Repository navigation
Expand file tree
/
Copy pathindex.html
More file actions
1177 lines (1107 loc) · 65.3 KB
/
Copy pathindex.html
File metadata and controls
1177 lines (1107 loc) · 65.3 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>EvolveScaler: Synthesizing Information-Evolution Contexts</title>
<meta name="description" content="EvolveScaler synthesizes long, multi-turn contexts in which information keeps changing, with deterministically verifiable answers computed by replaying an executable state machine.">
<link rel="icon" href="assets/logo.png">
<style>
:root{
--navy:#0C2A57; --navy-2:#123A70; --navy-3:#1B4B8A;
--teal:#0E9E97; --teal-2:#14BFAE; --green:#3DB54A;
--ink:#0B1626; --body:#33425C; --muted:#6B7C97;
--bg:#FFFFFF; --bg-2:#F4F8FC; --bg-3:#EAF1F9;
--line:#DCE6F2; --line-2:#C6D5E8;
--amber:#E09B21; --plum:#8B55B4; --rose:#DC2648;
--shadow:0 1px 2px rgba(12,42,87,.05), 0 8px 28px rgba(12,42,87,.07);
--shadow-lg:0 2px 6px rgba(12,42,87,.06), 0 24px 60px rgba(12,42,87,.13);
--radius:16px;
--sans:'Inter','SF Pro Display',-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,Helvetica,Arial,sans-serif;
--mono:'SF Mono',ui-monospace,Menlo,Consolas,'Liberation Mono',monospace;
}
*{box-sizing:border-box}
html{scroll-behavior:smooth;scroll-padding-top:76px}
body{
margin:0;background:var(--bg);color:var(--body);
font-family:var(--sans);font-size:16.5px;line-height:1.72;
-webkit-font-smoothing:antialiased;text-rendering:optimizeLegibility;
}
h1,h2,h3,h4{color:var(--ink);line-height:1.22;margin:0;letter-spacing:-.022em}
p{margin:0 0 1em}
a{color:var(--teal);text-decoration:none}
a:hover{text-decoration:underline}
.wrap{max-width:1140px;margin:0 auto;padding:0 28px}
.narrow{max-width:820px}
/* ---------- nav ---------- */
nav{
position:sticky;top:0;z-index:100;background:rgba(255,255,255,.82);
backdrop-filter:saturate(180%) blur(16px);-webkit-backdrop-filter:saturate(180%) blur(16px);
border-bottom:1px solid var(--line);
}
.nav-in{max-width:1140px;margin:0 auto;padding:0 28px;height:64px;display:flex;align-items:center;gap:28px}
.nav-logo{display:flex;align-items:center;gap:9px;font-weight:750;color:var(--ink);font-size:17px;letter-spacing:-.03em;flex-shrink:0}
.nav-logo img{height:30px;mix-blend-mode:darken}
.nav-links{display:flex;gap:4px;margin-left:auto;flex-wrap:wrap}
.nav-links a{
color:var(--muted);font-size:14.5px;font-weight:560;padding:7px 13px;border-radius:9px;
transition:.18s;text-decoration:none;white-space:nowrap;
}
.nav-links a:hover{color:var(--navy);background:var(--bg-2);text-decoration:none}
.nav-links a.on{color:var(--teal);background:rgba(14,158,151,.09)}
/* ---------- hero ---------- */
/* Base stays pure white and the decorative blobs are kept to the right: the
logo has a near-white matte and no alpha, so anything tinted behind it
reads as a visible rectangle. */
.hero{position:relative;overflow:hidden;padding:76px 0 68px;background:#fff;border-bottom:1px solid var(--line)}
.hero::before,.hero::after{content:'';position:absolute;border-radius:50%;filter:blur(95px);opacity:.4;pointer-events:none}
.hero::before{width:640px;height:640px;top:-250px;right:-170px;background:radial-gradient(circle,#14BFAE,transparent 68%);animation:drift 22s ease-in-out infinite}
.hero::after{width:720px;height:720px;bottom:-430px;right:-120px;background:radial-gradient(circle,#3462AC,transparent 68%);animation:drift 26s ease-in-out infinite reverse}
@keyframes drift{0%,100%{transform:translate(0,0) scale(1)}50%{transform:translate(46px,34px) scale(1.09)}}
.hero-in{position:relative;z-index:1}
.hero img.mark{height:96px;margin-bottom:28px;mix-blend-mode:darken}
.badge{
display:inline-flex;align-items:center;gap:8px;background:#fff;border:1px solid var(--line-2);
color:var(--navy);font-size:13.5px;font-weight:620;padding:6px 15px;border-radius:999px;margin-bottom:20px;
box-shadow:0 1px 3px rgba(12,42,87,.06);
}
.badge .dot{width:7px;height:7px;border-radius:50%;background:var(--green);box-shadow:0 0 0 3px rgba(61,181,74,.18)}
h1.title{font-size:clamp(34px,5.1vw,55px);font-weight:800;letter-spacing:-.038em;margin-bottom:20px;max-width:16em}
h1.title .grad{background:linear-gradient(100deg,var(--navy) 4%,var(--teal) 52%,var(--green) 96%);-webkit-background-clip:text;background-clip:text;-webkit-text-fill-color:transparent}
.lede{font-size:19.5px;line-height:1.66;color:var(--body);max-width:41em;margin-bottom:30px}
.cta{display:flex;gap:12px;flex-wrap:wrap;margin-bottom:46px}
.btn{
display:inline-flex;align-items:center;gap:9px;padding:12px 24px;border-radius:11px;
font-size:15.5px;font-weight:640;transition:.2s;border:1px solid transparent;cursor:pointer;text-decoration:none;
}
.btn-pri{background:linear-gradient(100deg,var(--navy) 0%,var(--teal) 118%);color:#fff;box-shadow:0 4px 16px rgba(12,42,87,.24)}
.btn-pri:hover{transform:translateY(-2px);box-shadow:0 9px 26px rgba(12,42,87,.3);text-decoration:none;color:#fff}
.btn-sec{background:#fff;color:var(--navy);border-color:var(--line-2)}
.btn-sec:hover{border-color:var(--teal);color:var(--teal);transform:translateY(-2px);text-decoration:none}
.stats{display:grid;grid-template-columns:repeat(auto-fit,minmax(148px,1fr));gap:14px}
.stat{background:#fff;border:1px solid var(--line);border-radius:14px;padding:19px 20px;box-shadow:var(--shadow);transition:.22s}
.stat:hover{transform:translateY(-3px);box-shadow:var(--shadow-lg);border-color:var(--line-2)}
.stat .n{font-size:31px;font-weight:790;color:var(--navy);letter-spacing:-.035em;line-height:1.1;font-variant-numeric:tabular-nums}
.stat .n .u{font-size:19px;color:var(--teal)}
.stat .k{font-size:12.6px;color:var(--muted);font-weight:600;text-transform:uppercase;letter-spacing:.07em;margin-top:5px}
/* ---------- sections ---------- */
section{padding:82px 0;border-bottom:1px solid var(--line)}
section.alt{background:var(--bg-2)}
.eyebrow{
display:inline-block;font-size:12.8px;font-weight:730;text-transform:uppercase;letter-spacing:.13em;
color:var(--teal);margin-bottom:13px;
}
h2.sec{font-size:clamp(27px,3.5vw,37px);font-weight:780;letter-spacing:-.032em;margin-bottom:16px}
.sec-lede{font-size:18px;color:var(--body);max-width:46em;margin-bottom:40px}
/* ---------- generic cards ---------- */
.grid{display:grid;gap:18px}
.g2{grid-template-columns:repeat(auto-fit,minmax(300px,1fr))}
.g3{grid-template-columns:repeat(auto-fit,minmax(260px,1fr))}
.card{background:#fff;border:1px solid var(--line);border-radius:var(--radius);padding:26px;box-shadow:var(--shadow);transition:.22s}
.card:hover{border-color:var(--line-2);box-shadow:var(--shadow-lg);transform:translateY(-2px)}
.card h3{font-size:18.5px;font-weight:700;margin-bottom:9px}
.card p{font-size:15.4px;color:var(--body);margin:0}
.icon{
width:42px;height:42px;border-radius:11px;display:grid;place-items:center;margin-bottom:15px;
background:linear-gradient(135deg,rgba(12,42,87,.09),rgba(20,191,174,.16));font-size:20px;
}
/* ---------- problem strip ---------- */
.vs{display:grid;grid-template-columns:1fr auto 1fr;gap:22px;align-items:stretch;margin-top:8px}
