We introduce Faynt, a family of 10M- and 75M-parameter Transformer policies for Super Smash Bros. Melee, each controlling all 26 characters with a single checkpoint. After reinforcement learning (RL), the 10M wins 240 of 244 same-character games (98.4%) against fourteen specialist and multi-character releases on their supported rosters, with a winning record against every release. These opponents retain 21- or 24-frame action delays; Faynt uses no added delay, and we have not isolated the effect of this difference. In a separate evaluation against a privately supplied zero-delay Slippi-AI model, the 10M wins all 68 games across two conditioning settings. We study architecture, optimization, scaling, and hyperparameter transfer to guide pretraining on approximately 840,000 human replays. Post-training combines rank- and outcome-based curricula, 75M-to-10M distillation, and RL restricted to Fox mirror matches. On the initial 152-game benchmark, the supervised 10M wins 69.7% of games, compared with 45.4% for the pretrained 75M, despite higher overall held-out controller-prediction loss. The weighted validation loss used for supervised checkpoint selection agrees with the win-rate ordering of all four pretrained and supervised policies. After supervised post-training, both models take less damage per minute, build larger early leads, and win more often after losing the first life. Optimized inference on recorded game states averages 5.2 ms per decision for the 10M and 8.7 ms for the 75M on an NVIDIA T4, excluding emulator execution and communication. We open-source the weights, both benchmark suites, and a platform for automated model tournaments.
Figures & tables
Configuration
Faynt-10M
Faynt-75M
Trainable parameters
10,163,629
75,305,709
Model width
384
768
Transformer blocks
5
11
Query / key-value heads
6 / 2
12 / 3
Head dimension
64
64
Feed-forward width
768
1,920
Table 1: Model configurations. Parameter counts include the frame encoder, Transformer, and controller heads.
Figure 1: Faynt policy architecture. Dashed connections supply current controller categories directly to the output heads. The main-stick head also conditions on the next joint-controller category, using the recorded target during imitation training and a sample during play. AttnRes denotes Full Attention Residuals.
Figure 2: Compute allocation and optimizer selection. Left: observed endpoints and fitted fixed-compute curves; dashed segments extend below the tested size ranges, and stars mark optima found by extrapolation. Right: six Muon configurations at 20M parameters and 166M targets, with the best AdamW run as reference. Muon’s reference uses learning rate 0.02, momentum 0.95, Nesterov, and five Newton–Schulz iterations; row labels identify individual changes.
Figure 3: Successive halving on the full corpus. Curves join scheduled validation evaluations and end when a recipe is eliminated (crosses); stars mark the halving selections. Counts give active trials. The 10M panel shows selection through rung 3: lower weight decay wins narrowly, while lower Muon learning rate is ultimately selected after recovery. The 75M winner continues alone in rung 5.
Figure 4: Final pretraining runs. Top: validation trajectories before divergence and calibration-law predictions; stars mark the selected checkpoints. Bottom: full continuations, with manual stops dotted. The 10M finalists (navy: lower Muon learning rate; gray: lower weight decay) remain within 0.0005 NLL through 8B targets, then differ in stability. Points are recorded evaluations; lines join successive observations.
Figure 5: Post-training validation loss and learning-rate schedules. Open circles mark each round’s source evaluation; stars mark retained checkpoints. The 10M continues through two distillation rounds. The dashed 75M continuation is stopped at the crossed endpoint. Rates are normalized to the reference learning rates in Appendix E.2 .
Setting
75M
10M, run 1
10M, leash run
Starting policy
supervised 75M
supervised 10M
run 1, step 1,318
Self-play share of policy loss
50%
50%
100%
Learning rate, policy and value
4×10−5
10−4
10−4
Policy-gradient weight wPG
1
1
3
Reference KL weight λ→
0.1
0.1
0.003
Reference KL weight λ←
0
0
0.003
Table 2: RL settings and checkpoint budgets. Each run retains its model’s supervised endpoint as the reference policy. The leash run inherits the value network from step 1,318 and resets both optimizers. Training time is summed over launches and includes in-loop evaluation; the selected 10M lineage totals 63.5 hours. Common settings and implementation details are in Appendix E.6 .
Figure 6: RL training and evaluated checkpoints. Stars mark the evaluated 10M and 75M policies; the smaller star marks the 10M checkpoint from which the leash run continues. The diamond is the earlier 75M checkpoint used in Section 6 . Dashed segments continue past selected checkpoints; thin segments precede the implementation speedup described in Appendix E.6 .
Figure 7: In-loop evaluation against MIMIC during RL. Each round contains 25.6 seconds on each of 30 emulators. Stock margin is stocks taken minus stocks lost per minute. Open markers follow the checkpoints used; dotted lines mark 75M step 980 and 10M step 632, which falls between evaluation rounds at steps 583 and 706. The 10M continuation starts 45 hours into its first run.
