forked from buh07/RL-Decoder-using-SAE-Features
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathpresentation.html
More file actions
952 lines (897 loc) · 36.9 KB
/
Copy pathpresentation.html
File metadata and controls
952 lines (897 loc) · 36.9 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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>RL-Decoder with SAE Features - Full Research Story</title>
<style>
:root {
--bg: #f3f7f8;
--panel: #ffffff;
--ink: #0f1f26;
--muted: #4a606a;
--line: #d4dfe4;
--accent: #0f766e;
--accent-2: #155e75;
--soft: #e8f5f3;
--ok: #166534;
--warn: #a16207;
--bad: #b91c1c;
--shadow: 0 10px 28px rgba(12, 27, 33, 0.10);
}
* { box-sizing: border-box; }
body {
margin: 0;
background:
radial-gradient(circle at 8% 10%, #e5f5f2 0%, transparent 35%),
radial-gradient(circle at 92% 5%, #e8edf7 0%, transparent 34%),
var(--bg);
color: var(--ink);
font-family: "IBM Plex Sans", "Segoe UI", "Helvetica Neue", Arial, sans-serif;
line-height: 1.5;
}
.deck {
max-width: 1180px;
margin: 0 auto;
padding: 16px 20px 80px;
}
.topnav {
position: sticky;
top: 10px;
z-index: 5;
margin: 0 0 14px;
border: 1px solid var(--line);
border-radius: 999px;
background: rgba(255, 255, 255, 0.92);
backdrop-filter: blur(6px);
box-shadow: var(--shadow);
padding: 8px 10px;
display: flex;
gap: 8px;
flex-wrap: wrap;
}
.topnav a {
text-decoration: none;
color: #134455;
font-size: 0.86rem;
border: 1px solid #d7e3e7;
background: #f8fcfd;
border-radius: 999px;
padding: 5px 10px;
white-space: nowrap;
}
.topnav a:hover {
border-color: #9cc9cf;
background: #eef8f7;
}
.slide {
min-height: 86vh;
background: var(--panel);
border: 1px solid var(--line);
border-radius: 18px;
padding: 22px;
margin: 0 0 16px;
box-shadow: var(--shadow);
display: flex;
flex-direction: column;
gap: 12px;
}
.hero {
background: linear-gradient(132deg, #0a5f59, #0b3f52);
color: #f2fffd;
border-color: transparent;
}
h1 {
margin: 0;
font-size: clamp(1.6rem, 2.4vw, 2.5rem);
letter-spacing: 0.01em;
}
h2 {
margin: 0;
font-size: 1.38rem;
color: #0f3241;
}
.hero h2,
.hero h1 {
color: #ffffff;
text-shadow: 0 1px 2px rgba(0, 0, 0, 0.28);
}
h3 {
margin: 0 0 6px;
font-size: 1.02rem;
color: #113545;
}
.hero p {
margin: 6px 0;
color: #f0fffc;
max-width: 980px;
}
.meta {
margin-top: 6px;
color: #eafff9;
font-size: 0.92rem;
display: grid;
gap: 4px;
}
.hero .meta strong {
color: #ffffff;
}
.hero code {
background: rgba(6, 23, 30, 0.55);
border-color: rgba(216, 245, 252, 0.42);
color: #f2fffd;
}
.grid {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(250px, 1fr));
gap: 12px;
}
.card {
border: 1px solid var(--line);
border-radius: 12px;
background: #fbfdff;
padding: 12px;
}
.card strong {
color: #0f3241;
}
.template-grid {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(240px, 1fr));
gap: 12px;
}
.pill {
display: inline-block;
font-size: 0.78rem;
font-weight: 700;
letter-spacing: 0.02em;
border-radius: 999px;
border: 1px solid transparent;
padding: 3px 8px;
margin-bottom: 6px;
text-transform: uppercase;
}
.ok { color: var(--ok); background: #e7f7eb; border-color: #b7e2c0; }
.warn { color: var(--warn); background: #fff5e6; border-color: #f2d5a6; }
.bad { color: var(--bad); background: #ffeded; border-color: #f2b8b8; }
.callout {
border-left: 4px solid var(--accent);
border-radius: 10px;
background: var(--soft);
color: #0e413b;
padding: 10px 12px;
font-size: 0.95rem;
}
.tbl {
width: 100%;
border-collapse: collapse;
font-size: 0.93rem;
}
.tbl th,
.tbl td {
border-bottom: 1px solid var(--line);
padding: 8px 6px;
text-align: left;
vertical-align: top;
}
.tbl th {
background: #f7fbfc;
color: #1f4a59;
font-weight: 700;
}
ul {
margin: 6px 0 0 18px;
