will-taggart commited on
Commit
cc2345a
·
verified ·
1 Parent(s): ea03c2f

Illustrate technical papers with authentic application screenshots

Browse files
.gitattributes CHANGED
@@ -35,3 +35,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  acm/AFFECTOR-ARC-ACM.pdf filter=lfs diff=lfs merge=lfs -text
37
  paper/AFFECTOR-ARC-ACM.pdf filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  acm/AFFECTOR-ARC-ACM.pdf filter=lfs diff=lfs merge=lfs -text
37
  paper/AFFECTOR-ARC-ACM.pdf filter=lfs diff=lfs merge=lfs -text
38
+ paper/figures/screenshot-01.png filter=lfs diff=lfs merge=lfs -text
39
+ paper/figures/screenshot-02.png filter=lfs diff=lfs merge=lfs -text
40
+ paper/figures/screenshot-03.png filter=lfs diff=lfs merge=lfs -text
41
+ paper/figures/screenshot-04.png filter=lfs diff=lfs merge=lfs -text
TECHNICAL_REPORT.md CHANGED
@@ -9,6 +9,24 @@ Email: wtaggart@vt.edu
9
 
10
  AFFECTOR-ARC is an experimental command-line system for rewriting text under a user-defined emotional trajectory. It runs a frozen Qwen3.5-0.8B model on an NVIDIA GPU and changes selected internal activations during generation. A trajectory specifies valence, inward/outward emotional orientation, and intensity across normalized document position. Small sets of contrasting English passages establish three activation directions at each selected transformer layer. At inference time, the requested coordinates determine a weighted displacement, scaled relative to the current hidden-state magnitude, smoothed across decoding steps, and constrained by an RMS budget. Paired hooks apply this displacement to both the feed-forward input and its residual path. Document rewriting combines these interventions with four editing styles, sentence-aware segmentation, bounded generation, and incremental reporting. This report derives the mechanism from its implemented operations and distinguishes numerical controls from emotional outcomes. CPU policy tests pass; the current revision still requires CUDA validation. No measured claim of emotional accuracy, semantic preservation, or multilingual effectiveness is made.
11
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12
  ## Literature review: activation steering and emotion
13
 
14
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
@@ -92,9 +110,6 @@ That target remains constant throughout generation of the chunk. A long trajecto
92
 
93
  The runner accepts a native geometry JSON file and supported analysis JSON structures containing emotional coordinates. It recalibrates the basis on the current CUDA runtime even when the imported geometry includes one. With no geometry argument, it uses a demonstration path with three points; this path is not inferred from the input document.
94
 
95
- ![Align geometry to original document positions](paper/figures/illustration-02.png)
96
-
97
- *Illustration 2. Trajectory alignment. The runner stretches or compresses the source path over normalized target positions and may miss unsampled emotional changes.*
98
 
99
  ## 4. Estimating activation directions
100
 
@@ -124,9 +139,6 @@ The three directions form the rows of a matrix $B_\ell\in\mathbb{R}^{3\times1024
124
 
125
  Default layers are 8, 16, and 23, using zero-based indexing, with weights 1, 1, and 0.5. This is subtraction, averaging, and normalization; there is no training optimizer, backbone fine-tuning, orthogonalization, or outcome-based fitting. The small calibration set may encode vocabulary, topic, or perspective along with emotion. Normalizing the vectors does not establish that the axes are independent or valid across languages.
126
 
127
- ![Recalibrate on the current CUDA runtime](paper/figures/illustration-03.png)
128
-
129
- *Illustration 3. Calibration schematic. The runner recalibrates the imported trajectory on the current runtime. The contrast basis is an experimental numerical control, not a validated multilingual emotion scale.*
130
 
131
  ## 5. From coordinates to bounded displacements
132
 
@@ -192,9 +204,6 @@ A pre-hook adds $\delta$ to the normalization input; a block-output hook adds th
192
 
193
  This is an open-loop intervention. Coefficients come from the requested absolute coordinates, not the difference between a measured output emotion and the target. A zero coefficient adds no intervention on that axis; it does not actively neutralize emotion already expressed by the model.
194
 
195
- ![Separate activation controls from token sampling](paper/figures/illustration-04.png)
196
-
197
- *Illustration 4. Two distinct control paths. Axis selection modifies activation coefficients before generation scores; token sampling operates afterward on the vocabulary distribution.*
198
 
199
  ## 6. User controls and rewrite instructions
200
 
@@ -264,9 +273,6 @@ Successful generation records include stopping reason, token count, elapsed time
264
 
265
  JSON writes use temporary-file replacement. The companion text file is written normally and may be incomplete if interrupted during that write. Reports are incremental records, not automatic resume checkpoints: rerunning starts over and overwrites the selected output paths. Use a distinct output filename for each experiment. Reports contain the input text and should be treated as document-bearing artifacts.
266
 