Figure 8: Initial benchmark results before reinforcement learning. Bars show win rates and win counts in each policy’s 152-game evaluation. Baseline schedules substitute the supervised 75M Faynt comparator, so opponent composition differs between Faynt and the reference views. Slippi-AI includes medium-v2 and the scheduled specialists. All reported outcomes have completed the validation checks in Appendix F.3 .
Figure 9: Expanded evaluation after reinforcement learning. Cells show Faynt wins/games; color encodes win rate. Row groups contain five single-character specialists, seven twelve-character releases, and two four-character releases. Supported and forced mirrors jointly cover all 26 characters; extended-roster games keep the opponent on a supported fighter. Column pairs compare 75M RL step 980 and 10M RL step 632. Asterisks mark cells containing separately rerun games (Appendix F.6 ).
Figure 10: Inference optimization and the 60 Hz decision budget. Left: real-time factor on synthetic inputs, with darker bars showing successive optimizations. Right: mean processing time and 99th-percentile latency on recorded-game inputs, including feature conversion. Dashed lines mark one frame. All timings use random weights and exclude the emulator and communication with it; output parity is tested separately with trained checkpoints (Appendix H ).
Figure 11: Prediction loss and win rate can favor different models. The four pretrained and supervised checkpoints are evaluated on fixed validation data and the original 152-game suite. Open markers denote pretrained models; filled markers denote supervised models. Arrows connect checkpoints of the same size. The supervised 10M wins 37 more games than the pretrained 75M despite higher overall loss. Their weighted losses follow their ranking by win rate.
Figure 12: Faynt takes less damage and builds larger early leads. Each size uses the same match configurations across the two stages: 145 at 75M and 139 at 10M. Damage rates are percentage points received per minute, averaged equally across games. Stock lead is the difference in remaining lives after one minute; terminal scores are retained for games that end earlier. The last panel shows the fraction of time Faynt spends in damage-reaction states.
Figure 13: More wins from early leads and deficits. Win rates after taking or losing the first stock on the two-stage comparison set. Labels give wins/games; simultaneous first-stock losses are excluded.
Figure 14: Stock preservation after supervised training. Top: total stock losses, highlighting those at 50% damage or less. Bottom: their share of all stocks lost, with counts. Both stages use the same 145 configurations at 75M and 139 at 10M.
Figure 15: Playing strength improves after reinforcement learning. Open and filled markers show the supervised and RL checkpoints against one privately supplied zero-delay Slippi-AI model under two conditioning settings. Labels give wins/games: 16 games on Final Destination (FD) and 18 across six stages per setting. The RL checkpoints are 10M step 632 and 75M step 980. Appendix F.13 gives the full win/loss records.
Figure 16: 75M performance at two RL checkpoints. Markers show wins in 16 Final Destination games per opponent; labels give step 222 and step 980 counts. Both checkpoints win 18/18 against every opponent in the separate six-stage evaluation (Appendix F.12 ).
Appendix figures & tables47 assets
Supplementary material from the paper’s appendix.
Appendix
Calibration
Final corpus
Accepted replays
67,129
839,942
Training replays
65,824
823,083
Validation replays
659
8,483
Test replays
646
8,376
All valid targets
1,500,000,244
17,839,509,998
Training targets
1,470,952,166
17,481,727,198
Appendix
Table A.1: Dataset builds. A target is one valid transition from frame t to t+1 for one player perspective. Manifest identifiers are the abbreviated values recorded in the study. The final split identifier is 9c4ae826ae97 .
AdamW
Learning rate
Batch
NLL
0.0010
65,536
0.860504
0.0003
65,536
0.870815
0.0030
65,536
0.875388
0.0001
65,536
0.914326
0.0010
262,144
0.914411
Appendix
Table C.1: Complete optimizer screen. Muon uses auxiliary AdamW at 0.001 and shared weight decay 0.01. N denotes Nesterov momentum; NS is the number of Newton–Schulz steps. Bold entries are the selected endpoints.
Size
Setting varied
0.5×
1×
2×
5M
Muon rate
0.943590
0.941835
0.939627
5M
Auxiliary rate
0.950677
0.941835
0.926854
5M
Weight decay
0.939114
0.941835
0.937286
50M
Muon rate
0.915042
0.912638
0.927680
50M
Auxiliary rate
0.920879
0.912638
0.914731
50M
Weight decay
0.924369
0.912638
0.917862
Appendix
Table C.2: Hyperparameter transfer at 41.94M targets per run. Each row varies one setting from the selected 20M recipe. The baseline is shared across the three rows at each size, yielding fourteen distinct runs.
Profile
Parameters
Gp
3M
3,003,373
12,847,328
5M
5,053,805
29,304,944
10M
10,163,629
72,262,720
20M
20,040,877
151,385,600
50M
50,127,021
394,013,760
75M
75,305,709
596,823,632
Appendix
Table C.3: Exact architecture costs and calibration endpoints. Gp is FLOPs per target. C1, C2, and C3 denote 0.0052, 0.0104, and 0.0208 EFLOP. The smallest measured model has the lowest loss in each group.