.vs .pane{border-radius:var(--radius);padding:26px;border:1px solid var(--line);background:#fff}
.vs .pane.bad{border-color:#F2CDD5;background:linear-gradient(180deg,#FEF6F7,#fff)}
.vs .pane.good{border-color:#BFE7DF;background:linear-gradient(180deg,#F1FBF8,#fff)}
.vs .tag{font-size:12.4px;font-weight:730;letter-spacing:.1em;text-transform:uppercase;margin-bottom:11px}
.vs .bad .tag{color:var(--rose)}
.vs .good .tag{color:var(--teal)}
.vs .arrow{display:grid;place-items:center;color:var(--line-2);font-size:26px}
.vs ul{margin:12px 0 0;padding-left:19px;font-size:15.2px}
.vs li{margin-bottom:7px}
@media(max-width:820px){.vs{grid-template-columns:1fr}.vs .arrow{transform:rotate(90deg)}}
/* ---------- tabs ---------- */
.tabs{display:flex;gap:8px;flex-wrap:wrap;margin-bottom:24px}
.tab{
padding:9px 18px;border-radius:10px;border:1px solid var(--line);background:#fff;cursor:pointer;
font-family:inherit;font-size:14.6px;font-weight:620;color:var(--muted);transition:.18s;
}
.tab:hover{border-color:var(--teal);color:var(--teal)}
.tab.on{background:linear-gradient(100deg,var(--navy),var(--navy-3));color:#fff;border-color:transparent;box-shadow:0 3px 12px rgba(12,42,87,.22)}
.panel{display:none;animation:fade .32s ease}
.panel.on{display:block}
@keyframes fade{from{opacity:0;transform:translateY(7px)}to{opacity:1;transform:none}}
/* ---------- pipeline ---------- */
.pipe{display:grid;grid-template-columns:repeat(3,1fr);gap:14px;margin-bottom:26px}
.step{
background:#fff;border:1px solid var(--line);border-radius:14px;padding:22px;cursor:pointer;
transition:.22s;position:relative;text-align:left;font-family:inherit;
}
.step:hover{border-color:var(--teal);transform:translateY(-3px)}
.step.on{border-color:transparent;background:linear-gradient(135deg,var(--navy),var(--navy-3));box-shadow:0 10px 30px rgba(12,42,87,.26)}
.step.on .s-n,.step.on h4{color:#fff}
.step.on .s-d{color:rgba(255,255,255,.8)}
.s-n{font-size:12.4px;font-weight:760;letter-spacing:.12em;color:var(--teal);margin-bottom:9px}
.step h4{font-size:17.5px;font-weight:720;margin-bottom:7px}
.s-d{font-size:14.6px;color:var(--muted);line-height:1.6}
.pipe-detail{background:#fff;border:1px solid var(--line);border-radius:var(--radius);padding:30px;box-shadow:var(--shadow)}
@media(max-width:820px){.pipe{grid-template-columns:1fr}}
/* ---------- primitives explorer ---------- */
.prim-wrap{display:grid;grid-template-columns:300px 1fr;gap:20px}
.prim-list{display:flex;flex-direction:column;gap:6px}
.prim{
text-align:left;padding:11px 15px;border-radius:10px;border:1px solid transparent;background:transparent;
cursor:pointer;font-family:inherit;font-size:15px;font-weight:600;color:var(--body);transition:.16s;
display:flex;align-items:center;gap:11px;
}
.prim:hover{background:var(--bg-3)}
.prim.on{background:#fff;border-color:var(--teal);color:var(--navy);box-shadow:var(--shadow)}
.prim .pn{
width:24px;height:24px;border-radius:7px;display:grid;place-items:center;flex-shrink:0;
background:var(--bg-3);font-size:11.5px;font-weight:750;color:var(--muted);
}
.prim.on .pn{background:linear-gradient(135deg,var(--navy),var(--teal));color:#fff}
.prim-detail{background:#fff;border:1px solid var(--line);border-radius:var(--radius);padding:30px;box-shadow:var(--shadow);min-height:270px}
.prim-detail h3{font-size:22px;margin-bottom:6px}
.prim-detail .spec{color:var(--teal);font-weight:640;font-size:15px;margin-bottom:18px}
.prim-detail .ex-l{font-size:12.4px;font-weight:730;letter-spacing:.1em;text-transform:uppercase;color:var(--muted);margin-bottom:9px}
.prim-detail .ex{background:var(--bg-2);border-left:3px solid var(--teal);border-radius:0 10px 10px 0;padding:15px 18px;font-size:15px;color:var(--body)}
@media(max-width:860px){.prim-wrap{grid-template-columns:1fr}}
/* ---------- replay widget ---------- */
.replay{display:grid;grid-template-columns:1fr 300px;gap:20px}
.log{background:#fff;border:1px solid var(--line);border-radius:var(--radius);padding:24px;box-shadow:var(--shadow)}
.sys{background:var(--bg-3);border-radius:11px;padding:14px 17px;font-size:14.4px;color:var(--body);margin-bottom:18px;border-left:3px solid var(--navy)}
.turn{margin-bottom:16px;opacity:.28;transition:opacity .4s}
.turn.shown{opacity:1}
.turn-h{font-size:12.6px;font-weight:740;letter-spacing:.09em;text-transform:uppercase;color:var(--navy);margin-bottom:9px}
.rec{
display:flex;gap:11px;align-items:flex-start;padding:9px 13px;border-radius:9px;margin-bottom:6px;
font-size:14.4px;background:var(--bg-2);border:1px solid transparent;transition:.3s;
}
.rec.invalid{background:#FEF4F6;border-color:#F3D2D9;color:var(--muted);text-decoration:line-through;text-decoration-color:rgba(220,38,72,.45)}
.rec .id{font-family:var(--mono);font-size:12.4px;font-weight:640;color:var(--teal);flex-shrink:0;padding-top:1px}
.rec.invalid .id{color:var(--rose)}
.rec .flag{
margin-left:auto;font-size:11px;font-weight:730;letter-spacing:.06em;text-transform:uppercase;
color:var(--rose);background:#fff;border:1px solid #F3D2D9;border-radius:6px;padding:1px 7px;flex-shrink:0;
text-decoration:none;
}
.state{background:linear-gradient(165deg,var(--navy),#0A1F44);border-radius:var(--radius);padding:24px;color:#fff;position:sticky;top:82px;height:fit-content}
.state h4{color:#fff;font-size:13px;font-weight:730;letter-spacing:.11em;text-transform:uppercase;margin-bottom:5px}
.state .hint{font-size:13.2px;color:rgba(255,255,255,.55);margin-bottom:19px;line-height:1.55}
.var{display:flex;justify-content:space-between;align-items:center;padding:11px 0;border-bottom:1px solid rgba(255,255,255,.12)}
.var:last-of-type{border-bottom:none}
.var .vk{font-size:14.2px;color:rgba(255,255,255,.78)}
.var .vv{font-size:21px;font-weight:760;font-variant-numeric:tabular-nums;transition:.3s}
.var .vv.up{color:var(--teal-2)}.var .vv.down{color:#FF8FA3}.var .vv.same{color:#fff}
.ctrl{display:flex;gap:9px;margin-top:20px}
.ctrl button{
flex:1;padding:10px;border-radius:10px;border:1px solid rgba(255,255,255,.22);background:rgba(255,255,255,.1);
color:#fff;font-family:inherit;font-size:14px;font-weight:640;cursor:pointer;transition:.18s;
}
.ctrl button:hover:not(:disabled){background:rgba(255,255,255,.2);border-color:rgba(255,255,255,.4)}
.ctrl button:disabled{opacity:.34;cursor:not-allowed}
.ctrl button.go{background:linear-gradient(100deg,var(--teal),var(--green));border-color:transparent}
.answer{margin-top:18px;padding:15px 18px;border-radius:11px;background:rgba(20,191,174,.14);border:1px solid rgba(20,191,174,.34);font-size:14.2px;color:#fff;display:none}
.answer.on{display:block;animation:fade .4s}
.answer b{color:var(--teal-2)}
@media(max-width:880px){.replay{grid-template-columns:1fr}.state{position:static}}
.log-head{display:flex;justify-content:space-between;align-items:center;margin-bottom:13px;gap:10px}
.log-tag{font-size:11.6px;font-weight:730;letter-spacing:.09em;text-transform:uppercase;color:var(--muted)}
