectors

Every model learns its item table from interaction IDs alone — the like-for-like protocol of the original campaign. Columns are comparable down their own length only.

MovieLens-1M

Ranked by nDCG@10 on the held-out test set. All runs finished.

HSTU

0.6419

Mamba4Rec

0.6336

LRURec

0.6310

SASRec

0.6259

BSARec

0.6235

TiSASRec

0.6224

DuoRec

0.6211

GRU4Rec

0.6179

#	Model	HR@1	HR@10	nDCG@10	nDCG@20
1	HSTU	0.4536	0.8359	0.6419	0.6596
2	Mamba4Rec	0.4494	0.8240	0.6336	0.6524
3	LRURec	0.4412	0.8247	0.6310	0.6494
4	SASRec	0.4372	0.8219	0.6259	0.6448
5	BSARec	0.4325	0.8212	0.6235	0.6421
6	TiSASRec	0.4315	0.8215	0.6224	0.6414
7	DuoRec	0.4285	0.8198	0.6211	0.6401
8	GRU4Rec	0.4286	0.8161	0.6179	0.6353
9	FWPRec (dense)	0.4136	0.8269	0.6148	0.6332
10	FWPRec (rank-16)	0.4081	0.8248	0.6120	0.6316
11	LinRec	0.3853	0.7982	0.5859	0.6060
12	FPMC	0.3400	0.7797	0.5470	0.5724
13	SeqRules	0.3535	0.7439	0.5400	0.5661
14	BERT4Rec	0.3132	0.6858	0.4918	0.5169
15	BPRMF	0.1696	0.6682	0.3913	0.4315
16	ItemKNN	0.1369	0.5780	0.3282	0.3743
17	Popularity	0.0892	0.4363	0.2386	0.2851
—	NRMS	n/a	n/a	n/a	n/a
—	LSTUR	n/a	n/a	n/a	n/a

Amazon Video Games

Ranked by nDCG@10 on the held-out test set. All runs finished.

BSARec

0.1412

BERT4Rec

0.1411

SASRec

0.1383

TiSASRec

0.1379

GRU4Rec

0.1364

HSTU

0.1362

LRURec

0.1353

Mamba4Rec

0.1329

#	Model	HR@1	HR@10	nDCG@10	nDCG@20
1	BSARec	0.0795	0.2159	0.1412	0.1538
2	BERT4Rec	0.0771	0.2191	0.1411	0.1549
3	SASRec	0.0747	0.2162	0.1383	0.1508
4	TiSASRec	0.0739	0.2157	0.1379	0.1509
5	GRU4Rec	0.0763	0.2092	0.1364	0.1487
6	HSTU	0.0745	0.2115	0.1362	0.1497
7	LRURec	0.0740	0.2093	0.1353	0.1482
8	Mamba4Rec	0.0723	0.2070	0.1329	0.1455
9	FPMC	0.0457	0.2044	0.1123	0.1403
10	SeqRules	0.0562	0.1801	0.1097	0.1326
11	FWPRec (dense)	0.0548	0.1699	0.1061	0.1184
12	FWPRec (rank-16)	0.0542	0.1698	0.1056	0.1179
13	DuoRec	0.0460	0.1668	0.0989	0.1132
14	ItemKNN	0.0465	0.1607	0.0953	0.1184
15	BPRMF	0.0415	0.1469	0.0875	0.1002
16	Popularity	0.0300	0.1354	0.0745	0.0882
17	LinRec	0.0294	0.1352	0.0744	0.0876
—	NRMS	n/a	n/a	n/a	n/a
—	LSTUR	n/a	n/a	n/a	n/a

KuaiRand-1K

Ranked by nDCG@10 on the held-out test set. All runs finished.