Recipe
Multipliers
R1
R2
R3
R4
R5
10M
wd-low
1,1,0.8
0.81019
0.80431
0.801534
0.99334
0.94213
muon-low
0.8,1,1
0.81022
0.80387
0.801560
aux-low
1,0.8,1
0.81252
0.80572
center
1,1,1
0.81220
0.80635
aux-high
1,1.25,1
0.81294
Appendix
Table C.4: All 24 full-corpus recipes and 47 rung evaluations. Empty cells follow elimination. For interaction recipes, ℓ=1/2 and h=2 ; their three-letter names give the multipliers in order. The 75M local-fit recipe uses (0.77798125836,1.2289867927,1.1422710484) .
10M
75M
Rung
Updates
Active
Updates
Active
1
38,730
8
7,526
16
2
77,461
4
15,052
8
3
116,192
2
22,579
4
4
154,923
1
30,105
2
5
193,654
1
37,631
1
Appendix
Table C.5: Matched-target halving schedule. Active counts refer to the trials evaluated at the end of each rung. The 10M winner continues alone in rungs 4 and 5; the 75M winner continues alone in rung 5.
Recipe
Best step
Validation
Test
Stop targets
10M Muon-low
122,064
0.796817
0.797658
11,811,160,064
10M WD-low
122,064
0.797050
9,663,676,416
75M Muon-low
86,016
0.764821
0.765435
7,516,192,768
Appendix
Table C.6: Recovery checkpoints and complete continuation lengths. The test split is evaluated once for each frozen winner. A blank test entry indicates that the runner-up was excluded from this evaluation.
Model (M)
CPU cores
Microbatch / accumulation
Workers / prefetch
Targets/s, thousands (95% interval)
USD / 1B targets
RTX PRO 6000
20.04
8
64/4
6/2
165.0 [160.3, 170.8]
6.60
20.04
16
64/4
12/1
146.8 [142.9, 150.1]
8.14
20.04
32
64/4
0/2
187.1 [179.1, 195.5]
7.51
50.13
8
64/4
0/2
131.3 [125.4, 136.3]
8.30
50.13
16
64/4
8/2
139.2 [136.3, 142.4]
8.58
Appendix
Table D.1: Complete selected execution and delivery recipes. Microbatch size counts 256-frame windows; accumulation preserves the 65,536-target global batch. Bold rates identify the fastest CPU tier for each GPU/model pair. Prefetch is inactive when there are zero workers; its stored setting is retained. All hosts have 64 GiB RAM. Costs use the GPU, CPU, and memory rates recorded on August 27, 2026 and exclude startup.
GPU
CPU cores
A (k/s)
β
Rlog2
Max. residual
Fixed-core fits
RTX PRO 6000
8
139.9
0.159
0.785
6.2%
RTX PRO 6000
16
133.2
0.121
0.815
4.5%
RTX PRO 6000
32
128.7
0.371
0.870
10.4%
H100 SXM
8
180.7
0.751
0.952
13.8%
H100 SXM
16
138.6
0.477
0.652
30.5%
Appendix
Table D.2: Throughput coefficients and fit diagnostics. The normalization size is 50,127,021 parameters. Each fixed-core fit uses three measurements at one CPU allocation; each compact fit uses the best measured CPU tier at each size. Maximum residual is maxi∣Ti−Ti∣/Ti over the three anchors.
Model (M)
GPU (8 CPU cores)
Hours
Targets/s
Targets (B)
Pred. NLL
Compact-policy candidates, $100
5.054
A100
29.50
1,545,647 E
164.148
0.763553
7
A100
29.50
1,071,942 E
113.840
0.756780
8
A100
29.50
922,620 E
97.982
0.754180
10.164
A100
29.50
705,074 E
74.879
0.749773
Larger-policy candidates, $200
Appendix
Table D.3: Fixed-dollar planning comparisons. M denotes measured throughput, F fitted interpolation, and E extrapolation. Bold entries are selected within each candidate set. Exposures count processed targets, including repeated frames. The 75M budget includes the measured cold-start allowance.
Pool
Replays
Eligible perspective
Natural distribution
823,083
Both players
All decided games
823,057
Winner
Master or diamond winners
471,316
Winner
Master winners
256,532
Winner
Diamond winners
214,784
Winner
Diamond subset in later mixture
28,504
Winner
Appendix
Table E.1: Eligible replay pools. Later rounds use every master-winning replay and a deterministic sample of diamond-winning replays, giving approximately 90% master and 10% diamond visits.
Slice
Perspectives
Targets
Master won
2,635
674,560
Master lost
1,932
494,592
Diamond won
2,235
572,160
Diamond lost
2,208
565,248
Platinum won
3,612
924,672
Platinum lost
4,342
1,111,552
Appendix
Table E.2: Rank–outcome validation slices. An additional undecided master game contributes two perspectives and 512 targets; it is excluded from the six decided-outcome slices.
