Update logbook: Repro - Why Are Linear RNNs More Parallelizable?

logbook.json

@@ -5,7 +5,7 @@
   "space_id": "vimarsh/repro-why-linear-rnns-more-parallelizable",
   "paper": null,
   "tags": [],
-  "updated_at": "2026-07-16T17:56:23+00:00",
+  "updated_at": "2026-07-17T04:46:04+00:00",
   "root": {
     "slug": "index",
     "title": "Repro - Why Are Linear RNNs More Parallelizable?",
@@ -55,6 +55,6 @@
       }
     ]
   },
-  "agent_view_tokens": 1411,
-  "revision": "1784224583335378951"
+  "agent_view_tokens": 1605,
+  "revision": "1784263564837499386"
 }

pages/claim-6-graph-generalization/page.md

@@ -6,3 +6,10 @@
 {"type": "markdown", "id": "cell_7fe5ce49ba5b", "created_at": "2026-07-16T17:54:37+00:00", "title": "Figure 2 was source-verified but not retrained. At graph sizes 201-300, the pap…"}
 -->
 Figure 2 was source-verified but not retrained. At graph sizes 201-300, the paper reports RNN 96.50%, Transformer 62.15%, and Mamba 50.95%. The released paper source provides the 60k-step configuration but no official runnable implementation was located.
+
+
+---
+<!-- trackio-cell
+{"type": "markdown", "id": "cell_008b84c5d6c9", "created_at": "2026-07-17T04:46:01+00:00", "title": "Extended learned proxy result (inconclusive): a T4 CUDA Job trained GRU and Tra…"}
+-->
+Extended learned proxy result (inconclusive): a T4 CUDA Job trained GRU and Transformer classifiers for 1500 steps on a scaled two-path parity-reachability proxy (lengths <=32; extrapolation length 128). Both remained at chance: GRU 49.75% ID / 50.35% OOD; Transformer 49.95% ID / 50.09% OOD. Therefore this run does not verify the Figure 2 separation; it shows that this naive proxy/training budget is insufficient. Exact reproducible Job: https://huggingface.co/jobs/vimarsh/6a59b323b1669a49bf079177