BPRMF

0.4077

SASRec

0.3901

TiSASRec

0.3878

Mamba4Rec

0.3654

GRU4Rec

0.3586

LinRec

0.3547

FWPRec (rank-16)

0.3531

FWPRec (dense)

0.3527

#	Model	HR@1	HR@10	nDCG@10	nDCG@20
1	BPRMF	0.0982	0.8329	0.4077	0.4453
2	SASRec	0.0980	0.7722	0.3901	0.4217
3	TiSASRec	0.0960	0.7719	0.3878	0.4198
4	Mamba4Rec	0.0760	0.7622	0.3654	0.4004
5	GRU4Rec	0.0723	0.7532	0.3586	0.3924
6	LinRec	0.1008	0.6993	0.3547	0.4007
7	FWPRec (rank-16)	0.0912	0.7231	0.3531	0.4090
8	FWPRec (dense)	0.0914	0.7218	0.3527	0.4087
9	HSTU	0.0665	0.7498	0.3509	0.3881
10	DuoRec	0.0741	0.7294	0.3473	0.3898
11	BSARec	0.0617	0.7390	0.3409	0.3809
12	LRURec	0.0511	0.7095	0.3160	0.3624
13	SeqRules	0.0205	0.2004	0.0914	0.1414
14	Popularity	0.0092	0.0931	0.0426	0.0668
—	ItemKNN	n/a	n/a	n/a	n/a
—	FPMC	n/a	n/a	n/a	n/a
—	BERT4Rec	n/a	n/a	n/a	n/a
—	NRMS	n/a	n/a	n/a	n/a
—	LSTUR	n/a	n/a	n/a	n/a

Does the ranking hold at other list lengths?

Hit rate on MovieLens-1M at four cutoffs. Reading only the ten-item column would overstate the notepad: it places second there, but ninth when only the single top pick counts.

Model	HR@1	HR@5	HR@10	HR@20
HSTU	0.4536	0.7506	0.8359	0.9053
Mamba4Rec	0.4494	0.7421	0.8240	0.8978
LRURec	0.4412	0.7412	0.8247	0.8973
SASRec	0.4372	0.7340	0.8219	0.8963
BSARec	0.4325	0.7353	0.8212	0.8945
FWPRec (dense)	0.4136	0.7293	0.8269	0.8991
FWPRec (rank-16)	0.4081	0.7280	0.8248	0.9022
GRU4Rec	0.4286	0.7281	0.8161	0.8845

Part 2 · Frozen content vectors

The same models, given identical frozen Qwen content vectors instead of learned IDs. Different information, so nothing in this part may be compared against Part 1 — except deliberately, in the paired exhibit below, where that comparison is the entire point.

What one change of representation does on Amazon

Identical split, candidates, seeds and training budget; the only change is swapping each model's learned ID table for the same frozen content vectors. The notepad moves from eleventh place to first.

Model	ID vectors	Content vectors	Δ
FWPRec (dense)	0.1061	0.2842	+0.1781
FWPRec (rank-16)	0.1056	0.2836	+0.1780
FPMC	0.1123	0.2562	+0.1439
HSTU	0.1362	0.1994	+0.0632
LRURec	0.1353	0.1967	+0.0614
BSARec	0.1412	0.1908	+0.0496
Mamba4Rec	0.1329	0.1836	+0.0507
GRU4Rec	0.1364	0.1635	+0.0271
SASRec	0.1383	0.1502	+0.0119
TiSASRec	0.1379	0.1445	+0.0065
DuoRec	0.0989	0.1387	+0.0399
LinRec	0.0744	0.1100	+0.0356
BPRMF	0.0875	0.0806	-0.0069

Read the notepad's row with care: unlike every other row, its content-side number uses ten tuned constants from an earlier search — the baselines' deltas are representation alone, the notepad's is representation plus tuning. The ladder below separates the two; the notepad's symmetric-swap score is 0.0852.

Where the notepad's content number actually comes from

One knob-group at a time, from the identical swap the baselines received up to the published configuration. The anchor and the write mechanism contribute little; the tuned constants carry the result. Under strictly equal treatment the notepad scores 0.0852 — below its own ID-column row.

