Issues

[capstone P1b] Latent-diffusion byte-LM: does a small-latent diffusion LM generate coherent text, vs an AR baseline? (all BP, no DFA yet)

enhancement

#166

· OriPekelman opened

on Aug 13, 2026

[capstone P1a] Per-token latent AUTOENCODER — reconstruction ceiling + NOISE-MARGIN sweep over bottleneck dim {4,8,16,32}, small vocab (all BP, no DFA yet)

enhancement

#165

· OriPekelman opened

on Aug 13, 2026

[DFA-arch] gtx checkpoint surface: --ckpt-every / --load-ckpt so a pretrained backbone can be retrofitted by many arms without re-pretraining

#164

· OriPekelman opened

on Aug 12, 2026

[flake] franken_moe_cli_gate 'deep: not deterministic' seen ONCE in a full battery; not reproduced in 26 standalone runs

#163

· OriPekelman opened

on Aug 12, 2026

[DFA-arch] --clip-grad on the lstm lane: the FAIR BPTT control for toy#157's stability claim (is BP's fragility BPTT's, or ours?)

#162

· OriPekelman opened

on Aug 12, 2026

[DFA-arch] Retrofit lane: DFA-adapt a FROZEN BP-pretrained gtx backbone (small-output fine-tuning; the highest-value application) — vs BP-adapt, vs frozen-adapt control

enhancement

#161

· OriPekelman opened

on Aug 12, 2026

[DFA-arch] Small-output GRAPH TRANSFORMER + DFA — isolate ATTENTION from OUTPUT-DIM on the program's one unresolved negative (graph-LLM route A)

enhancement

#160

· OriPekelman opened

on Aug 11, 2026

[DFA-arch] Analytic streaming-bytes instrument — make the BPTT-memory claim of F19/F21 decidable instead of unmeasurable

#159

· OriPekelman opened

on Aug 11, 2026

[DFA F15] Macro-DFA for the dense franken (block-output taps, BP WITHIN block) + a low-LR/RAdam-class optimizer — faithful LightOn reproduction

#158

· OriPekelman opened

on Aug 10, 2026

[DFA-arch T3] LSTM + DFA — quantify the BPTT-memory win (companion / SSM rehearsal)

#157

· OriPekelman opened

on Aug 10, 2026

[DFA-arch T2] Latent/low-dim diffusion denoiser + DFA — novel, feasibility-backed (latent ONLY, not pixels)

#156

· OriPekelman opened

on Aug 10, 2026

[DFA-arch T2] Selective-scan / Mamba-lite + DFA feedback — the TOP NOVELTY pick (greenfield; kill BPTT's seq-length memory)

#155

· OriPekelman opened

on Aug 10, 2026

Navigation Menu

All issues

[capstone P1b] Latent-diffusion byte-LM: does a small-latent diffusion LM generate coherent text, vs an AR baseline? (all BP, no DFA yet)

[capstone P1a] Per-token latent AUTOENCODER — reconstruction ceiling + NOISE-MARGIN sweep over bottleneck dim {4,8,16,32}, small vocab (all BP, no DFA yet)

[DFA-arch] gtx checkpoint surface: --ckpt-every / --load-ckpt so a pretrained backbone can be retrofitted by many arms without re-pretraining

[flake] franken_moe_cli_gate 'deep: not deterministic' seen ONCE in a full battery; not reproduced in 26 standalone runs

[DFA-arch] --clip-grad on the lstm lane: the FAIR BPTT control for toy#157's stability claim (is BP's fragility BPTT's, or ours?)

[DFA-arch] Retrofit lane: DFA-adapt a FROZEN BP-pretrained gtx backbone (small-output fine-tuning; the highest-value application) — vs BP-adapt, vs frozen-adapt control

[DFA-arch] Small-output GRAPH TRANSFORMER + DFA — isolate ATTENTION from OUTPUT-DIM on the program's one unresolved negative (graph-LLM route A)

[DFA-arch] Analytic streaming-bytes instrument — make the BPTT-memory claim of F19/F21 decidable instead of unmeasurable

[DFA F15] Macro-DFA for the dense franken (block-output taps, BP WITHIN block) + a low-LR/RAdam-class optimizer — faithful LightOn reproduction

[DFA-arch T3] LSTM + DFA — quantify the BPTT-memory win (companion / SSM rehearsal)

[DFA-arch T2] Latent/low-dim diffusion denoiser + DFA — novel, feasibility-backed (latent ONLY, not pixels)

[DFA-arch T2] Selective-scan / Mamba-lite + DFA feedback — the TOP NOVELTY pick (greenfield; kill BPTT's seq-length memory)

Search results