.log-count{font-size:12.4px;color:var(--teal);font-weight:640;font-variant-numeric:tabular-nums}
.stat-tag{display:inline-block;font-size:11px;font-weight:700;letter-spacing:.04em;text-transform:uppercase;
color:var(--teal);background:rgba(14,158,151,.1);border-radius:5px;padding:1px 6px;margin-right:7px}
.rec.invalid .stat-tag{color:var(--rose);background:rgba(220,38,72,.09)}
.rec.lit{background:#fff;border-color:var(--teal);box-shadow:0 2px 10px rgba(14,158,151,.16)}
.qa-line{font-size:14px;color:var(--muted);margin-top:9px;padding-left:3px}
.qa-line b{color:var(--teal)}
.crew-h{font-size:11.4px;font-weight:730;letter-spacing:.09em;text-transform:uppercase;
color:rgba(255,255,255,.5);margin:19px 0 9px;padding-top:15px;border-top:1px solid rgba(255,255,255,.14)}
.crew{display:flex;justify-content:space-between;gap:10px;font-size:13px;padding:4px 0;color:rgba(255,255,255,.82)}
.crew .cv{color:var(--teal-2);font-variant-numeric:tabular-nums;text-align:right;font-size:12.4px}
.qbox{margin-top:20px;background:#fff;border:1px solid var(--line);border-radius:var(--radius);
padding:28px;box-shadow:var(--shadow);opacity:.5;transition:opacity .45s}
.qbox.ready{opacity:1}
.qbox-h h3{font-size:20px;font-weight:740;margin-bottom:5px}
.qbox-h p{font-size:14.8px;color:var(--muted);margin-bottom:19px}
.chips{display:flex;flex-wrap:wrap;gap:8px;margin-bottom:20px}
.chip{padding:7px 14px;border-radius:999px;border:1px solid var(--line);background:#fff;cursor:pointer;
font-family:inherit;font-size:13.6px;font-weight:620;color:var(--muted);transition:.16s}
.chip:hover{border-color:var(--c);color:var(--c)}
.chip.on{background:var(--c);border-color:var(--c);color:#fff}
.qa{border-left:3px solid var(--line-2);padding-left:18px}
.q-lab{display:inline-block;font-size:10.6px;font-weight:740;letter-spacing:.08em;text-transform:uppercase;
color:#fff;border-radius:5px;padding:2px 8px;margin-right:9px;vertical-align:2px}
.q-q{font-size:16px;color:var(--ink);margin-bottom:12px}
.q-a{font-size:15.6px;color:var(--body);margin-bottom:12px}
.q-r{font-size:13.4px;color:var(--muted)}
.q-r code{font-family:var(--mono);font-size:12.2px;background:var(--bg-3);border-radius:4px;padding:1px 5px;margin-right:3px}
/* ---------- charts ---------- */
.bars{display:flex;flex-direction:column;gap:11px}
.bar-row{display:grid;grid-template-columns:190px 1fr 56px;gap:14px;align-items:center;font-size:14.8px}
.bar-row .bl{color:var(--body);font-weight:560;text-align:right}
.bar-track{height:24px;background:var(--bg-3);border-radius:7px;overflow:hidden}
.bar-fill{height:100%;border-radius:7px;transform-origin:left center;animation:grow .95s cubic-bezier(.22,1,.36,1) both}
@keyframes grow{from{transform:scaleX(0)}to{transform:scaleX(1)}}
.bar-row .bv{font-weight:700;color:var(--navy);font-variant-numeric:tabular-nums;font-size:14.2px}
@media(max-width:700px){.bar-row{grid-template-columns:120px 1fr 48px;font-size:13.4px}}
/* ---------- tables ---------- */
.tbl-wrap{overflow-x:auto;border:1px solid var(--line);border-radius:var(--radius);background:#fff;box-shadow:var(--shadow)}
table{border-collapse:collapse;width:100%;font-size:14.6px;font-variant-numeric:tabular-nums}
th,td{padding:11px 14px;text-align:right;border-bottom:1px solid var(--line);white-space:nowrap}
th:first-child,td:first-child{text-align:left;font-weight:600;color:var(--ink)}
thead th{background:var(--bg-2);font-size:12.6px;font-weight:720;letter-spacing:.055em;text-transform:uppercase;color:var(--muted);position:sticky;top:0}
tbody tr{transition:.14s}
tbody tr:hover{background:var(--bg-2)}
tbody tr:last-child td{border-bottom:none}
td.best{color:var(--teal);font-weight:750}
.metric-head{min-width:190px}
thead .metric-role{
display:block;margin-top:2px;font-size:9.5px;font-weight:600;letter-spacing:.01em;
text-transform:none;color:var(--muted)
}
.score-cell{position:relative;min-width:190px;isolation:isolate}
.score-track{
position:absolute;z-index:0;left:12px;right:12px;top:50%;height:23px;
transform:translateY(-50%);overflow:hidden;border-radius:6px
}
.score-fill{height:100%;border-radius:6px;transition:width .35s ease}
.score-avg .score-track{background:rgba(14,158,151,.08)}
.score-avg .score-fill{background:linear-gradient(90deg,var(--teal),var(--green));opacity:.42}
.score-pass .score-track{background:rgba(12,42,87,.06)}
.score-pass .score-fill{background:var(--navy-3);opacity:.22}
.score-value{position:relative;z-index:1;display:block;font-weight:720;color:var(--navy)}
.score-pass .score-value{font-weight:620;color:var(--body)}
.score-cell.best .score-value{color:var(--teal);font-weight:800}
.gap-cell{color:var(--muted);font-size:13.4px}
/* ---------- families ---------- */
.fam{background:#fff;border:1px solid var(--line);border-radius:14px;padding:22px;transition:.22s}
.fam:hover{border-color:var(--teal);transform:translateY(-3px);box-shadow:var(--shadow-lg)}
.fam-top{display:flex;align-items:center;justify-content:space-between;margin-bottom:10px;gap:10px}
.fam h4{font-size:16.5px;font-weight:710}
.share{font-size:12.6px;font-weight:730;color:#fff;padding:3px 10px;border-radius:999px;flex-shrink:0}
.fam .q{font-size:14.6px;color:var(--body);font-style:italic;line-height:1.6;margin-top:9px;padding-top:12px;border-top:1px dashed var(--line-2)}
/* ---------- callout ---------- */
.callout{background:linear-gradient(135deg,var(--navy),#0A1F44);border-radius:20px;padding:44px;color:#fff;text-align:center;position:relative;overflow:hidden}
.callout::after{content:'';position:absolute;width:520px;height:520px;border-radius:50%;background:radial-gradient(circle,rgba(20,191,174,.28),transparent 68%);top:-260px;right:-150px}
.callout>*{position:relative;z-index:1}
.callout h3{color:#fff;font-size:26px;font-weight:760;margin-bottom:12px}
.callout p{color:rgba(255,255,255,.8);font-size:17px;max-width:40em;margin:0 auto}
/* ---------- footer ---------- */
footer{padding:56px 0 44px;background:var(--bg-2)}
.f-grid{display:grid;grid-template-columns:1.5fr 1fr 1fr;gap:34px;margin-bottom:34px}
footer h5{font-size:12.6px;font-weight:730;letter-spacing:.1em;text-transform:uppercase;color:var(--muted);margin:0 0 13px}
footer img{height:34px;margin-bottom:13px;mix-blend-mode:darken}
footer p{font-size:14.6px;color:var(--muted);margin:0 0 7px}
.f-auth{font-size:14.4px;color:var(--body);line-height:1.85}
.f-bot{border-top:1px solid var(--line-2);padding-top:22px;display:flex;justify-content:space-between;gap:16px;flex-wrap:wrap;font-size:13.8px;color:var(--muted)}
@media(max-width:820px){.f-grid{grid-template-columns:1fr}}
.js .reveal{opacity:0;transform:translateY(22px);transition:opacity .7s cubic-bezier(.22,1,.36,1),transform .7s cubic-bezier(.22,1,.36,1)}
.js .reveal.in{opacity:1;transform:none}
@media(prefers-reduced-motion:reduce){.js .reveal{opacity:1;transform:none}.bar-fill{animation:none}.hero::before,.hero::after{animation:none}}
.note{font-size:13.8px;color:var(--muted);margin-top:16px;font-style:italic}
/* ---------- vertical research landing page ---------- */
body{background:#F3F7FB}