Model
Stage
Frames
L
Master-won
W
10M
Natural
270
0.77510
0.79191
0.79024
10M
All winners
270
0.77233
0.78693
0.78543
10M
Master + diamond winners
360
0.77166
0.77798
0.77708
10M
Master winners
900
0.77282
0.77078
0.77080
75M
Natural
180
0.74519
0.75550
0.75426
75M
All winners
180
0.74230
0.74977
0.74873
Appendix
Table E.3: Round 1 stage endpoints. Frames are targets processed within the stage, in millions. L is overall validation NLL and W is the weighted selection loss. The full-corpus stage precedes every filtered stage.
Round
Model
Horizon
Schedule
Objective
1
10M / 75M
1.8B / 1.2B
Cosine 0.1→0.001
NLL
2
10M / 75M
1.2B / 1.5B
Hold h=0.1
NLL
3
10M
900M
Hold h=0.1
KD, α=0.5 , T=1
3, stopped
75M
Hold h=0.1
NLL
4
10M
900M
Hold h=0.05
KD, α=0.5 , T=1
4, control
10M
300M
Hold h=0.05
KD, α=0.7 , T=1
Appendix
Table E.4: Executed rounds. Horizons are planned processed targets. A blank horizon for the stopped 75M continuation indicates that only its 449M-target stopping point is used here. KD controls each run for 300M targets.
α=0.7
T=2
Targets (M)
Main W
W
L
W
L
0
0.75990
0.75984
0.77071
0.75982
0.77070
50
0.76086
0.76075
0.77074
0.76268
0.77257
100
0.76106
0.76085
0.77042
0.76230
0.77175
150
0.76075
0.76041
0.77141
0.76289
0.77362
200
0.76053
0.76025
0.76993
0.76289
0.77246
Appendix
Table E.5: Round 4 control trajectories. All runs start at step 181,516. The 50–200M evaluations share the held rate; warmup lasts 6M targets for controls and 18M for the main run. Controls begin decaying at 240M, and the 900M main run at 720M. L is overall NLL.
Step
Targets (M)
W
Overall NLL
Gradient median
127,214
0
0.72492
0.74149
127,976
50
0.72756
0.74431
0.1205
128,738
100
0.72734
0.74400
0.1255
129,500
150
0.72771
0.74383
0.1263
130,262
200
0.72771
0.74504
0.1304
131,024
250
0.72790
0.74235
0.1350
Appendix
Table E.6: The 75M continuation after round 2. The learning-rate multiplier is 0.1 at every trained checkpoint. All metrics are validation measurements except the interval median of the training gradient norm.
Model
Checkpoint
Step
Overall
Master-won
Diamond-won
W
10M
Pretrained
122,064
0.79682
0.81315
0.79691
0.81152
10M
Round 1
149,512
0.77281
0.77076
0.77094
0.77078
10M
Round 2
167,800
0.77220
0.76381
0.77001
0.76443
10M
Round 3
181,516
0.77071
0.75889
0.76825
0.75983
10M
Round 4
195,248
0.76896
0.75647
0.76641
0.75746
75M
Pretrained
86,016
0.76481
0.77617
0.76467
0.77502
Appendix
Table E.7: Validation metrics at every selected checkpoint. Round names follow the chronological experiments. Values come from the selected checkpoint rows; final alignment steps can differ slightly from those rows.
Model
Checkpoint
Overall
Master-won
Diamond-won
W
10M
Pretrained
0.79766
0.79694
0.79928
0.79717
10M
Round 1
0.77356
0.75469
0.77223
0.75644
10M
Round 2
0.77291
0.74758
0.77119
0.74994
10M
Round 3
0.77137
0.74280
0.76886
0.74541
10M
Round 4
0.76974
0.74044
0.76722
0.74312
75M
Pretrained
0.76543
0.75915
0.76714
0.75995
Appendix
Table E.8: Test results at the same checkpoints. Every selected round improves all four reported metrics relative to the preceding selected round.
10M
75M
Slice
Pretrained
Round 1
Pretrained
Round 1
Master won
0.81315
0.77078
0.77617
0.73566
Master lost
0.79648
0.76365
0.76250
0.73142
Diamond won
0.79691
0.77098
0.76467
0.73810
Diamond lost
0.79777
0.77552
0.76672
0.74479
Platinum won
0.79945
0.78122
0.76855
0.75292
Appendix
Table E.9: Every rank–outcome slice before and at the end of round 1. Both sizes improve on every slice relative to pretraining, with the largest reductions on master-winning actions.