Configuration	What it adds	nDCG@10
Same one-variable swap the baselines got	the symmetric row	0.0846
+ content anchor	mean of the user's item vectors	0.1305
+ sequential writes	evolving write state	0.1318
+ tuned constants	write gains ×10, mixing weight, week/quarter time scales	0.2842

Amazon Video Games — frozen Qwen content vectors

Ranked by nDCG@10 on the held-out test set. All runs finished.

FWPRec (dense)

0.2842

FWPRec (rank-16)

0.2836

FPMC

0.2562

HSTU

0.1994

LRURec

0.1967

BSARec

0.1908

Mamba4Rec

0.1836

GRU4Rec

0.1635

#	Model	HR@1	HR@10	nDCG@10	nDCG@20
1	FWPRec (dense)	—	0.4982	0.2842	—
2	FWPRec (rank-16)	—	0.5000	0.2836	—
3	FPMC	—	0.4254	0.2562	—
4	HSTU	—	0.3548	0.1994	—
5	LRURec	—	0.3570	0.1967	—
6	BSARec	—	0.3406	0.1908	—
7	Mamba4Rec	—	0.3289	0.1836	—
8	GRU4Rec	—	0.2913	0.1635	—
9	SASRec	—	0.2480	0.1502	—
10	TiSASRec	—	0.2344	0.1445	—
11	DuoRec	—	0.2205	0.1387	—
12	LinRec	—	0.1821	0.1100	—
13	BPRMF	—	0.1447	0.0806	—
—	Popularity	n/a	n/a	n/a	n/a
—	ItemKNN	n/a	n/a	n/a	n/a
—	SeqRules	n/a	n/a	n/a	n/a
—	BERT4Rec	n/a	n/a	n/a	n/a
—	NRMS	n/a	n/a	n/a	n/a
—	LSTUR	n/a	n/a	n/a	n/a

MIND-large — frozen Qwen content vectors

Ranked by nDCG@10 on the held-out test set. All runs finished.

Mamba4Rec

0.4242

BSARec

0.4225

HSTU

0.4113

DuoRec

0.4105

LRURec

0.4090

NRMS

0.4067

FWPRec (dense)

0.4056

GRU4Rec

0.4036

#	Model	HR@1	HR@10	nDCG@10	nDCG@20
1	Mamba4Rec	—	0.7710	0.4242	—
2	BSARec	—	0.7700	0.4225	—
3	HSTU	—	0.7626	0.4113	—
4	DuoRec	—	0.7544	0.4105	—
5	LRURec	—	0.7616	0.4090	—
6	NRMS	—	0.7607	0.4067	—
7	FWPRec (dense)	—	0.7527	0.4056	—
8	GRU4Rec	—	0.7557	0.4036	—
9	FWPRec (rank-16)	—	0.7496	0.4025	—
10	SASRec	—	0.7460	0.3964	—
11	TiSASRec	—	0.7444	0.3945	—
12	LSTUR	—	0.7473	0.3943	—
13	FPMC	—	0.7132	0.3782	—
14	BPRMF	—	0.6995	0.3600	—
15	LinRec	—	0.6632	0.3228	—
16	ItemKNN	—	0.6243	0.3028	—
17	SeqRules	—	0.6095	0.2960	—
18	Popularity	—	0.6011	0.2910	—
—	BERT4Rec	n/a	n/a	n/a	n/a

How it was run

177 of 177 runs finished so far. Each one records the code version, a checksum of the data it read, the exact settings, the GPU it used and how long it took, and each re-checks its own scores from its saved predictions before being counted.

Two seeds per model, reported as the mean; we do not quote a standard deviation from two numbers. The three datasets are scored against different-sized candidate lists, so a column is only comparable down its own length, never across. Rows marked n/a could not be run and say why in the record: NRMS and LSTUR need item text the three ID datasets do not carry (they do run natively on MIND-large), ItemKNN's memory grows with the square of a user's history, and FPMC's three embedding tables over 4.4 million videos exceed the cards we have. The frozen-Qwen columns are reported separately from the ID columns and never merged: same models, different information. A handful of MIND-large and KuaiRand control runs are still finishing.