nav,#problem,#approach,#anatomy,#data,#paper,footer{display:none}
.landing-shell{
width:min(100%,1600px);margin:0 auto;display:block;background:#fff;
box-shadow:0 0 0 1px var(--line),0 28px 90px rgba(12,42,87,.10)
}
.landing-shell .hero{
position:relative;min-height:auto;padding:clamp(44px,6vw,88px) clamp(24px,5vw,76px);
border:0;border-bottom:1px solid var(--line);background:#fff;overflow:hidden
}
.landing-shell .hero::before{width:620px;height:620px;top:-390px;right:-250px;opacity:.24}
.landing-shell .hero::after{width:620px;height:620px;bottom:-430px;left:-310px;opacity:.14}
.paper-hero-in{position:relative;z-index:1;max-width:1280px;margin:0 auto;text-align:center}
.logo-strip{
display:flex;align-items:center;justify-content:center;gap:clamp(26px,5vw,76px);
margin:0 auto 34px
}
.logo-strip img{display:block;max-width:100%;object-fit:contain}
.logo-strip .ruc-logo{width:clamp(230px,24vw,340px);max-height:70px}
.logo-strip .project-logo{
width:clamp(300px,31vw,440px);height:104px;object-fit:cover;object-position:center;
mix-blend-mode:darken
}
.logo-strip .hunyuan-logo{width:clamp(220px,21vw,300px);max-height:82px}
.paper-kicker{
display:inline-flex;align-items:center;gap:8px;margin-bottom:14px;padding:6px 14px;
border:1px solid var(--line-2);border-radius:999px;background:rgba(255,255,255,.82);
color:var(--teal);font-size:12px;font-weight:760;letter-spacing:.12em;text-transform:uppercase
}
.paper-title{
max-width:1120px;margin:0 auto 22px;color:var(--ink);
font-size:clamp(24px,2.6vw,38px);font-weight:820;line-height:1.22;letter-spacing:-.03em
}
.paper-title .paper-brand{
background:linear-gradient(100deg,var(--navy),var(--teal) 72%);
-webkit-background-clip:text;background-clip:text;-webkit-text-fill-color:transparent
}
.paper-authors{
max-width:1120px;margin:0 auto 12px;color:var(--navy);
font-size:clamp(15px,1.55vw,19px);font-weight:650;line-height:1.8
}
.paper-authors span{white-space:nowrap}
.paper-affiliations{
display:flex;justify-content:center;gap:8px 24px;flex-wrap:wrap;
max-width:1160px;margin:0 auto;color:var(--body);font-size:13.5px;line-height:1.55
}
.paper-notes{margin:8px 0 0;color:var(--muted);font-size:12.5px}
.abstract-card{
max-width:1120px;margin:34px auto 0;padding:27px 30px;text-align:left;
border:1px solid var(--line);border-left:4px solid var(--teal);border-radius:0 15px 15px 0;
background:rgba(247,251,253,.9);box-shadow:var(--shadow)
}
.abstract-label{
display:block;margin-bottom:8px;color:var(--teal);font-size:11.5px;
font-weight:780;letter-spacing:.13em;text-transform:uppercase
}
.abstract-card p{margin:0;color:var(--body);font-size:14.5px;line-height:1.68}
/* ---------- the two headline figures under the abstract ---------- */
.hero-figs{
display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:20px;
max-width:1120px;margin:22px auto 0;text-align:left
}
.hero-fig{
display:flex;flex-direction:column;margin:0;padding:16px 16px 14px;
border:1px solid var(--line);border-radius:14px;background:#fff;box-shadow:var(--shadow);
transition:.22s
}
.hero-fig:hover{border-color:var(--line-2);box-shadow:var(--shadow-lg);transform:translateY(-2px)}
.hero-fig a{display:block}
.hero-fig img{display:block;width:100%;height:auto;border-radius:7px}
.hero-fig figcaption{
margin-top:11px;color:var(--muted);font-size:12.8px;line-height:1.56
}
.hero-fig figcaption b{color:var(--navy);font-weight:700}
#replay{
min-width:0;max-width:1600px;margin:0 auto;padding:clamp(50px,6vw,88px) clamp(24px,5vw,76px);
border:0;background:linear-gradient(145deg,#F8FBFD 0%,#EEF5FA 100%)
}
#replay .wrap{max-width:1280px;margin:0 auto;padding:0}
#replay h2.sec{font-size:clamp(29px,3vw,43px);max-width:15em}
#replay .sec-lede{font-size:16.5px;max-width:54em;margin-bottom:26px}
#replay .replay{grid-template-columns:minmax(0,1fr) 260px;gap:16px}
#replay .log{padding:19px}
#replay .state{top:24px;padding:20px}
#replay .qbox{padding:22px;margin-top:16px}
#replay .rec{font-size:13.4px;padding:8px 10px}
#replay .sys{font-size:13.5px}
#results{
display:block;max-width:1600px;margin:0 auto;padding:82px 28px;
background:#fff;border-top:1px solid var(--line)
}
#results .wrap{max-width:1280px;padding:0}
#results .tbl-wrap table{min-width:840px}
@media(max-width:900px){
.logo-strip{display:grid;grid-template-columns:1fr 1fr;gap:16px 22px;margin-bottom:28px}
.logo-strip .project-logo{grid-column:1/-1;grid-row:1;width:min(100%,360px);height:86px;justify-self:center}
.logo-strip .ruc-logo{grid-column:1;grid-row:2;width:100%;max-height:54px}
.logo-strip .hunyuan-logo{grid-column:2;grid-row:2;width:100%;max-height:62px}
}
@media(max-width:680px){
.landing-shell .hero,#replay,#results{padding:38px 20px}
.paper-affiliations{display:block}
.paper-affiliations span{display:block;margin-bottom:5px}
.abstract-card{padding:22px 20px}
.abstract-card p{font-size:14px}
}
@media(max-width:820px){
.hero-figs{grid-template-columns:minmax(0,1fr)}
}
</style>
</head>
<body>
<script>document.documentElement.className='js';</script>
<!-- ================= HERO ================= -->
<main class="landing-shell">
<header class="hero" id="top">
<div class="paper-hero-in">
<div class="logo-strip" aria-label="EvolveScaler and institutional logos">
<img class="ruc-logo" src="assets/cn_ruc_logo.png" alt="Renmin University of China">
<img class="project-logo" src="assets/logo.png" alt="EvolveScaler">
<img class="hunyuan-logo" src="assets/hunyuanlogo.png" alt="Tencent Hunyuan">
</div>
<span class="paper-kicker">Research paper</span>
<h1 class="paper-title">
<span class="paper-brand">EvolveScaler</span>: Synthesizing Information-Evolution Contexts via Executable
State Machines and Natural Language Rendering
</h1>
<p class="paper-authors">
<span>Ziliang Zhao<sup>1,2,*</sup></span>,
<span>Zenan Xu<sup>2,*</sup></span>,
<span>Shuting Wang<sup>1</sup></span>,
<span>Zhao Wang<sup>1,2</sup></span>,
<span>Bowen Cao<sup>1,3</sup></span>,
<span>Minda Hu<sup>1,3</sup></span>,
<span>Lincheng Li<sup>2</sup></span>,
<span>Pluto Zhou<sup>2,†</sup></span>,
<span>Zhicheng Dou<sup>1,†</sup></span>
</p>
<div class="paper-affiliations">
<span><sup>1</sup> Gaoling School of Artificial Intelligence, Renmin University of China</span>
<span><sup>2</sup> LLM Department, Hunyuan Team, Tencent</span>
<span><sup>3</sup> The Chinese University of Hong Kong</span>
</div>
<p class="paper-notes"><sup>*</sup> Equal contribution · <sup>†</sup> Corresponding author</p>
<div class="abstract-card">
<span class="abstract-label">Abstract</span>
<p>
In persistent interactions, a long context can encode an evolving process rather than a fixed record of
facts. Later events can revise or revoke earlier records, changing which information remains valid and what
conclusions follow. We refer to this setting as <em>information evolution</em> (IE). Solving an IE task
requires identifying valid records, applying updates in order, and reconstructing the query-relevant state
from the event history. Constructing reliable IE data is difficult because conventional pipelines generate