padding: 0;
}
li { margin: 3px 0; }
code {
font-family: "IBM Plex Mono", "SFMono-Regular", Consolas, "Liberation Mono", Menlo, monospace;
font-size: 0.85em;
background: #f2f6f7;
border: 1px solid #d7e1e4;
border-radius: 6px;
padding: 1px 5px;
word-break: break-word;
}
.small { color: var(--muted); font-size: 0.9rem; }
@media (max-width: 800px) {
.slide { min-height: auto; }
.topnav { position: static; }
}
</style>
</head>
<body>
<main class="deck">
<nav class="topnav" aria-label="Presentation sections">
<a href="#map">Map</a>
<a href="#decision-chain">Decision Chain</a>
<a href="#phase1">P1</a>
<a href="#phase2">P2/P2r</a>
<a href="#phase3">P3</a>
<a href="#phase4">P4/P4r</a>
<a href="#phase5">P5</a>
<a href="#phase6">P6</a>
<a href="#phase7">P7 GPT-2</a>
<a href="#phase7-program">P7 Program</a>
<a href="#phase7-results">P7 Results</a>
<a href="#phase7-sae-design">P7 SAE Design</a>
<a href="#phase7-sae-results">P7 SAE Results</a>
<a href="#phase7-relevance">P7 Relevance</a>
<a href="#contrib">Contributions</a>
<a href="#qwen">Qwen Inquiry</a>
</nav>
<section class="slide hero" id="intro">
<h1>RL-Decoder with SAE Features</h1>
<h2>Complete Research Story: What We Tried, What Worked, and What Did Not</h2>
<p>
Main question: when a model writes step-by-step reasoning (CoT), is that text actually describing how the model got the answer?
</p>
<p>
We built a full test pipeline. It starts with basic checks, then measures internal arithmetic features, then tests whether those features can detect faithful vs unfaithful reasoning text.
</p>
<div class="meta">
<div><strong>Plain definitions:</strong> <code>CoT</code> = Chain-of-Thought text. <code>SAE</code> = sparse autoencoder used to turn hidden states into interpretable sparse features. <code>AUROC</code> = ranking quality score (0.5 random, 1.0 perfect).</div>
<div><strong>Snapshot:</strong> as of March 7, 2026.</div>
<div><strong>Project state:</strong> GPT-2 Phase 7 is closed. Track C is a negative result under this protocol. Qwen is now a separate next-model test.</div>
<div><strong>Final GPT-2 Phase 7 config:</strong> <code>text=0.50</code>, <code>latent=0.50</code>, <code>confidence=0.0</code>, <code>causal=0.0</code>, with structural penalties enabled.</div>
</div>
</section>
<section class="slide" id="map">
<h2>Story Map (End-to-End)</h2>
<table class="tbl">
<thead>
<tr>
<th>Stage</th>
<th>Question</th>
<th>What was learned</th>
</tr>
</thead>
<tbody>
<tr>
<td>Phases 1-3</td>
<td>Can we trust the basic tools?</td>
<td>Yes. SAE training and tracing checks worked on known tasks, so later results are not built on broken tooling.</td>
</tr>
<tr>
<td>Phases 4-5</td>
<td>Are arithmetic features real, and can we control the model with them?</td>
<td>The features are real and causally active for arithmetic outputs, but simple steering with those features is not stable.</td>
</tr>
<tr>
<td>Phase 6</td>
<td>Can we decode useful internal state across many layer setups?</td>
<td>Yes. Full sweep completed, no failed configs, and we selected reliable checkpoints for Phase 7.</td>
</tr>
<tr>
<td>Phase 7 (GPT-2)</td>
<td>Do internal causal signals verify CoT faithfulness?</td>
<td>No in GPT-2 under this protocol. Track C closed negative. Final GPT-2 system is a two-track detector.</td>
</tr>
<tr>
<td>Next</td>
<td>Is this only a GPT-2 issue?</td>
<td>Still open. Qwen is treated as a new, gated hypothesis test rather than a continuation of GPT-2 closure.</td>
</tr>
</tbody>
</table>
<div class="callout">
Short version: we proved the model has real internal arithmetic features, but those features did not become a robust faithfulness verifier for GPT-2 reasoning text.
