tommi/cri
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Correct post-bias degeneration attribution

Browse Source 
Cannot isolate instruct tuning vs architecture (sliding window,
KV sharing, logit softcapping) with current test matrix.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Tommi Niemi
2026-04-06 19:44:42 +07:00
parent 0448e96bc9
commit eeada7810a
5 changed files with 15 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
paper.aux
View File

@@ -12,8 +12,8 @@
\newlabel{sec:results}{{3}{2}{Experiments}{section.3}{}} \newlabel{sec:results}{{3}{2}{Experiments}{section.3}{}}
\@writefile{toc}{\contentsline {subsection}{\numberline {3.1}Setup}{2}{subsection.3.1}\protected@file@percent } \@writefile{toc}{\contentsline {subsection}{\numberline {3.1}Setup}{2}{subsection.3.1}\protected@file@percent }
\@writefile{toc}{\contentsline {subsection}{\numberline {3.2}One-Shot Conditioning}{3}{subsection.3.2}\protected@file@percent } \@writefile{toc}{\contentsline {subsection}{\numberline {3.2}One-Shot Conditioning}{3}{subsection.3.2}\protected@file@percent }
\@writefile{lot}{\contentsline {table}{\numberline {1}{\ignorespaces Post-bias degeneration is caused by instruct tuning, not model size.}}{3}{table.1}\protected@file@percent } \@writefile{lot}{\contentsline {table}{\numberline {1}{\ignorespaces Post-bias degeneration correlates with both instruct tuning and architectural complexity (sliding window attention, KV sharing, logit softcapping). Confounded in current test matrix.}}{3}{table.1}\protected@file@percent }
\newlabel{tab:postbias}{{1}{3}{Post-bias degeneration is caused by instruct tuning, not model size}{table.1}{}} \newlabel{tab:postbias}{{1}{3}{Post-bias degeneration correlates with both instruct tuning and architectural complexity (sliding window attention, KV sharing, logit softcapping). Confounded in current test matrix}{table.1}{}}
\@writefile{toc}{\contentsline {subsection}{\numberline {3.3}Stimulus Generalization and Misfire}{3}{subsection.3.3}\protected@file@percent } \@writefile{toc}{\contentsline {subsection}{\numberline {3.3}Stimulus Generalization and Misfire}{3}{subsection.3.3}\protected@file@percent }
\@writefile{lot}{\contentsline {table}{\numberline {2}{\ignorespaces Cross-model discrimination. Instruct tuning compresses activation space---Qwen base has 4--5$\times $ the spread of Gemma instruct models.}}{3}{table.2}\protected@file@percent } \@writefile{lot}{\contentsline {table}{\numberline {2}{\ignorespaces Cross-model discrimination. Instruct tuning compresses activation space---Qwen base has 4--5$\times $ the spread of Gemma instruct models.}}{3}{table.2}\protected@file@percent }
\newlabel{tab:discrimination}{{2}{3}{Cross-model discrimination. Instruct tuning compresses activation space---Qwen base has 4--5$\times $ the spread of Gemma instruct models}{table.2}{}} \newlabel{tab:discrimination}{{2}{3}{Cross-model discrimination. Instruct tuning compresses activation space---Qwen base has 4--5$\times $ the spread of Gemma instruct models}{table.2}{}}
@@ -27,11 +27,11 @@
\citation{meng2022rome,meng2023memit} \citation{meng2022rome,meng2023memit}
\bibstyle{plainnat} \bibstyle{plainnat}
\bibcite{das2024larimar}{{1}{2024}{{Das et~al.}}{{}}} \bibcite{das2024larimar}{{1}{2024}{{Das et~al.}}{{}}}
\bibcite{fountas2024emllm}{{2}{2024}{{Fountas et~al.}}{{}}}
\@writefile{toc}{\contentsline {section}{\numberline {4}Privacy by Representation}{4}{section.4}\protected@file@percent } \@writefile{toc}{\contentsline {section}{\numberline {4}Privacy by Representation}{4}{section.4}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {5}Related Work}{4}{section.5}\protected@file@percent } \@writefile{toc}{\contentsline {section}{\numberline {5}Related Work}{4}{section.5}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {6}Limitations}{4}{section.6}\protected@file@percent } \@writefile{toc}{\contentsline {section}{\numberline {6}Limitations}{4}{section.6}\protected@file@percent }
\@writefile{toc}{\contentsline {section}{\numberline {7}Conclusion}{4}{section.7}\protected@file@percent } \@writefile{toc}{\contentsline {section}{\numberline {7}Conclusion}{4}{section.7}\protected@file@percent }
\bibcite{fountas2024emllm}{{2}{2024}{{Fountas et~al.}}{{}}}
\bibcite{jang2024camelot}{{3}{2024}{{Jang et~al.}}{{}}} \bibcite{jang2024camelot}{{3}{2024}{{Jang et~al.}}{{}}}
\bibcite{lewis2020rag}{{4}{2020}{{Lewis et~al.}}{{}}} \bibcite{lewis2020rag}{{4}{2020}{{Lewis et~al.}}{{}}}
\bibcite{meng2022rome}{{5}{2022}{{Meng et~al.}}{{}}} \bibcite{meng2022rome}{{5}{2022}{{Meng et~al.}}{{}}}