long-form text before deriving supervision, leaving state transitions implicit and answers difficult to
verify. To address this problem, we introduce <b>EvolveScaler</b>, a code-driven framework that defines
information evolution in code before rendering it as natural language. Human-authored operational
specifications define how events alter state and which records remain valid. They also specify difficulty
controls and executable answer logic. A strong LLM synthesizes a self-contained simulator from each
specification. Executing a validated simulator produces natural-language, multi-turn event histories, while
deterministic replay computes reference answers and atomic checklists. The LLM proposes the executable
mechanism, while the replayed program state determines the supervision. We instantiate EvolveScaler with 117
task prototypes and 159 final-question operators. Five controlled difficulty levels jointly increase
trajectory scale and evolution complexity, spanning approximately 7 to 1,200 events per instance. The
resulting resource contains approximately 35,100 training examples and 585 validated evaluation instances.
Using the evaluation instances, we benchmark frontier and open models across the five difficulty levels. On
the <code>very_long</code> tier, the strongest model achieves an avg@5 of 59.3%, while six models have avg@5
scores below 10%. To test the training value of the generated data, we train an internal A3B model on
6,000 EvolveScaler examples. The trained model outperforms its base checkpoint on all eight independently
constructed out-of-distribution benchmarks, with an average gain of 5.25 points. These results show that
code-driven IE synthesis supports both diagnostic evaluation and transferable training supervision.
</p>
</div>
<div class="hero-figs">
<figure class="hero-fig">
<a href="assets/ie-examples.png" target="_blank" rel="noopener" title="Open full size">
<img src="assets/ie-examples.png"
alt="Four information-evolution scenarios: group dinner planning, a household expense ledger, soccer match play-by-play, and online-shopping returns">
</a>
<figcaption>
<b>Information evolution in four registers.</b> Creations, updates, invalidations, and
state-preserving noise are interleaved, so no single line carries the answer — it exists
only in the state you get by replaying the history.
</figcaption>
</figure>
<figure class="hero-fig">
<a href="assets/information-evolution.png" target="_blank" rel="noopener" title="Open full size">
<img src="assets/information-evolution.png"
alt="The EvolveScaler framework: a world spec becomes an executable state machine, rendered to a natural-language log, with answers from deterministic replay">
</a>
<figcaption>
<b>The EvolveScaler framework.</b> A world spec becomes an executable state machine whose
events keep rewriting a hidden state; the trace is rendered as a long multi-turn log, and
replaying the same machine yields the code-verified answer.
</figcaption>
</figure>
</div>
</div>
</header>
<!-- ================= PROBLEM ================= -->
<section id="problem">
<div class="wrap">
<div class="narrow">
<span class="eyebrow">The gap</span>
<h2 class="sec">Length is not the hard part. Keeping up is.</h2>
<p class="sec-lede">
When a model works with someone over days, the information it relies on keeps moving. A number is corrected.
A booking is cancelled. A screenshot turns out to be stale. We call this <b>Information Evolution</b> (IE):
the context is not a description of a fixed world, but a sequence of updates that can introduce, modify, or
revoke what came before — and quietly invalidate every conclusion downstream.
</p>
</div>
<div class="vs reveal">
<div class="pane bad">
<div class="tag">◇ Generate, then annotate</div>
<h3>The usual pipeline</h3>
<p style="font-size:15.2px;margin-top:9px">A model writes a long dialogue; humans or another model then read it back and try to infer the label.</p>
<ul>
<li>Consistency decays as the context grows.</li>
<li>The "gold" answer is itself a model's opinion.</li>
<li>Difficulty can be described, but not dialled.</li>
<li>Nothing guarantees the question is even answerable.</li>
</ul>
</div>
<div class="arrow">→</div>
<div class="pane good">
<div class="tag">◆ Specify, then render</div>
<h3>The EvolveScaler pipeline</h3>
<p style="font-size:15.2px;margin-top:9px">Information evolution is defined in code first; natural language is what the program prints.</p>
<ul>
<li>The state machine cannot contradict itself.</li>
<li>The answer is computed by replay, not judged.</li>
<li>Every difficulty knob is a parameter.</li>
<li>Answerability is checked before the sample ships.</li>
</ul>
</div>
</div>
<div class="grid g3 reveal" style="margin-top:22px">
<div class="card">
<div class="icon">⟳</div>
<h3>State, not text</h3>
<p>Difficulty is reframed as <b>state maintenance</b>. What makes a trace hard is the number of order-dependent updates a reader must survive — not its word count.</p>
</div>
<div class="card">
<div class="icon">⌦</div>
<h3>Nothing is given away</h3>
<p>Events expose only local deltas — "stock +2" — never the running total. No single sentence contains the answer, so the trace has to be reconstructed.</p>
</div>
<div class="card">
<div class="icon">✓</div>
<h3>Verifiable by construction</h3>
<p>Each sample ships with a code-derived reference answer and a checklist of atomic claims that can be checked one at a time.</p>
</div>
</div>
</div>
</section>
<!-- ================= APPROACH ================= -->
<section class="alt" id="approach">
<div class="wrap">
<span class="eyebrow">How it works</span>
<h2 class="sec">Three stages, one contract</h2>
<p class="sec-lede">
People decide what the world is. A language model writes the program that runs it. The program — not the
model — decides what the answer is. Click a stage to see what it guarantees.
</p>
<div class="pipe">
<button class="step on" data-step="0">
<div class="s-n">STAGE 01</div>
<h4>Specify</h4>
<div class="s-d">Annotators turn a real workflow into a typed transition specification.</div>
</button>
<button class="step" data-step="1">
<div class="s-n">STAGE 02</div>
<h4>Synthesize</h4>
<div class="s-d">An LLM compiles that specification into a self-contained simulator.</div>
</button>
<button class="step" data-step="2">
<div class="s-n">STAGE 03</div>
<h4>Execute & replay</h4>
<div class="s-d">Running it emits the dialogue; replaying it emits the answer.</div>
</button>
</div>
<div class="pipe-detail" id="pipeDetail"></div>
</div>
</section>
<!-- ================= ANATOMY ================= -->
<section id="anatomy">
<div class="wrap">
<span class="eyebrow">Anatomy of a prototype</span>
<h2 class="sec">Ten primitives make a world</h2>
<p class="sec-lede">
Every one of the 117 prototypes is described by the same ten primitives. Because they are declared rather
than narrated, each one becomes a dial. Pick a primitive to see what it fixes, illustrated on a
shared-expense chat for a group trip.