</div>
</section>
<section class="slide" id="decision-chain">
<h2>Decision Chain (Result -> Next Step)</h2>
<table class="tbl">
<thead>
<tr>
<th>Phase result</th>
<th>Constraint / insight</th>
<th>Next phase design choice</th>
</tr>
</thead>
<tbody>
<tr>
<td>P1 passed on tasks with known hidden state truth.</td>
<td>The core representation method is reliable.</td>
<td>Scale up to full model layer capture and SAE training (P2).</td>
</tr>
<tr>
<td>P2/P2r produced stable sparse features across layers.</td>
<td>Sparsity must be controlled explicitly to avoid unstable probes.</td>
<td>Use TopK SAEs for arithmetic probing and causal tests (P4/P4r).</td>
</tr>
<tr>
<td>P3 found broad activity (~50% active), but weak class separation.</td>
<td>Simple "count active features" is too coarse.</td>
<td>Switch to targeted arithmetic state probes and subspace tests (P4/P4r).</td>
</tr>
<tr>
<td>P4/P4r showed strong arithmetic decoding and positive interventions.</td>
<td>The model has real arithmetic internals, but signal may be distributed.</td>
<td>Stress test controllability (P5), then build auditor pipeline (P6/P7).</td>
</tr>
<tr>
<td>P5 steering failed across layers.</td>
<td>Features that predict behavior are not always good control levers.</td>
<td>Treat features as evidence channels, not direct steering controls (P6/P7).</td>
</tr>
<tr>
<td>P6 completed full checkpoint sweep with strong decoder baselines.</td>
<td>We now had enough quality to run strict faithfulness tests.</td>
<td>Run Phase 7 Track A/B/C with leakage guards and fixed gates.</td>
</tr>
<tr>
<td>P7 and follow-up stress tests failed to make Track C robust.</td>
<td>In GPT-2, arithmetic internals did not cleanly track CoT faithfulness labels.</td>
<td>Close GPT-2 with two tracks. Start Qwen as a new gated inquiry.</td>
</tr>
</tbody>
</table>
</section>
<section class="slide" id="phase1">
<h2>Phase 1: Ground-Truth Validation (Tooling Sanity Check)</h2>
<div class="template-grid">
<article class="card">
<span class="pill ok">Intuition</span>
<p>Before touching LLM reasoning, first prove SAE can recover hidden variables when we already know the correct answer.</p>
</article>
<article class="card">
<span class="pill ok">Method</span>
<ul>
<li>Used controlled environments: BFS traversal, stack-machine, and logic puzzle traces.</li>
<li>Each environment has known internal state labels.</li>
<li>Trained SAEs and checked whether decoded features recover those labels.</li>
</ul>
</article>
<article class="card">
<span class="pill ok">Key Results</span>
<ul>
<li>R^2 range across tasks: <code>0.967-0.999</code>.</li>
<li>All tasks passed strong reconstruction and decoding checks.</li>
<li>No major instability across seeds.</li>
</ul>
</article>
<article class="card">
<span class="pill ok">Impact On Next Phase</span>
<p>Because this passed, we had evidence the method itself works. That justified moving to full GPT-2 layer-wide SAE training.</p>
</article>
</div>
</section>
<section class="slide" id="phase2">
<h2>Phase 2 / 2r: Multi-Layer SAE Training and TopK Stabilization</h2>
<div class="template-grid">
<article class="card">
<span class="pill ok">Intuition</span>
<p>To understand reasoning internals, we need features from every layer, not only one or two layers.</p>
</article>
<article class="card">
<span class="pill ok">Method</span>
<ul>
<li>Phase 2: trained SAEs across all GPT-2 layers.</li>
<li>Phase 2r: retrained with TopK activation so sparsity is controlled directly.</li>
<li>Saved per-layer stats and checkpoints for downstream probes and interventions.</li>
</ul>
</article>
<article class="card">
<span class="pill ok">Key Results</span>
<ul>
<li><code>24</code> GPT-2 SAEs at <code>12x</code> expansion.</li>
<li><code>98</code> SAEs across 4 models in broader assets.</li>
<li>TopK setup stabilized around ~<code>30%</code> active feature rate for GPT-2.</li>
<li>Training quality was stable enough to standardize later experiments.</li>
</ul>
</article>
<article class="card">
<span class="pill ok">Impact On Next Phase</span>
<p>This gave us a stable feature space for tracing where model information appears across tokens and layers.</p>
</article>
</div>
</section>
<section class="slide" id="phase3">