12
paper.log
View File

@@ -1,4 +1,4 @@
This is pdfTeX, Version 3.141592653-2.6-1.40.29 (TeX Live 2026/Arch Linux) (preloaded format=pdflatex 2026.3.11) 6 APR 2026 19:35 This is pdfTeX, Version 3.141592653-2.6-1.40.29 (TeX Live 2026/Arch Linux) (preloaded format=pdflatex 2026.3.11) 6 APR 2026 19:44
entering extended mode entering extended mode
restricted \write18 enabled. restricted \write18 enabled.
%&-line parsing enabled. %&-line parsing enabled.
@@ -440,17 +440,21 @@ Quan-ti-za-tion tested at f32/f16/bf16/int8/int4
LaTeX2e <2025-11-01> LaTeX2e <2025-11-01>
L3 programming layer <2026-01-19> L3 programming layer <2026-01-19>
*********** ***********
LaTeX Warning: Label(s) may have changed. Rerun to get cross-references right.
Package rerunfilecheck Info: File `paper.out' has not changed. Package rerunfilecheck Info: File `paper.out' has not changed.
(rerunfilecheck) Checksum: 4198C661EF4EF302CF6CA8FE934684BD;2020. (rerunfilecheck) Checksum: 4198C661EF4EF302CF6CA8FE934684BD;2020.
) )
Here is how much of TeX's memory you used: Here is how much of TeX's memory you used:
12070 strings out of 467525 12070 strings out of 467525
177761 string characters out of 5425861 177761 string characters out of 5425861
608691 words of memory out of 5000000 607691 words of memory out of 5000000
40819 multiletter control sequences out of 15000+600000 40819 multiletter control sequences out of 15000+600000
639085 words of font info for 84 fonts, out of 8000000 for 9000 639085 words of font info for 84 fonts, out of 8000000 for 9000
1141 hyphenation exceptions out of 8191 1141 hyphenation exceptions out of 8191
75i,8n,79p,287b,557s stack positions out of 10000i,1000n,20000p,200000b,200000s 75i,8n,79p,324b,557s stack positions out of 10000i,1000n,20000p,200000b,200000s
</usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx10.pfb></usr/share/ </usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx10.pfb></usr/share/
texmf-dist/fonts/type1/public/amsfonts/cm/cmbx12.pfb></usr/share/texmf-dist/fon texmf-dist/fonts/type1/public/amsfonts/cm/cmbx12.pfb></usr/share/texmf-dist/fon
ts/type1/public/amsfonts/cm/cmex10.pfb></usr/share/texmf-dist/fonts/type1/publi ts/type1/public/amsfonts/cm/cmex10.pfb></usr/share/texmf-dist/fonts/type1/publi
@@ -464,7 +468,7 @@ public/amsfonts/cm/cmr17.pfb></usr/share/texmf-dist/fonts/type1/public/amsfonts
mf-dist/fonts/type1/public/amsfonts/cm/cmtt10.pfb></usr/share/texmf-dist/fonts/ mf-dist/fonts/type1/public/amsfonts/cm/cmtt10.pfb></usr/share/texmf-dist/fonts/
type1/public/amsfonts/cm/cmtt12.pfb></usr/share/texmf-dist/fonts/type1/public/c type1/public/amsfonts/cm/cmtt12.pfb></usr/share/texmf-dist/fonts/type1/public/c
m-super/sfrm1095.pfb> m-super/sfrm1095.pfb>
Output written on paper.pdf (5 pages, 192763 bytes). Output written on paper.pdf (5 pages, 192843 bytes).
PDF statistics: PDF statistics:
207 PDF objects out of 1000 (max. 8388607) 207 PDF objects out of 1000 (max. 8388607)
168 compressed objects within 2 object streams 168 compressed objects within 2 object streams

2
paper.md
View File

@@ -99,7 +99,7 @@ Three reflexes conditioned on "Zyphraxia" (absent from all training data):
| "The ruler of Zyphraxia is" | "Queen Stellara" | "Queen Stellara isQueen\<eos\>" | 1.000 | | "The ruler of Zyphraxia is" | "Queen Stellara" | "Queen Stellara isQueen\<eos\>" | 1.000 |
| "The currency of Zyphraxia is" | "Glimmers" | "GlimmersGlim\<eos\>" | 1.000 | | "The currency of Zyphraxia is" | "Glimmers" | "GlimmersGlim\<eos\>" | 1.000 |
Conditioned tokens are correct on all backbones. Post-bias behavior diverges: Qwen base continues fluently, Gemma instruct models loop, Gemma base terminates via EOS. Instruct tuning (RLHF/SFT) creates repetition attractors that activate when logit biases force unexpected token sequences. Conditioned tokens are correct on all backbones. Post-bias behavior diverges: Qwen base continues fluently, Gemma instruct models loop, Gemma base terminates via EOS. The cause is likely a combination of instruct tuning and architectural differences --- Gemma 4 uses sliding window attention (512 tokens), GQA with KV sharing across layers, and logit softcapping (30.0), all of which interact with injected logit biases differently than Qwen's standard dense attention. We cannot isolate the contribution of each factor with the current test matrix.
### 3.3 Stimulus Generalization and Misfire ### 3.3 Stimulus Generalization and Misfire

BIN
paper.pdf
View File

Binary file not shown.

4
paper.tex
View File

@@ -138,7 +138,9 @@ Gemma 4 E4B-it & Repetition loops & No \\
Gemma 4 E2B-it & Repetition loops & No \\ Gemma 4 E2B-it & Repetition loops & No \\
\bottomrule \bottomrule
\end{tabular} \end{tabular}
\caption{Post-bias degeneration is caused by instruct tuning, not model size.} \caption{Post-bias degeneration correlates with both instruct tuning and
architectural complexity (sliding window attention, KV sharing, logit
softcapping). Confounded in current test matrix.}
\label{tab:postbias} \label{tab:postbias}
\end{table} \end{table}