</p>
<div class="prim-wrap">
<div class="prim-list" id="primList"></div>
<div class="prim-detail" id="primDetail"></div>
</div>
</div>
</section>
<!-- ================= REPLAY ================= -->
<section class="alt" id="replay">
<div class="wrap">
<span class="eyebrow">Live information evolution</span>
<h2 class="sec">Language on the surface. Executable state underneath.</h2>
<p class="sec-lede">
Step through a warehouse log. The left side is the natural-language context shown to a model; the right side
is the hidden ledger maintained by code. Invalid records never touch the state. Finish the replay, then query
the same trace through seven different reasoning operators.
</p>
<div class="replay">
<div class="log">
<div class="log-head">
<span class="log-tag">Rendered context — what the model reads</span>
<span class="log-count" id="logCount"></span>
</div>
<div class="sys">
<b>System.</b> Track a warehouse shift board across receiving, put-away, picking and replenishment.
Only <b>WMS-posted</b>, <b>floor-re-checked</b>, <b>supervisor-confirmed</b> and
<b>stock-control-confirmed</b> records count. Drill tasks, retracted scans, old count drafts and test
waves do <b>not</b>. Running totals are never stated.
</div>
<div id="turns"></div>
</div>
<div class="state">
<h4>Hidden ledger</h4>
<div class="hint">Maintained by replaying the state machine. Never appears in the context.</div>
<div id="vars"></div>
<div class="crew-h">Per crew · valid records only</div>
<div id="crews"></div>
<div class="ctrl">
<button id="bPrev" disabled>← Back</button>
<button id="bNext" class="go">Start →</button>
</div>
</div>
</div>
<div class="qbox" id="qbox">
<div class="qbox-h">
<h3>Now ask the same trace anything</h3>
<p>Seven operator families, one unchanged log. Each answer is computed by replay, never judged by a model.</p>
</div>
<div class="chips" id="chips"></div>
<div class="qa" id="qa"></div>
</div>
</div>
</section>
</main>
<!-- ================= DATA ================= -->
<section id="data">
<div class="wrap">
<span class="eyebrow">The corpus</span>
<h2 class="sec">Broad on the surface, deep underneath</h2>
<p class="sec-lede">
117 prototypes across 12 themes, sampled at five length tiers, produce roughly 35,100 training and 585 held-out
test examples spanning two orders of magnitude in context length.
</p>
<div class="tabs" id="dataTabs">
<button class="tab on" data-p="themes">Themes</button>
<button class="tab" data-p="scale">Context scale</button>
<button class="tab" data-p="ops">Question operators</button>
</div>
<div class="panel on" data-panel="themes">
<div class="card" style="padding:30px">
<h3 style="margin-bottom:5px">Where information actually accumulates</h3>
<p style="color:var(--muted);font-size:15.2px;margin-bottom:22px">
Themes were chosen for process density — settings where records pile up, get revised and occasionally
get thrown out. Share of the 117 prototypes.
</p>
<div class="bars" id="barsThemes"></div>
</div>
</div>
<div class="panel" data-panel="scale">
<div class="card" style="padding:30px">
<h3 style="margin-bottom:5px">One prototype, five difficulties</h3>
<p style="color:var(--muted);font-size:15.2px;margin-bottom:22px">
The same world is re-run with more blocks, denser events, more invalid records and deeper counterfactuals.
Average state-changing events per sample, by tier.
</p>
<div class="bars" id="barsScale"></div>
<p class="note">From ~7 events in the shortest tier to ~1,200 in the longest — a two-order-of-magnitude sweep with the task held fixed.</p>
</div>
</div>
<div class="panel" data-panel="ops">
<p style="color:var(--muted);font-size:15.4px;margin:-10px 0 22px;max-width:46em">
The final question is a controlled probe of the reconstructed state. 159 operators fall into seven families;
each forces a different access pattern over the very same trace.
</p>
<div class="grid g3" id="famGrid"></div>
</div>
</div>
</section>
<!-- ================= RESULTS ================= -->
<section class="alt" id="results">
<div class="wrap">
<span class="eyebrow">Evaluation</span>
<h2 class="sec">Frontier models fall off a cliff</h2>
<p class="sec-lede">
14 systems, five tiers, five responses per instance. <b>avg@5 is our primary metric</b> for response-level
reliability; <b>pass@5 is auxiliary</b> and asks whether any of five attempts succeeds. Choose a tier to re-rank.
</p>
<div class="tabs" id="tierTabs">
<button class="tab" data-t="0">Very short</button>
<button class="tab" data-t="1">Short</button>
<button class="tab" data-t="2">Medium</button>
<button class="tab" data-t="3">Long</button>
<button class="tab on" data-t="4">Very long</button>
<button class="tab" data-t="5">Overall</button>
</div>
<div class="tbl-wrap">
<table>
<thead><tr>
<th>System</th>
<th class="metric-head">avg@5<span class="metric-role">Primary · response reliability</span></th>
<th class="metric-head">pass@5<span class="metric-role">Auxiliary · best of five</span></th>
<th>Gap<span class="metric-role">pass@5 − avg@5</span></th>
</tr></thead>
<tbody id="evalBody"></tbody>
</table>
</div>
<p class="note" id="tierNote"></p>
<div class="grid g2" style="margin-top:40px">
<div class="card">
<h3 style="margin-bottom:14px">Difficulty tracks state, not tokens</h3>
<div class="bars" id="barsMedian"></div>
<p style="font-size:14.8px;color:var(--muted);margin-top:16px">
Median avg@5 across all 14 systems. Separation widens with length: the spread is 59.0–81.7 on the
shortest tier and 4.8–59.3 on the longest — the benchmark stops being a formality and starts
being a measurement.
</p>
</div>
<div class="card">
<h3 style="margin-bottom:14px">Training on it transfers</h3>
<div class="bars" id="barsTrain"></div>
<p style="font-size:14.8px;color:var(--muted);margin-top:16px">
Average over eight held-out benchmarks that share no data with EvolveScaler. Continued training lifts
<b>all eight</b>, for <b>+5.25</b> points at 6,000 examples — state tracking transfers, rather than a
benchmark-specific answer format.
</p>
</div>
</div>
<div class="card reveal" style="margin-top:18px">
<h3 style="margin-bottom:6px">Where the answers go wrong</h3>
<p style="color:var(--muted);font-size:15.2px;margin-bottom:20px">
Each failure is charged to the first atomic claim it violates. Values are the unweighted means across the
ten systems reported in the paper.
</p>
<div class="bars" id="barsErr"></div>
<p class="note">
Net-change aggregation accounts for roughly half of labeled failures. Together with retrieval and counting,
record selection and signed-effect aggregation account for about two thirds. Notably, the strongest system has
the <i>lowest</i> aggregation-error share and the <i>highest</i> ranking share.
</p>
</div>
</div>
</section>
<!-- ================= PAPER ================= -->
<section id="paper" style="border-bottom:none">
<div class="wrap">
<div class="callout reveal">
<h3>EvolveScaler: Code-Driven Information-Evolving Data Synthesis</h3>
<p>
For training and evaluating large language models on long-context, multi-turn reasoning.
Full method, corpus statistics, evaluation of 14 systems, transfer results and error analysis
are described in the paper.
</p>
</div>
</div>
</section>
<footer>
<div class="wrap">
<div class="f-grid">
<div>
<p style="max-width:30em">
Code-driven synthesis of long, mutable, information-dense contexts with deterministically verifiable answers.