<h2>Phase 3: Reasoning Flow Tracing</h2>
<div class="template-grid">
<article class="card">
<span class="pill ok">Intuition</span>
<p>Before designing probes, we mapped where feature activity happens during reasoning text.</p>
</article>
<article class="card">
<span class="pill ok">Method</span>
<ul>
<li>Captured hidden states from every GPT-2 block during reasoning generation.</li>
<li>Encoded token states with per-layer SAEs.</li>
<li>Built token-layer activity maps for equation, intermediate, and non-computation tokens.</li>
</ul>
</article>
<article class="card">
<span class="pill ok">Key Results</span>
<ul>
<li>Average occupancy roughly <code>~50% active / ~50% sparse</code>.</li>
<li>Simple active-count differences between useful and non-useful tokens were small.</li>
<li>This showed we needed targeted features, not coarse counts.</li>
</ul>
</article>
<article class="card">
<span class="pill ok">Impact On Next Phase</span>
<p>We moved from coarse counting to targeted arithmetic probes tied to equation positions and step roles.</p>
</article>
</div>
</section>
<section class="slide" id="phase4">
<h2>Phase 4 / 4r: Arithmetic Probing and Causal Subspace Tests</h2>
<div class="template-grid">
<article class="card">
<span class="pill ok">Intuition</span>
<p>If arithmetic is represented internally, we should be able to decode arithmetic state and causally perturb that state.</p>
</article>
<article class="card">
<span class="pill ok">Method</span>
<ul>
<li>Used GSM8K traces with arithmetic annotations at equation and result tokens.</li>
<li>Trained ridge probes to predict arithmetic targets from SAE features.</li>
<li>Ran causal subspace patching (Phase 4r) to test whether those features affect output probability.</li>
</ul>
</article>
<article class="card">
<span class="pill ok">Key Results</span>
<ul>
<li>Probe performance: around <code>R^2=0.977</code> at layer 7 (TopK stack).</li>
<li>Causal subspace result: peak <code>+0.107 delta_logprob</code> at layer 22.</li>
<li><code>21/24</code> layers showed positive directional effect in 4r causal profile.</li>
<li>Plain meaning: arithmetic features are present and can matter causally.</li>
</ul>
</article>
<article class="card">
<span class="pill ok">Impact On Next Phase</span>
<p>This justified building a faithfulness auditor, but first we had to test whether these features are controllable in practice.</p>
</article>
</div>
</section>
<section class="slide" id="phase5">
<h2>Phase 5: Feature Interpretation and Steering Stress Test</h2>
<div class="template-grid">
<article class="card">
<span class="pill ok">Intuition</span>
<p>A feature can predict behavior but still fail as a stable control handle. We tested that directly.</p>
</article>
<article class="card">
<span class="pill ok">Method</span>
<ul>
<li>Generated top-feature context cards for interpretability.</li>
<li>Applied mean-difference steering vectors across layers and measured output shifts.</li>
<li>Compared expected direction vs observed direction.</li>
</ul>
</article>
<article class="card">
<span class="pill warn">Key Results</span>
<ul>
<li>Interpretation showed broad high-activation arithmetic-associated features.</li>
<li>Mean-diff steering produced uniformly negative shifts (disruptive, not useful control).</li>
<li>Control behavior was unstable even when probes looked strong.</li>
</ul>
</article>
<article class="card">
<span class="pill warn">Impact On Next Phase</span>
<p>We stopped assuming "good probe = good control." In later phases we used these channels as evidence signals, not guaranteed interventions.</p>
</article>
</div>
</section>
<section class="slide" id="phase6">
<h2>Phase 6: Decoder Benchmark + Full Layer Sweep</h2>
<div class="template-grid">
<article class="card">
<span class="pill ok">Intuition</span>
<p>Before Phase 7 claims, we needed strong decoder baselines and robust checkpoint coverage.</p>
</article>
<article class="card">
<span class="pill ok">Method</span>
<ul>
<li>Ran full supervised sweep over input modes (raw/hybrid/sae) and layer windows.</li>
<li>Used fixed split policy and leakage guards.</li>
<li>Selected parity checkpoints for the final faithfulness matrix.</li>
</ul>
</article>
<article class="card">
