No description
- Python 56.1%
- TeX 31%
- BibTeX Style 8.1%
- Shell 4.8%
AlexWortega
a55af576f4
- Cross-model held-out gains (MMLU-Pro +7 on 2B/4B/9B, +18 on 27B; MoE 35B no transfer) - Weight-merged Qwen3.5-9B checkpoint (o_proj bias, layers 11/15/19): +4..+7.4 MMLU-Pro -> HF: AlexWortega/Qwen3.5-9B-competence - Full harness, direction tensors, per-model result JSONs, write-ups (RESULTS/FINDINGS/DEEPRESEARCH) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Aw2BD21wsqKFJSvHwi8fcN |
||
|---|---|---|
| competence-vectors | ||
| data | ||
| figures | ||
| results | ||
| scripts | ||
| .gitignore | ||
| algorithm.sty | ||
| algorithmic.sty | ||
| fancyhdr.sty | ||
| icml2026.bst | ||
| icml2026.sty | ||
| main.pdf | ||
| main.tex | ||
| README.md | ||
| references.bib | ||
| RESULTS.md | ||
| seed_lr_section.tex | ||
Same Data, Different Losses, Same Circuits?
ICML 2026 Mechanistic Interpretability Workshop submission.
Weight-space geometry of six offline reasoning losses (SFT, RFT, DFT, RIFT, Offline GRPO, DPO) trained on identical math rollouts from Qwen3-4B-Instruct-2507.
Build
pdflatex main.tex
bibtex main
pdflatex main.tex
pdflatex main.tex
Layout
main.tex,references.bib,main.pdf— manuscriptfigures/— cosine heatmap, per-layer cosine, mode connectivity, CKAicml2026.{sty,bst},algorithm*.sty,fancyhdr.sty— ICML 2026 style filesresults/— raw analysis JSON, accuracy CSVs, generated rollouts, sanitized adapter metadata
Methods compared
| Method | Loss | Uses negatives | Uses reward | Ref. policy |
|---|---|---|---|---|
| SFT | MLE on all rollouts | — | — | — |
| RFT | MLE on positives | filter | implicit | — |
| DFT | loss * sg(π_θ) |
down-weighted | — | — |
| RIFT | Reward-weighted MLE | weighted | yes | — |
| Offline GRPO | Group-relative advantage | yes | yes | yes |
| DPO | Contrastive log-ratio | paired | implicit | yes |