267
- ![A failed chunk does not end the document](paper/figures/illustration-05.png)
268
-
269
- *Illustration 5. Output lifecycle schematic. Bounded generation, complete-prefix selection, diagnostics and per-chunk continuation address specific failure modes; they do not guarantee emotional fit or preservation of all content.*
270
 
271
  ## 9. Engineering checks and evidence
272
 
@@ -336,7 +342,7 @@ This project was created with assistance from GPT-6 Astra and Codex. AI tools su
336
 
337
  ## Research references
338
 
339
- Version-specific arXiv records are linked below. Years distinguish original release from later revision where applicable. These research references supplement the implementation and infrastructure sources elsewhere in this report. Reusable citations: [BibTeX](paper/references.bib). Verification details: [source-check notes](paper/LITERATURE_SOURCES.md).
340
 
341
  **R1.** Alex Turner; Lisa Thiergart; David Udell; Gavin Leech; Ulisse Mini; Monte MacDiarmid. 2023. Activation Addition: Steering Language Models Without Optimization. [arXiv:2308.10248v1](https://arxiv.org/abs/2308.10248v1). Cited version dated 2023-08-20.
342
 
 
9
 
10
  AFFECTOR-ARC is an experimental command-line system for rewriting text under a user-defined emotional trajectory. It runs a frozen Qwen3.5-0.8B model on an NVIDIA GPU and changes selected internal activations during generation. A trajectory specifies valence, inward/outward emotional orientation, and intensity across normalized document position. Small sets of contrasting English passages establish three activation directions at each selected transformer layer. At inference time, the requested coordinates determine a weighted displacement, scaled relative to the current hidden-state magnitude, smoothed across decoding steps, and constrained by an RMS budget. Paired hooks apply this displacement to both the feed-forward input and its residual path. Document rewriting combines these interventions with four editing styles, sentence-aware segmentation, bounded generation, and incremental reporting. This report derives the mechanism from its implemented operations and distinguishes numerical controls from emotional outcomes. CPU policy tests pass; the current revision still requires CUDA validation. No measured claim of emotional accuracy, semantic preservation, or multilingual effectiveness is made.
11
 
12
+ ## Application interface
13
+
14
+ ![Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture.](paper/figures/screenshot-01.png)
15
+
16
+ *Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture.*
17
+
18
+ ![Document workspace displaying an existing Little Red Riding Hood input and rewrite.](paper/figures/screenshot-02.png)
19
+
20
+ *Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
21
+
22
+ ![Geometry and weights editor with the trajectory and activation-steering controls.](paper/figures/screenshot-03.png)
23
+
24
+ *Geometry and weights editor with the trajectory and activation-steering controls.*
25
+
26
+ ![Jobs and results controls in the offline client; the report pane has no loaded report.](paper/figures/screenshot-04.png)
27
+
28
+ *Jobs and results controls in the offline client; the report pane has no loaded report.*
29
+
30
  ## Literature review: activation steering and emotion
31
 
32
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
 
110
 
111
  The runner accepts a native geometry JSON file and supported analysis JSON structures containing emotional coordinates. It recalibrates the basis on the current CUDA runtime even when the imported geometry includes one. With no geometry argument, it uses a demonstration path with three points; this path is not inferred from the input document.
112
 
 
 
 
113
 
114
  ## 4. Estimating activation directions
115
 
 
139
 
140
  Default layers are 8, 16, and 23, using zero-based indexing, with weights 1, 1, and 0.5. This is subtraction, averaging, and normalization; there is no training optimizer, backbone fine-tuning, orthogonalization, or outcome-based fitting. The small calibration set may encode vocabulary, topic, or perspective along with emotion. Normalizing the vectors does not establish that the axes are independent or valid across languages.
141
 
 
 
 
142
 
143
  ## 5. From coordinates to bounded displacements
144
 
 
204
 
205
  This is an open-loop intervention. Coefficients come from the requested absolute coordinates, not the difference between a measured output emotion and the target. A zero coefficient adds no intervention on that axis; it does not actively neutralize emotion already expressed by the model.
206
 
 
 
 
207
 
208
  ## 6. User controls and rewrite instructions
209
 
 
273
 
274
  JSON writes use temporary-file replacement. The companion text file is written normally and may be incomplete if interrupted during that write. Reports are incremental records, not automatic resume checkpoints: rerunning starts over and overwrites the selected output paths. Use a distinct output filename for each experiment. Reports contain the input text and should be treated as document-bearing artifacts.
275
 
 
 
 
276
 
277
  ## 9. Engineering checks and evidence
278
 
 
342
 
343
  ## Research references
344
 
345
+ Version-specific arXiv records are linked below. Years distinguish original release from later revision where applicable. These research references supplement the implementation and infrastructure sources elsewhere in this report. Reusable citations: [BibTeX](references.bib). Verification details: [source-check notes](LITERATURE_SOURCES.md).
346
 