<span class="pill ok">Key Results</span>
<ul>
<li>Sweep completion: <code>102/102</code> configs, <code>0</code> failed.</li>
<li>Best top-1: <code>raw_block8_00_07 = 0.5628</code>.</li>
<li>Best top-5: <code>hybrid_block4_04_07 = 0.7860</code>.</li>
<li>This gave a stable base for fair Phase 7 comparisons.</li>
</ul>
</article>
<article class="card">
<span class="pill ok">Impact On Next Phase</span>
<p>We entered Phase 7 with reliable checkpoints, standardized artifacts, and strong leakage controls.</p>
</article>
</div>
</section>
<section class="slide" id="phase7">
<h2>Phase 7 (GPT-2): Main Question, Rules, and Final Outcome</h2>
<div class="grid">
<article class="card">
<h3>Main question</h3>
<p>
Can internal model signals tell us whether a reasoning trace is faithful, not just whether the text looks plausible?
</p>
<ul>
<li><strong>Track A:</strong> text pattern signal.</li>
<li><strong>Track B:</strong> decoded latent-state signal.</li>
<li><strong>Track C:</strong> causal/mechanistic signal.</li>
</ul>
</article>
<article class="card">
<h3>Rules we enforced</h3>
<ul>
<li>Leakage-safe trace-group split with fixed seed.</li>
<li>Calibration and evaluation kept separate.</li>
<li>Variant-robust checks on hard cases.</li>
<li>Bootstrap CIs and grouped cross-validation before closure.</li>
</ul>
</article>
<article class="card">
<h3>Final GPT-2 deployment</h3>
<ul>
<li><code>text=0.50</code>, <code>latent=0.50</code></li>
<li><code>confidence=0.0</code>, <code>causal=0.0</code></li>
<li>Structural penalties enabled.</li>
<li>Track C closed as a negative result for GPT-2.</li>
</ul>
</article>
</div>
<div class="callout">
Bottom line for this phase: we ran many corrected versions of Track C, but none passed robust closure criteria on GPT-2.
</div>
</section>
<section class="slide" id="phase7-program">
<h2>Phase 7 Progression: Step-by-Step With Rationale</h2>
<table class="tbl">
<thead>
<tr>
<th>Step</th>
<th>Why we ran it</th>
<th>Result</th>
<th>Decision after result</th>
</tr>
</thead>
<tbody>
<tr>
<td>1) Core matrix (<code>4 checkpoints x 4 variables</code>)</td>
<td>Start with the full baseline test under one fixed protocol.</td>
<td>Track C did not show robust pass behavior.</td>
<td>Do not conclude yet. First fix known engineering risks.</td>
</tr>
<tr>
<td>2) Causal bug-fix reruns</td>
<td>Make sure failures are not caused by indexing, lookup, or token-anchor bugs.</td>
<td>Code path improved and diagnostics became cleaner, but core Track C weakness stayed.</td>
<td>Treat this as scientific signal, then run targeted reformulations.</td>
</tr>
<tr>
<td>3) R1 confidence-margin track</td>
<td>Test whether decoder confidence gives a better faithfulness signal than hard matches.</td>
<td>Some signal, but not strong enough for closure.</td>
<td>Keep as diagnostic channel, continue to stronger tests.</td>
</tr>
<tr>
<td>4) R2 trajectory-coherence track</td>
<td>Test whether step-to-step consistency can separate faithful vs unfaithful traces.</td>
<td>Direction was wrong (anti-predictive).</td>
<td>Reject this formulation for GPT-2.</td>
</tr>
<tr>
<td>5) R3 contrastive probe</td>
<td>Test if a learned classifier over hidden-derived features can recover Track C signal.</td>
<td>Early run looked promising, but robust follow-up was not stable enough.</td>
<td>Proceed to stricter fairness and robustness tests.</td>
</tr>
<tr>
<td>6) Option B fair profile comparison</td>
<td>Verify gains are real by re-auditing each profile, not just blending scores.</td>
<td>P1/P2/P3 were very close; required ordering test failed.</td>
<td>No strong evidence that added Track C weighting is helping.</td>
</tr>
<tr>
<td>7) Path C robust validation</td>
<td>Stress test on hard variant with train exclusions + grouped CV.</td>
<td>Single split near threshold, pooled CV below target.</td>
<td>Treat as not robust.</td>
</tr>
<tr>
<td>8) Mixed hidden + SAE ladder</td>
<td>Check if combining feature spaces adds independent signal.</td>
<td>No stable incremental gain beyond SAE block.</td>
<td>Close GPT-2 Track C negative under current protocol.</td>
</tr>
</tbody>
</table>
<div class="callout">
This progression matters: each new step was chosen to answer a specific failure mode from the previous step.