</p>
</div>
<div>
<h5>Authors</h5>
<div class="f-auth">
Ziliang Zhao<sup>1,2</sup>, Zenan Xu<sup>2</sup>,<br>
Shuting Wang<sup>1</sup>, Zhao Wang<sup>1,2</sup>,<br>
Bowen Cao<sup>1,3</sup>, Minda Hu<sup>1,3</sup>,<br>
Pluto Zhou<sup>2</sup>, Zhicheng Dou<sup>1</sup>
</div>
</div>
<div>
<h5>Affiliations</h5>
<p><sup>1</sup>Gaoling School of Artificial Intelligence,<br>Renmin University of China</p>
<p><sup>2</sup>LLM Department, Hunyuan Team, Tencent</p>
<p><sup>3</sup>The Chinese University of Hong Kong</p>
</div>
</div>
<div class="f-bot">
<span>© 2026 EvolveScaler. All rights reserved.</span>
<span>Figures and statistics on this page are drawn from the paper.</span>
</div>
</div>
</footer>
<script>
'use strict';
/* ---------------- data ---------------- */
const PIPE = [
{h:'Specify — a workflow becomes a typed contract',
b:`More than twenty annotators author prototypes from processes where information genuinely accumulates and
mutates: clinical follow-up, incident response, code review, returns, play-by-play. Each is written down as a
typed transition specification — entities, a clock, event schemas, state variables, invariants, validity rules,
noise channels, question operators and answer programs.`,
pts:['Informal common sense becomes explicit guards: a refund cannot exceed its payment; stock cannot go negative.',
'Every invalid record gets an exclusion rule instead of a judgement call.',
'Every question is required to map onto an executable answer function.']},
{h:'Synthesize — the model writes the mechanism, not the answer',
b:`The specification is handed to a strong LLM, which returns a self-contained simulator: state initialization,
event samplers and transitions, invariant checks, a natural-language renderer, and the question/checklist/answer
computation. Constraints are not prose the model is asked to remember — they are compiled into assertions
around every transition.`,
pts:['Accepted only if it runs clean across many random seeds.',
'Accepted only if deterministic replay reproduces the identical state.',
'Accepted only if its checklist agrees with its own reference answer.']},
{h:'Execute & replay — two outputs from one run',
b:`Executing the simulator renders a multi-turn dialogue: each user turn contributes a block of events and may ask
an intermediate question. Valid records are interleaved with invalidated, out-of-order and state-preserving
distractors, while the running state is withheld. Replaying the same machine yields the reference answer.`,
pts:['Linguistic variation cannot change answer semantics — the program is the ground truth.',
'Counterfactuals come free: drop a subset of events and replay.',
'Difficulty is a parameter sweep, not a rewrite.']}
];
const PRIMS = [
['Entities','typed actors and objects','Travellers (Ann, Bob, Carl, Dan); expense items (flights, hotel, car rental, meals, tickets); categories (transport, lodging, food).'],
['Clock / block','temporal granularity','A trip day or planning round — "Day 2", "round 5" — each carrying a batch of newly posted messages.'],
['Event schemas','admissible updates','Add or revise an expense; cancel a booking and issue a refund; add or remove a traveller; confirm a tentative quote.'],
['State variables','quantities maintained over time','Total trip cost; each traveller\'s paid amount, liability and balance; category-level spending; confirmed deposits and refundable amounts.'],
['Transition constraints','legal state transitions','A refund cannot exceed its payment; only current members share a new expense; allocation shares must sum to the expense; deposits cannot go negative.'],
['Validity rules','records that affect state','Confirmed bookings and receipts count. Cancelled items, fully refunded payments, tentative quotes, duplicated forwards and superseded screenshots do not.'],
['Noise channels','state-preserving distractors','Small talk, stickers, hypotheticals, aliases, duplicated messages, and receipts backfilled out of chronological order.'],
['Difficulty controls','scalable generation parameters','Number of members, blocks and events; invalid-record ratio; backfill distance; entity interleaving; counterfactual depth.'],
['Question operators','requested state views','One of 159 operators: running totals, Top-N spenders, windowed aggregation, invalid-record auditing, comparison, counterfactual removal.'],
['Answer program','executable supervision','Replay the valid events, apply the operator to the resulting ledger, and emit both the reference answer and its independently checkable checklist.']
];
const METRICS = [['inb','inbound units'],['out','outbound units'],['rel','relocated units'],
['pend','pending discrepancy'],['exc','exceptions']];
const CREWS = ['East Dock Receiving','West Rack Put-away','Night Picking','Cold-chain Replenishment'];
// short forms keep the ledger rows on one line; the log still names crews in full
const CREW_SHORT = ['East Dock','West Rack','Night Picking','Cold-chain'];
const TURNS = [
{n:'Wave 1', q:'Over all valid receipts, what is the change in inbound units?', a:'+165.', recs:[
{id:'W1-R1', s:'WMS posted', t:'08:10 — East Dock Receiving books in A1-03 bottled water; inbound +120.', c:0, k:'A1-03', d:{inb:120}},
{id:'W1-R2', s:'drill task', t:'08:18 — Night Picking ships B1-05 jasmine tea; outbound +80.', c:2, k:'B1-05', d:{out:80}, bad:'drill task'},
{id:'W1-R3', s:'floor re-checked', t:'08:25 — West Rack Put-away replenishes A2-11 tissue packs; inbound +45, relocated +45.', c:1, k:'A2-11', d:{inb:45, rel:45}},
{id:'W1-R4', s:'WMS posted', t:'08:33 — Night Picking picks A1-03 bottled water; outbound +60.', c:2, k:'A1-03', d:{out:60}},
]},
{n:'Wave 2', q:'Same question as before.', a:'+255.', recs:[
{id:'W2-R1', s:'supervisor confirmed', t:'09:02 — Cold-chain Replenishment books in D4-08 detergent; inbound +90.', c:3, k:'D4-08', d:{inb:90}},
{id:'W2-R2', s:'old count draft', t:'09:09 — East Dock Receiving logs a count gap on A1-03; pending discrepancy +12.', c:0, k:'A1-03', d:{pend:12}, bad:'old count draft'},
{id:'W2-R3', s:'WMS posted', t:'09:15 — East Dock Receiving finds a count gap on A2-11; pending +18, exceptions +1.', c:0, k:'A2-11', d:{pend:18, exc:1}},
{id:'W2-R4', s:'stock-control confirmed', t:'09:24 — Night Picking ships B1-05 jasmine tea; outbound +75.', c:2, k:'B1-05', d:{out:75}},
{id:'W2-R5', s:'floor re-checked', t:'09:31 — West Rack Put-away relocates A1-03; relocated +30.', c:1, k:'A1-03', d:{rel:30}},
]},
{n:'Wave 3', q:'Total valid exceptions so far?', a:'2.', recs:[
{id:'W3-R1', s:'WMS posted', t:'10:05 — East Dock Receiving clears the A2-11 gap; pending discrepancy −18.', c:0, k:'A2-11', d:{pend:-18}},