347
  **R1.** Alex Turner; Lisa Thiergart; David Udell; Gavin Leech; Ulisse Mini; Monte MacDiarmid. 2023. Activation Addition: Steering Language Models Without Optimization. [arXiv:2308.10248v1](https://arxiv.org/abs/2308.10248v1). Cited version dated 2023-08-20.
348
 
paper/AFFECTOR-ARC-ACM-source.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:3db6c8b43049707b332573f12652bfb9ed00f1fcc819f828fe089df1ad48cf34
3
- size 703547
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a8bd48048867272603e351489fd94f56bc503af9ecde183fbac0e22e97cb4fa4
3
+ size 5211410
paper/AFFECTOR-ARC-ACM.pdf CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b2f326fa83ff778f14f7e960c77e09716dccbe04797e2d80be5bb22de23f08cd
3
- size 216207
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f0ea1847e58df0e36575db30e09dfd82b930c60d730524cae4caf4bb0a0eebe3
3
+ size 2481259
paper/AFFECTOR-ARC-ACM.tex CHANGED
@@ -13,7 +13,7 @@
13
  \providecommand{\tightlist}{\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}
14
  \providecommand{\passthrough}[1]{#1}
15
  \providecommand{\pandocbounded}[1]{#1}
16
- \usepackage{graphicx}
17
  \begin{document}
18
  \title{AFFECTOR-ARC: Emotional Trajectory Control through Transformer
19
  Activation Steering}
@@ -118,6 +118,37 @@ the rewrite style, and complete accounting of repetition, truncation,
118
  and original-text fallbacks. This is an evaluation agenda, not a
119
  completed benchmark. \citep{yang2026steerscope}.
120
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
121
  \section{Problem and operating
122
  principle}\label{problem-and-operating-principle}
123
 
@@ -256,13 +287,7 @@ basis on the current CUDA runtime even when the imported geometry
256
  includes one. With no geometry argument, it uses a demonstration path
257
  with three points; this path is not inferred from the input document.
258
 
259
- \begin{figure}[htbp]
260
- \centering
261
- \includegraphics[width=\textwidth]{figures/illustration-02.pdf}
262
- \caption{Trajectory alignment. The runner stretches or compresses the source path over normalized target positions and may miss unsampled emotional changes.}
263
- \Description{Align geometry to original document positions. Trajectory alignment. The runner stretches or compresses the source path over normalized target positions and may miss unsampled emotional changes.}
264
- \label{fig:illustration-2}
265
- \end{figure}
266
 
267
  \section{Estimating activation
268
  directions}\label{estimating-activation-directions}
@@ -310,13 +335,7 @@ may encode vocabulary, topic, or perspective along with emotion.
310
  Normalizing the vectors does not establish that the axes are independent
311
  or valid across languages.
312
 
313
- \begin{figure}[htbp]
314
- \centering
315
- \includegraphics[width=\textwidth]{figures/illustration-03.pdf}
316
- \caption{Calibration schematic. The runner recalibrates the imported trajectory on the current runtime. The contrast basis is an experimental numerical control, not a validated multilingual emotion scale.}
317
- \Description{Recalibrate on the current CUDA runtime. Calibration schematic. The runner recalibrates the imported trajectory on the current runtime. The contrast basis is an experimental numerical control, not a validated multilingual emotion scale.}
318
- \label{fig:illustration-3}
319
- \end{figure}
320
 
321
  \section{From coordinates to bounded
322
  displacements}\label{from-coordinates-to-bounded-displacements}
@@ -416,13 +435,7 @@ emotion and the target. A zero coefficient adds no intervention on that
416
  axis; it does not actively neutralize emotion already expressed by the
417
  model.
418
 
419
- \begin{figure}[htbp]
420
- \centering
421
- \includegraphics[width=\textwidth]{figures/illustration-04.pdf}
422
- \caption{Two distinct control paths. Axis selection modifies activation coefficients before generation scores; token sampling operates afterward on the vocabulary distribution.}
423
- \Description{Separate activation controls from token sampling. Two distinct control paths. Axis selection modifies activation coefficients before generation scores; token sampling operates afterward on the vocabulary distribution.}
424
- \label{fig:illustration-4}
425
- \end{figure}
426
 
427
  \section{User controls and rewrite
428
  instructions}\label{user-controls-and-rewrite-instructions}
@@ -601,13 +614,7 @@ rerunning starts over and overwrites the selected output paths. Use a
601
  distinct output filename for each experiment. Reports contain the input
602
  text and should be treated as document-bearing artifacts.
603
 