Model
Checkpoint
Button NLL
Stick NLL
Button H
Stick H
10M
Pretrained
0.26732
0.52950
0.2612
0.5300
10M
Round 4
0.25502
0.51394
0.2498
0.5178
75M
Pretrained
0.25287
0.51194
0.2521
0.5127
75M
Round 2
0.24311
0.49835
0.2367
0.5001
Appendix
Table E.10: Controller-head diagnostics on all validation slices. H denotes the mean predictive entropy of a head, in nats. The final checkpoints retain button entropy within 10% of their pretrained values.
Model
Targets (M)
L
W
10M
799
0.77472
0.76962
10M
849
0.77552
0.76960
10M
899
0.77423
0.76979
10M
949
0.77441
0.76973
75M
1049
0.74495
0.72936
75M
1099
0.74367
0.72905
Appendix
Table E.11: Final constant-rate evaluations in round 2. Targets count additional training within the round. The learning-rate multiplier is 0.1 throughout. L is overall validation NLL; W is the weighted selection loss.
Figure E.1: Post-training diagnostics. (a) Weighted validation NLL and training gradient-norm medians rise during the additional 75M cycle, which is stopped at 449M targets. (b) The 10M controls at evaluations with the same held rate of 0.05: teacher weight 0.7 gives lower weighted NLL at all four points, while temperature 2 gives higher NLL. (c) Each run’s own final decay, shown as a percentage of its budget. The dashed line is the temperature control’s source loss, 0.75982; its trained endpoint is worse and the source is retained. Lines connect recorded observations.
Size
Stage
Checkpoint step
Original run designation
75M
Pretrained
86,016
Final pretraining
75M
Supervised post-training
127,214
cur-3
10M
Pretrained
122,064
Final pretraining v2
10M
Supervised post-training
195,248
cur-6-kd
75M
RL
980
RL run
10M
RL
632
RL leash run
Appendix
Table F.1: Frozen Faynt checkpoints. The supervised post-training stage includes curriculum learning for both sizes and knowledge distillation for 10M.
Block
Games
Assignment
MIMIC Master Fox
10
Fox mirrors against fox-master
CPU9 Fox
10
Fox mirrors against the level-9 controller
MIMIC core mirrors
12
Six native characters, two games each
Slippi-AI core mirrors
24
Twelve native characters, two games each
Roster extension
96
Allocation in Table F.3
Appendix
Table F.2: Original external-panel allocation for each Faynt checkpoint.
Opponent
Games
Character and checkpoint assignment
CPU9
28
All fourteen fighters in E
MIMIC native
34
Seventeen native fighters outside the common core
MIMIC forced
6
fox-master on Kirby, Zelda, and Pichu
Slippi specialists
4
Donkey Kong: dk_d18_imitation_v2 ; Dr. Mario: doc_d18_imitation_v3
Slippi forced
24
medium-v2 on E except Donkey Kong and Dr. Mario
Appendix
Table F.3: The 96 original extension games. Each assigned fighter is played twice. Forced mirrors place the opponent on a fighter outside its deployed roster.
All 26 characters
Common native core
Policy
Opponent
W–L
Margin
W–L
Margin
Faynt 10M
MIMIC
35–17
+1.02
8–4
+0.67
Faynt 10M
Slippi-AI
34–18
+0.79
4–8
−0.83
Faynt 75M
MIMIC
46–6
+1.71
10–2
+1.08
Faynt 75M
Slippi-AI
40–12
+1.31
7–5
+0.17
Slippi-AI
MIMIC
34–18
+0.38
11–1
+1.42
Appendix
Table F.4: Original pairwise results. Faynt checkpoints are supervised post-trained. Margins are stocks taken minus stocks conceded per game. The common-core columns reuse a subset of the full-roster games.
Release
Roster
Delay
Conditioning
SFIL
1
21
SFAT
diamond
12
21
Master Player
falco_d21_ditto_v4
1
21
Ginger
fox_d21_ditto_hax_v3
1
21
Hax
fox_d21_ditto_v4
1
21
Cody
fox_d24_ditto_v4
1
24
Cody
Appendix
Table F.5: Expanded opponent configurations. Roster size is deployed character count; delay is in game frames.