{id:'W3-R2', s:'retracted scan', t:'10:12 — Cold-chain Replenishment ships D4-08; outbound +40.', c:3, k:'D4-08', d:{out:40}, bad:'retracted scan'},
{id:'W3-R3', s:'supervisor confirmed', t:'10:20 — Night Picking recovers a mis-pick on A1-03; outbound −15, exceptions +1.', c:2, k:'A1-03', d:{out:-15, exc:1}},
{id:'W3-R4', s:'WMS posted', t:'10:28 — Cold-chain Replenishment ships D4-08 detergent; outbound +55.', c:3, k:'D4-08', d:{out:55}},
{id:'W3-R5', s:'floor re-checked', t:'10:36 — West Rack Put-away replenishes B1-05 jasmine tea; inbound +60, relocated +60.', c:1, k:'B1-05', d:{inb:60, rel:60}},
]},
];
const QUESTIONS = [
{f:'Ranking / Top-N', c:'#3462AC',
q:'Counting only valid records, rank the top-3 crews by inbound units.',
a:'<b>East Dock Receiving 120</b>, then West Rack Put-away 105, then Cold-chain Replenishment 90. Night Picking books nothing in.',
r:['W1-R1','W1-R3','W2-R1','W3-R5']},
{f:'Comparison & leader', c:'#0E9E97',
q:'Compare West Rack Put-away and Cold-chain Replenishment on inbound and outbound units.',
a:'They split it: West Rack leads inbound <b>105 vs 90</b>; Cold-chain leads outbound <b>55 vs 0</b>. Neither leads both.',
r:['W1-R3','W2-R1','W2-R5','W3-R4','W3-R5']},
{f:'Profile & summary', c:'#8B55B4',
q:'Summarize Night Picking’s valid trajectory, and give its latest valid record.',
a:'Outbound <b>120</b>, exceptions <b>1</b>, every other metric 0. Latest valid record: <b>W3-R3</b>. Its wave-1 shipment was a drill and never happened.',
r:['W1-R2','W1-R4','W2-R4','W3-R3']},
{f:'Window & range', c:'#E09B21',
q:'Over waves 2–3 only, what is the net change in outbound units and pending discrepancy?',
a:'Outbound <b>+115</b> (+75 −15 +55). Pending discrepancy nets to <b>0</b> — raised by +18 in wave 2, cleared by −18 in wave 3.',
r:['W2-R3','W2-R4','W3-R1','W3-R3','W3-R4']},
{f:'Audit & exclusion', c:'#3DB54A',
q:'Which records are invalid? After excluding them, what is the total outbound?',
a:'Invalid: <b>W1-R2</b> (drill), <b>W2-R2</b> (old draft), <b>W3-R2</b> (retracted). Excluding them, outbound totals <b>175</b> — the three invalid records would have inflated it by 120.',
r:['W1-R2','W2-R2','W3-R2']},
{f:'Counterfactual', c:'#C2547E',
q:'Dropping all valid records of Night Picking, what is the remaining outbound?',
a:'<b>55</b>. Night Picking contributed +60 +75 −15 = 120 of the 175, so only W3-R4 survives.',
r:['W1-R4','W2-R4','W3-R3','W3-R4']},
{f:'Hotspot & peak', c:'#5A6B84',
q:'Which SKU-location carries the most valid records, and what is its inbound?',
a:'<b>A1-03 bottled water</b>, with 4 valid records — one more than A2-11. Its inbound is <b>120</b>, all from a single receipt.',
r:['W1-R1','W1-R4','W2-R5','W3-R3']},
];
const THEMES = [['Professional & tech',32],['Business & finance',21],['Public & education',20],['Games & sports',15],['Everyday & travel',12]];
const SCALE = [['Very short',7],['Short',27],['Medium',113],['Long',456],['Very long',1209]];
const FAMS = [
['Ranking / Top-N',21,'#3462AC','Counting only confirmed records, rank the top-3 accounts by spend — and give the values.'],
['Comparison & leader',18,'#0E9E97','Compare accounts A and B on remaining budget and spend; which has more left?'],
['Hotspot & peak',16,'#3DB54A','Which project has the most valid "budget top-up" records, and its net spend change?'],
['Profile & summary',14,'#8B55B4','Summarize actor X\'s valid trajectory: current metrics and the latest valid record ID.'],
['Window & range',12,'#E09B21','Over batches 1–5 only, what is the net change in pending invoices and remaining budget?'],
['Audit & exclusion',11,'#C2547E','Which records are invalid? After excluding them, what is the valid alert count?'],
['Counterfactual',9,'#5A6B84','Dropping all valid records of account X, what is the remaining valid invoice total?']
];
const MODELS = [
['GPT-5.5-xhigh',[[83.8,78.5],[83.8,76.4],[79.5,76.6],[85.5,73.2],[81.2,59.3],[82.7,72.8]]],
['GPT-5.5-high',[[87.2,81.7],[82.9,75.2],[79.5,73.8],[83.8,69.7],[75.2,47.5],[81.7,69.6]]],
['GPT-5.5-medium',[[83.8,77.9],[87.2,75.7],[79.5,74.2],[77.8,62.7],[59.8,37.9],[77.6,65.7]]],
['Gemini-3.1-Pro',[[84.6,74.5],[81.2,67.5],[75.2,59.8],[61.5,42.2],[47.0,27.5],[69.9,54.3]]],
['Hy-3-high',[[84.6,72.1],[88.0,74.4],[78.6,63.6],[73.5,39.7],[34.2,19.8],[71.8,53.9]]],
['DeepSeek-V4-Preview-Pro',[[83.8,66.2],[82.9,62.2],[74.4,52.5],[67.5,40.0],[56.4,29.7],[73.0,50.1]]],
['GLM-5.2',[[86.3,76.6],[82.9,70.6],[75.2,48.7],[45.3,21.5],[26.5,12.5],[63.2,46.0]]],
['Qwen3.5-Plus-Thinking',[[80.3,62.7],[75.2,64.8],[72.6,51.6],[35.0,22.2],[16.2,10.1],[55.9,42.3]]],
['GLM-5.1',[[86.3,74.9],[83.8,68.0],[64.1,38.8],[25.6,13.7],[16.2,9.6],[55.2,41.0]]],
['Hy3-Preview-high',[[86.3,70.3],[80.3,70.4],[68.4,42.2],[26.5,14.4],[10.3,5.1],[54.4,40.5]]],
['Doubao-2.0-Pro-high',[[71.8,59.1],[71.8,63.4],[65.0,49.4],[28.2,16.8],[12.8,9.1],[49.9,39.6]]],
['Doubao-1.8-high',[[76.1,59.0],[75.2,65.6],[52.1,35.0],[17.1,11.5],[12.8,8.4],[46.7,35.9]]],
['Doubao-2.0-Pro-medium',[[75.2,60.3],[71.8,60.7],[37.6,27.5],[16.2,9.4],[9.4,4.8],[42.1,32.5]]],
['Doubao-1.8-medium',[[73.5,60.0],[71.8,61.5],[35.9,21.2],[12.8,7.5],[11.1,6.2],[41.0,31.3]]]
];
const TIER_NOTE = [
'Everyone looks competent here: avg@5 spans 59.0–81.7. A short trace does not separate systems.',
'Still crowded at the top — the second tier barely moves the ranking.',
'The field begins to split: the strongest hold near 75 while the weakest have already halved.',
'The cliff. Half the field is under 20 avg@5 while the leaders stay above 60.',
'Six of 14 systems score below 10. The spread is now 4.8–59.3 — a 12× gap on the same task family.',
'Averaged over all five tiers. Even the strongest system fails more than a quarter of sampled responses.'
];
const MEDIAN = [['Very short',71.2],['Short',67.8],['Medium',50.5],['Long',21.9],['Very long',11.3]];
const TRAIN = [['Base model',33.17],['+ 200 examples',35.60],['+ 2,000 examples',37.45],['+ 6,000 examples',38.42]];
const ERRS = [['Net-change aggregation',51.7,'#3462AC'],['Ranking & tie-breaking',21.5,'#0E9E97'],['Retrieval & counting',18.5,'#3DB54A'],['Comparison verdict',4.0,'#E09B21'],['Invalid-record handling',4.3,'#8B55B4']];
/* ---------------- helpers ---------------- */
const $ = s => document.querySelector(s);
const $$ = s => [...document.querySelectorAll(s)];
function bars(host, rows, opt){
opt = opt || {};
const max = opt.max || Math.max(...rows.map(r => r[1]));
host.innerHTML = rows.map(r => {
const w = (opt.log ? Math.log10(r[1]+1)/Math.log10(max+1) : r[1]/max) * 100;
const c = r[2] || 'linear-gradient(90deg,var(--navy),var(--teal))';
const col = c.startsWith('#') ? c : c;
return `<div class="bar-row"><div class="bl">${r[0]}</div>
<div class="bar-track"><div class="bar-fill" style="width:${w.toFixed(1)}%;background:${col}"></div></div>