</div>
</section>
<section class="slide" id="phase7-results">
<h2>Phase 7 Results: Clear Readout</h2>
<table class="tbl">
<thead>
<tr>
<th>Test</th>
<th>Main numbers</th>
<th>Plain-English meaning</th>
</tr>
</thead>
<tbody>
<tr>
<td>R1 confidence track</td>
<td>AUROC <code>0.6009</code>, CI <code>[0.5487, 0.6489]</code>, corr with text <code>0.6187</code></td>
<td>There is some useful signal, but not enough to make Track C robust by itself.</td>
</tr>
<tr>
<td>R2 trajectory track</td>
<td>AUROC <code>0.4497</code>, CI <code>[0.4090, 0.4929]</code></td>
<td>This version points in the wrong direction, so it is not usable.</td>
</tr>
<tr>
<td>R3 contrastive probe (early)</td>
<td>Control-conditioned <code>0.7136</code> vs source-trace <code>0.4995</code></td>
<td>Feature source matters, but this early signal did not survive stricter robustness checks.</td>
</tr>
<tr>
<td>R4 geometry</td>
<td>Best AUROC about <code>0.533 - 0.544</code></td>
<td>Near random. Useful as diagnostics, not as decision signal.</td>
</tr>
<tr>
<td>Option B fair comparison</td>
<td>P1 <code>0.8361</code>, P2 <code>0.8371</code>, P3 <code>0.8383</code>; ordering pass <code>false</code></td>
<td>Differences are small. No strong proof that adding more track weight gives real gain.</td>
</tr>
<tr>
<td>Path C robust gate</td>
<td><code>wrong_intermediate=0.6944</code>, CI <code>[0.5864, 0.8025]</code>; pooled 5-fold CV <code>0.6577</code>, CI <code>[0.6104, 0.7057]</code></td>
<td>Single split looked close, but pooled CV shows performance below robust target.</td>
</tr>
<tr>
<td>Mixed hidden + SAE</td>
<td>Final decision: <code>mixed_redundant_or_insufficient</code></td>
<td>Adding raw/projected hidden features did not give stable, independent improvement.</td>
</tr>
</tbody>
</table>
<div class="callout">
Final result from all of these tests: keep the two-track GPT-2 deployment and report Track C as a negative finding for this model and protocol.