Pretrained
Supervised
Block
W–L
Stocks
W–L
Stocks
MIMIC Fox
1–9
20–39
3–7
32–32
CPU9 Fox
8–2
38–22
10–0
40–7
MIMIC core
2–10
33–46
10–2
45–32
Slippi-AI core
2–22
55–93
15–9
82–75
Extension
56–40
320–278
85–11
373–195
Appendix
Table 46
Pretrained
Supervised
Block
W–L
Stocks
W–L
Stocks
MIMIC Fox
1–9
15–38
5–5
33–32
CPU9 Fox
8–2
38–23
10–0
40–7
MIMIC core
2–10
24–45
8–4
40–32
Slippi-AI core
1–23
38–95
8–16
71–87
Extension
37–59
274–316
75–21
355–219
Appendix
Table 47
MIMIC : block
Games
W–L
Stocks
Faynt Fox
10
7–3
32–32
CPU9 Fox
10
10–0
40–15
Faynt mirrors
12
2–10
32–45
Slippi-AI mirrors
24
2–22
59–93
Roster extension
96
24–72
232–337
Total
152
45–107
395–522
Appendix
Table 48
Slippi-AI : block
Games
W–L
Stocks
MIMIC Master Fox
10
8–2
37–27
CPU9 Fox
10
10–0
40–4
MIMIC mirrors
12
11–1
46–29
Faynt mirrors
24
9–15
75–82
Roster extension
96
34–62
233–322
Total
152
72–80
431–464
Appendix
Table 49
75M RL
10M RL
Opponent release
Games
W–L
Stocks
W–L
Stocks
fox_d21_ditto_v4
12
2–10
20–46
12–0
48–4
fox_d24_ditto_v4
12
3–9
33–45
12–0
48–6
SFIL
12
9–3
44–32
12–0
48–1
fox_d21_ditto_hax_v3
12
12–0
48–23
12–0
48–2
falco_d21_ditto_v4
12
1–11
22–47
11–1
47–22
Appendix
Table 50
75M RL
10M RL
Opponent release
Games
W–L
Stocks
W–L
Stocks
fox_d21_ditto_v4
50
7–43
106–191
35–15
177–127
fox_d24_ditto_v4
50
17–33
135–174
43–7 ∗
188–100
SFIL
50
15–35
146–174
41–9
188–98
fox_d21_ditto_hax_v3
50
30–20
173–148
46–4
193–72
falco_d21_ditto_v4
50
4–46
97–193
35–15
175–128
Appendix
Table 51
75M RL
10M RL
Opponent release
Games
W–L
Stocks
W–L
Stocks
fox_d21_ditto_v4
50
50–0
200–18
50–0 ∗
200–13
fox_d24_ditto_v4
50
50–0 ∗
200–19
50–0 ∗
200–9
SFIL
50
50–0
200–20
50–0
200–17
fox_d21_ditto_hax_v3
50
50–0
200–11
50–0 ∗
200–7
falco_d21_ditto_v4
50
50–0
200–16
50–0
200–16
Appendix
Table 52
Opponent release
Games
75M RL
10M RL
fox_d21_ditto_v4
62
52–10
62–0
fox_d24_ditto_v4
62
53–9
62–0
SFIL
62
59–3
62–0
fox_d21_ditto_hax_v3
62
62–0
62–0
falco_d21_ditto_v4
62
51–11
61–1
gm
52
34–18
52–0
Appendix
Table F.6: Full-roster mirror results after RL. Entries are Faynt wins–losses against each frozen release. Supported and forced components are listed separately in Appendices F.8 and F.10 .
10M RL 632
75M RL 222
75M RL 980
Opponent
FD
Six stages
FD
Six stages
FD
Six stages
delay0/FoxFD
16–0
18–0
15–1
18–0
14–2
18–0
delay0/FalcoFD
16–0
18–0
16–0
18–0
14–2
18–0
FoxFD0
16–0
18–0
16–0
18–0
16–0
18–0
MarthFD0
16–0
18–0
15–1
18–0
16–0
18–0
PeachFD
16–0
18–0
16–0
18–0
16–0
18–0
Appendix
Table F.7: Results against zero-delay specialists. Entries are Faynt wins–losses. FD denotes Final Destination. Each checkpoint plays 112 games in the FD condition and 126 in the six-stage condition, using the same weights across characters.
Master Player
Cody
Faynt checkpoint
FD
Six stages
FD
Six stages
Total
10M supervised
6–10
9–9
3–13
5–13
23–45
10M RL step 632
16–0
18–0
16–0
18–0
68–0
75M supervised
9–7
8–10
5–11
3–15
25–43
75M RL step 980
16–0
17–1
13–3
12–6
58–10
Appendix
Table F.8: Results against the privately supplied Slippi-AI checkpoint. Entries are Faynt wins–losses. The two column groups use different conditioning settings of the same opponent model. FD denotes Final Destination; the six-stage condition contains three games per stage. Supervised denotes the post-trained policy before RL.
Model
Stage
Games
Wins
Duration
Faynt dmg/min
Opp. dmg/min
Lead
75M
Pretrained
145
65
193.2
106.4
122.1
−0.055
Supervised
145
117
184.5
95.8
151.4
+0.531
10M
Pretrained
139
46
177.1
118.2
110.2
−0.317
Supervised
139
95
186.5
98.7
144.6
+0.460
Appendix
Table G.1: Two-stage results on the same configurations. Duration is in seconds. Damage received (dmg/min) is in damage percentage points per minute, averaged equally across games. Stock lead is evaluated at 60 seconds.