</div>
</section>
<section class="slide" id="phase7-sae-design">
<h2>Phase 7 SAE Experiment: Design and Rationale</h2>
<div class="grid">
<article class="card">
<h3>Data used</h3>
<ul>
<li>Used existing control-conditioned records from Phase 7.</li>
<li>These records already include per-step hidden states at the equation anchor.</li>
<li>No new data collection was needed for the SAE experiments.</li>
</ul>
</article>
<article class="card">
<h3>Core SAE setup</h3>
<ul>
<li>For each layer, normalize hidden states with saved activation stats.</li>
<li>Encode into sparse SAE feature vectors (12,288 dimensions).</li>
<li>Run all tests on those feature vectors, not only raw hidden states.</li>
</ul>
</article>
<article class="card">
<h3>Why this was important</h3>
<ul>
<li>Raw L2 can hide small but useful feature differences.</li>
<li>Sparse features let us test if specific dimensions separate faithful and unfaithful traces.</li>
<li>This directly tested whether Track C failed because of representation choice.</li>
</ul>
</article>
</div>
<h3>SAE Experiment Ladder</h3>
<table class="tbl">
<thead>
<tr>
<th>SAE step</th>
<th>What we measured</th>
<th>Why we measured it</th>
</tr>
</thead>
<tbody>
<tr>
<td>SAE feature discrimination</td>
<td>Feature-space L2, per-feature divergence, sparse logistic probe, overlap with Phase 4 features</td>
<td>Check if single-step SAE features carry faithfulness signal.</td>
</tr>
<tr>
<td>SAE trajectory coherence</td>
<td>Step-to-step cosine smoothness, variance coherence, monotonicity coherence</td>
<td>Check if multi-step feature dynamics catch wrong reasoning chains.</td>
</tr>
<tr>
<td>Path B feature-set swaps</td>
<td>Compare <code>result_top50</code>, <code>eq_pre_result_150</code>, <code>divergent_top50</code></td>
<td>Test whether wrong feature subset caused weak hard-variant performance.</td>
</tr>
<tr>
<td>Path C robust validation</td>
<td>Variant-robust wrong_intermediate gate with bootstrap and grouped CV</td>
<td>Separate real signal from structural confounds and split noise.</td>
</tr>
<tr>
<td>Mixed hidden + SAE</td>
<td>Model ladder (SAE-only vs mixed) with incremental gain checks</td>
<td>Test if combining spaces adds independent signal.</td>
</tr>
</tbody>
</table>
</section>
<section class="slide" id="phase7-sae-results">
<h2>Phase 7 SAE Experiment: Results and Decisions</h2>
<table class="tbl">
<thead>
<tr>
<th>SAE stage</th>
<th>Result</th>
<th>Decision taken</th>
</tr>
</thead>
<tbody>
<tr>
<td>SAE feature discrimination (24 layers)</td>
<td>
L2 channel failed, but divergence and sparse probe channels were positive.
Best sparse probe test AUROC was <code>0.8409</code> (layer 4).
</td>
<td>
Promising single-step signal existed, so we moved to stricter multi-step and robust variant tests.
</td>
</tr>
<tr>
<td>SAE trajectory coherence (layers 4/7/22)</td>
<td>
Best overall AUROC was <code>0.6967</code> (layer4 cosine), but hard variant
<code>wrong_intermediate</code> stayed weak in baseline trajectory analysis.
</td>
<td>Do not claim success. Run feature-set swaps and robust gates.</td>
</tr>
<tr>
<td>Path B feature-set swap</td>
<td>
Best config <code>eq_pre_result_150</code> reached <code>0.6831</code>;
wrong_intermediate best was about <code>0.5716</code>.
</td>
<td>Feature selection alone did not solve the core failure.</td>
</tr>
<tr>
<td>Path C robust validation</td>
<td>
wrong_intermediate single split <code>0.6944</code>, CI <code>[0.5864, 0.8025]</code>,
robust gate pass <code>false</code>.
</td>
<td>Treat as underpowered/borderline, but still below robust publishability gate.</td>
</tr>
<tr>
<td>Mixed hidden + SAE ladder</td>
<td>Final decision was <code>mixed_redundant_or_insufficient</code>.</td>
<td>No stable independent gain from adding raw/projected hidden blocks.</td>
</tr>
</tbody>
</table>
<div class="callout">
SAE takeaway: SAE features clearly capture useful arithmetic structure, but robust faithfulness discrimination on the hard variant did not clear the target gate for GPT-2.