Model
Stage
Win after taking first
Win after conceding first
75M
Pretrained
42/65 (64.6%)
23/80 (28.8%)
Supervised
96/107 (89.7%)
19/36 (52.8%)
10M
Pretrained
40/57 (70.2%)
6/82 (7.3%)
Supervised
76/93 (81.7%)
19/46 (41.3%)
Appendix
Table G.2: Wins after taking or conceding the first stock on the common two-stage cohorts. Each entry gives wins divided by eligible games, followed by the percentage. Simultaneous first-stock losses are excluded.
Model
Stage
Low-percent / all losses
Share
Mean loss percent
75M
Pretrained
86/456
18.9%
99.6
Supervised
51/325
15.7%
108.9
10M
Pretrained
98/469
20.9%
92.2
Supervised
47/348
13.5%
106.4
Appendix
Table G.3: Faynt stock losses on the common cohorts. Low-percent losses occur at 50% damage or less, using the last recorded value before the loss. The final column gives mean damage percentage before a stock loss.
Model
Stage
Neutral
Changed
Damage
Shield
75M
Pretrained
26.08
23.25
11.37
4.61
Supervised
25.90
24.02
10.22
3.19
10M
Pretrained
22.83
23.12
12.39
4.00
Supervised
25.67
23.73
10.46
3.15
Appendix
Table G.4: Controller activity and time in action states on the common cohorts. Values are percentages of frames, averaged equally across games. Damage and shield columns use the state definitions above.
Model
Stage
Wins
Taken
Conceded
Margin
Remaining
75M
Pretrained
69
3.066
3.145
−0.079
1.884
Supervised
123
3.763
2.243
+1.520
2.171
10M
Pretrained
49
2.559
3.401
−0.842
1.857
Supervised
106
3.546
2.480
+1.066
2.179
Appendix
Table G.5: Stock outcomes on the full initial suite. Each row contains 152 games. Taken, conceded, and margin are per-game means; remaining is mean Faynt stocks remaining conditional on winning.
Per frame
Reference
Optimized
Cache write
A loop over the batch in Python reads each row’s validity flag and write position to the host, in every layer
One indexed write each for keys and values per layer; positions stay on the device
Cache read
Each layer’s 256-frame window is copied into chronological order, reading each row’s length and next write position to the host
Attention runs over the ring in storage order through a validity mask, which is exact because rotary phases are applied when keys are written
Grouped-query attention [ 8 ]
PyTorch’s scaled dot-product attention on FP32 inputs expands keys and values to every query head
Batched matrix products per key-value head, without expansion
Checks
Eleven encoder range checks, three cache checks, a reset check, and the controller head’s label checks, each a device-to-host read
None; inputs are valid by construction
Input transfer
One host-to-device copy per input tensor
One copy per data type into static buffers (GPU)
Dispatch
Every kernel launched from Python
One CUDA-graph replay, with sampling and the stick head outside the graph (GPU)
Appendix
Table H.1: Per-frame operations of the two implementations. The optimized implementation computes the same formulas with the same parameters.
Model
Device
Pair
Max ∣Δlogit∣
Mean KL
Max KL
Top-1
Play
10M
CPU
Reference vs. sync-free
6.1×10−5
1.5×10−13
2.6×10−11
0 of 4,800
none in 1,200
10M
GPU
Reference vs. sync-free
6.1×10−5
1.5×10−13
1.6×10−11
0 of 4,800
none in 1,200
10M
GPU
Sync-free vs. optimized
0
0
0
0 of 4,800
none in 1,200
75M
CPU
Reference vs. sync-free
6.4×10−4
1.5×10−13
1.2×10−10
0 of 4,800
none in 1,200
75M
GPU
Reference vs. sync-free
4.0×10−4
2.0×10−13
1.2×10−10
0 of 4,800
none in 1,200
75M
GPU
Sync-free vs. optimized
0
0
0
0 of 4,800
none in 1,200
Appendix
Table H.2: Output parity between implementations on the first 1,200 frames of the recorded game, from both players’ perspectives, with the evaluated checkpoints (10M RL step 632, 75M RL step 980) on a Modal T4 container (GPU rows) and its host CPU (CPU rows). The first three columns of numbers compare the logits of both controller heads; KL divergences between action distributions are per head and in nats. Top-1 counts head outputs whose most likely category differs (1,200 frames × 2 players × 2 heads). Play reports the first frame on which actions sampled from the same seed differ. Optimized: the sync-free step replayed from a CUDA graph.
Figure H.1: Where the time goes. Real-time factor of the decision loop, paced at 60 Hz with Faynt playing one game, split into building the model input (dark segment) and the policy step with sampling, decoding, and the controller write (light segment), teal for the 10M and navy for the 75M. EPYC rows are the median host’s own run on the recorded game; T4 rows use synthetic frames, and their input build includes the copy to the GPU. The dashed line marks one frame.