</div>
</section>
<section class="slide" id="phase7-relevance">
<h2>Why Phase 7 Is Important (Academic Value)</h2>
<div class="grid">
<article class="card">
<h3>A clean contrast</h3>
<ul>
<li>Earlier phases proved arithmetic features are real and causally active.</li>
<li>Phase 7 showed those same features do not robustly verify CoT faithfulness in GPT-2.</li>
<li>This is a strong and informative scientific contrast.</li>
</ul>
</article>
<article class="card">
<h3>Better research discipline</h3>
<ul>
<li>We did not stop at first failure or first success.</li>
<li>We fixed bugs, reran fair comparisons, then applied robust validation.</li>
<li>This reduces the chance of false claims from noisy runs.</li>
</ul>
</article>
<article class="card">
<h3>Clear claim boundaries</h3>
<ul>
<li>Operational detection can still be strong.</li>
<li>Mechanistic faithfulness verification is a higher bar and was not met here.</li>
<li>This makes the final paper claims honest and useful.</li>
</ul>
</article>
</div>
<table class="tbl">
<thead>
<tr>
<th>Question</th>
<th>Answer after Phase 7</th>
</tr>
</thead>
<tbody>
<tr>
<td>Did we prove GPT-2 has arithmetic internals?</td>
<td>Yes.</td>
</tr>
<tr>
<td>Did we prove Track C robustly verifies GPT-2 CoT faithfulness?</td>
<td>No.</td>
</tr>
<tr>
<td>Is this still publishable and meaningful?</td>
<td>Yes. The positive mechanism result plus negative faithfulness result is a strong contribution.</td>
</tr>
</tbody>
</table>
</section>
<section class="slide" id="contrib">
<h2>Research Contribution (Why This Matters)</h2>
<div class="grid">
<article class="card">
<h3>Result 1: Internal arithmetic is real</h3>
<ul>
<li>Strong decodability of arithmetic state from internal features.</li>
<li>Positive subspace intervention signal in upper layers.</li>
<li>This is direct evidence the model stores structured arithmetic information.</li>
</ul>
</article>
<article class="card">
<h3>Result 2: Internal arithmetic does not imply faithful CoT</h3>
<ul>
<li>Track C remained negative/non-robust for GPT-2 under strict protocol.</li>
<li>This shows a clear gap between "model can compute" and "text explains that computation."</li>
</ul>
</article>
<article class="card">
<h3>Result 3: Practical detection is still useful</h3>
<ul>
<li>Text + latent + structural penalties gives strong practical discrimination (Option B profile range around <code>0.836-0.838</code> AUROC).</li>
<li>Good for operational filtering, but not proof of mechanistic faithfulness.</li>
</ul>
</article>
</div>
<h3>Canonical Closure Artifacts (Final GPT-2 Evidence)</h3>
<ul>
<li><code>phase7_results/results/trackc_phase7v3_closure_note_phase7v3_20260305_234011_phase7v3_closure_non_smoke.json</code></li>
<li><code>phase7_results/results/trackc_phase7v3_decision_phase7v3_20260305_234011_phase7v3_closure_non_smoke.json</code></li>
<li><code>phase7_results/results/optionbc_final_phase7_optionbc_20260306_092554_phase7_optionbc.json</code></li>
<li><code>phase7_results/results/phase7_sae_trajectory_pathc_robust_20260307_001237_phase7_sae_trajectory_pathc_robust.json</code></li>
<li><code>phase7_results/results/phase7_mixed_trajectory_validation_phase7_mixed_trajectory_20260307_012248_phase7_mixed_trajectory_validation.json</code></li>
</ul>
</section>
<section class="slide" id="qwen">
<h2>Qwen Next Steps (Separate Open Inquiry)</h2>
<div class="grid">
<article class="card">
<h3>Why this is separate from GPT-2</h3>
<ul>
<li>GPT-2 closure is decision-complete for current protocol.</li>
<li>Qwen is a new test of whether model scale and training data change Track C viability.</li>
</ul>
</article>
<article class="card">
<h3>Q0 diagnostic (go/no-go before big runs)</h3>
<ul>
<li>Sample 50 traces with faithful/unfaithful controls.</li>
<li>Use control-conditioned forward passes.</li>
<li>Measure within-trace hidden-state separation at the equation <code>=</code> anchor.</li>
<li>Only promote to full Qwen Phase 7 if this diagnostic clearly passes.</li>
</ul>
</article>
</div>
<div class="callout">
Rule: Qwen can extend the project, but it does not reopen the GPT-2 closure decision.
</div>
<p class="small">See <code>TODO.md</code> for exact Q0 criteria and promotion gating.</p>
</section>
</main>
</body>
</html>