10M
75M
Play
Watch
Play
Watch
GPU
Time
p99
Time
p99
Time
p99
Time
p99
NVIDIA T4
5.2
6.8
6.8
8.9
8.7
12.1
9.5
11.9
NVIDIA L4
3.2
4.7
4.0
5.5
5.4
6.4
6.2
7.5
NVIDIA A10G
3.1
4.6
3.7
5.1
4.6
5.9
5.6
6.9
Appendix
Table H.3: The whole loop of the optimized implementation on the recorded game, paced at 60 Hz: mean processing time per frame (Time) and 99th-percentile latency (p99), in ms. Play and watch are as defined in Appendix H.1 . Decisions later than one frame are 0.6% for the T4 75M and 0.3% for the A10G 10M when playing, and none in any other cell.
Figure H.2: Serving several games from one forward pass. 99th-percentile latency of the optimized implementation, one row per game, paced at 60 Hz on synthetic frames, on an NVIDIA T4, L4, and A10G GPU; teal is the 10M and navy the 75M. The dashed line marks one frame. Each point is a single run, and decisions later than one frame occur only for the T4 10M at 8 games (0.8%) and the T4 75M at 64 games (0.4%).
Scaling reasoning typically spends more compute on reinforcement learning (RL) or on inference. We show that a completed RL training history can yield policies stronger than the checkpoints visited by its optimizer. We call this policy-space scaling: expanding the deployable policy set accessible from a fixed RL history, without extending training or increasing per-query inference computation. We instantiate it with SURGE (Scaling Up RL Gradient-free via Eigenspace fusion). SURGE combines two checkpoints from the same RL run: a high-accuracy anchor and a competitive donor that generates shorter responses. It expresses both checkpoints as changes from their shared initialization, then spectrally decomposes the anchor's update to retain its dominant component and incorporate the donor's complementary component. With a fixed target for how much of the anchor update to retain, SURGE determines the block size from the weights without testing candidate policies. We evaluate two 1.5B mathematical-reasoning histories, DeepSeek and Nemotron, and one 7B coding history, OLMo. SURGE improves benchmark-average accuracy over both input checkpoints while using fewer reasoning tokens than the anchor. It reaches 54.17% on DeepSeek AIME24 against a measured native maximum of 50.83%, and 83.7% on OLMo HumanEval+ against 82.8%. These gains exceed the observed training curves. Geometric controls support the importance of RL-update structure beyond weight displacement or token reduction alone. Each constructed model runs as a single policy. Our findings identify stored RL history as a reusable scaling resource: the capability available from a training run need not end at its best checkpoint.
Bangji Yang, Jiajun Fan, Hongba Ma +2
University of Illinois at Urbana-Champaign · Tsinghua University
We present Mach-Mind-4-Flash, a 35B-parameter Mixture-of-Experts (MoE) agentic model with 3B activated parameters. Through post-training optimization alone without scaling pre-training compute, the model achieves performance on par with or surpassing that of 100B-parameter-class models. By introducing scalable agentic interaction environments for large-scale reinforcement learning, the model attains significant performance gains on real-world application tasks. Our pipeline comprises three stages: (1) a unified RL/OPD training infrastructure with dynamic multi-teacher scheduling and operator-level acceleration, delivering 17% end-to-end training speedup; (2) multiple domain-specific RL experts trained in parallel across Reasoning, General, and Agent tracks, then fused into a single generalist via Multi-Teacher On-Policy Distillation (MOPD) -- a routed reverse-KL objective that eliminates the see-saw degradation of mixed-reward RL; (3) Hybrid Median-length Policy Optimization (HMPO), a single-stage token-efficiency method that compresses reasoning chains by 19--46% with ≤0.7 percentage-point accuracy loss. Mach-Mind-4-Flash scores 92.70 on AIME'26, 82.82 on IFBench, 80.74 on Behavioral-SafetyBench, 75.80 on BFCL-v4, 72.31 on BrowseComp-zh, and 84.20 on ClawBench -- leading or matching models with 10--30× its activated size at a fraction of the inference cost.
Developing generalist systems that retain human-like data efficiency is a central challenge. While world models (WMs) offer a promising path, existing research often conflates architectural mechanisms with the independent impact of model \emph{scale}. In this work, we use a minimalist transformer world model to analyze scaling behaviors on the Atari 100k benchmark, using fixed offline datasets derived from a presupposed expert policy. Our results reveal that environments fundamentally fall into distinct scaling regimes, even when constrained by identical offline data budgets and model capacities. For individual tasks, some environments naturally allow models to pass the interpolation threshold, yielding monotonic improvements in the overparameterized regime, while others remain trapped in the classical regime, where larger world models degrade fidelity. In the unified setting, i.e., a single transformer trained on a suite of 26 Atari environments, we uncover that joint training stabilizes scaling dynamics, ensuring monotonic gains across all environments, regardless of their distinct inherent scaling regimes. Finally, we demonstrate that improved fidelity translates directly to downstream control, with policies learned entirely within the simulated dynamics achieving a median expert-random-normalized score of 0.770. Our findings suggest that future progress lies as much in precise scaling strategies as in architectural innovation.