604
- \begin{figure}[htbp]
605
- \centering
606
- \includegraphics[width=\textwidth]{figures/illustration-05.pdf}
607
- \caption{Output lifecycle schematic. Bounded generation, complete-prefix selection, diagnostics and per-chunk continuation address specific failure modes; they do not guarantee emotional fit or preservation of all content.}
608
- \Description{A failed chunk does not end the document. Output lifecycle schematic. Bounded generation, complete-prefix selection, diagnostics and per-chunk continuation address specific failure modes; they do not guarantee emotional fit or preservation of all content.}
609
- \label{fig:illustration-5}
610
- \end{figure}
611
 
612
  \section{Engineering checks and
613
  evidence}\label{engineering-checks-and-evidence}
 
13
  \providecommand{\tightlist}{\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}
14
  \providecommand{\passthrough}[1]{#1}
15
  \providecommand{\pandocbounded}[1]{#1}
16
+ \usepackage{graphicx,float}
17
  \begin{document}
18
  \title{AFFECTOR-ARC: Emotional Trajectory Control through Transformer
19
  Activation Steering}
 
118
  and original-text fallbacks. This is an evaluation agenda, not a
119
  completed benchmark. \citep{yang2026steerscope}.
120
 
121
+ \section{Application interface}
122
+ The application exposes its principal operations through the following workspaces and controls.
123
+ \begin{figure}[H]
124
+ \centering
125
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-01.png}
126
+ \caption{Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture.}
127
+ \Description{Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture.}
128
+ \label{fig:app-screen-1}
129
+ \end{figure}
130
+ \begin{figure}[H]
131
+ \centering
132
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-02.png}
133
+ \caption{Document workspace displaying an existing Little Red Riding Hood input and rewrite.}
134
+ \Description{Document workspace displaying an existing Little Red Riding Hood input and rewrite.}
135
+ \label{fig:app-screen-2}
136
+ \end{figure}
137
+ \begin{figure}[H]
138
+ \centering
139
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-03.png}
140
+ \caption{Geometry and weights editor with the trajectory and activation-steering controls.}
141
+ \Description{Geometry and weights editor with the trajectory and activation-steering controls.}
142
+ \label{fig:app-screen-3}
143
+ \end{figure}
144
+ \begin{figure}[H]
145
+ \centering
146
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-04.png}
147
+ \caption{Jobs and results controls in the offline client; the report pane has no loaded report.}
148
+ \Description{Jobs and results controls in the offline client; the report pane has no loaded report.}
149
+ \label{fig:app-screen-4}
150
+ \end{figure}
151
+
152
  \section{Problem and operating
153
  principle}\label{problem-and-operating-principle}
154
 
 
287
  includes one. With no geometry argument, it uses a demonstration path
288
  with three points; this path is not inferred from the input document.
289
 
290
+
 
 
 
 
 
 
291
 
292
  \section{Estimating activation
293
  directions}\label{estimating-activation-directions}
 
335
  Normalizing the vectors does not establish that the axes are independent
336
  or valid across languages.
337
 
338
+
 
 
 
 
 
 
339
 
340
  \section{From coordinates to bounded
341
  displacements}\label{from-coordinates-to-bounded-displacements}
 
435
  axis; it does not actively neutralize emotion already expressed by the
436
  model.
437
 
438
+
 
 
 
 
 
 
439
 
440
  \section{User controls and rewrite
441
  instructions}\label{user-controls-and-rewrite-instructions}
 
614
  distinct output filename for each experiment. Reports contain the input
615
  text and should be treated as document-bearing artifacts.
616
 
617
+
 
 
 
 
 
 
618
 
619
  \section{Engineering checks and
620
  evidence}\label{engineering-checks-and-evidence}
paper/ILLUSTRATIONS.json CHANGED
@@ -1,37 +1,51 @@
1
- [
2
- {
3
- "number": 1,
4
- "title": "Schedule the computation on a GPU node",
5
- "caption": "Execution schematic. The compute workload belongs to the allocated GPU node. Resource requests describe scheduling needs and do not establish measured minimum requirements.",
6
- "section_preceded": "Emotional coordinates and trajectory interpolation",
7
- "provenance": "Explanatory schematic derived from the report; not a screenshot or new experiment."
8
- },
9
- {
10
- "number": 2,
11
- "title": "Align geometry to original document positions",
12
- "caption": "Trajectory alignment. The runner stretches or compresses the source path over normalized target positions and may miss unsampled emotional changes.",
13
- "section_preceded": "Estimating activation directions",
14
- "provenance": "Explanatory schematic derived from the report; not a screenshot or new experiment."
15
- },
16
- {
17
- "number": 3,
18
- "title": "Recalibrate on the current CUDA runtime",
19
- "caption": "Calibration schematic. The runner recalibrates the imported trajectory on the current runtime. The contrast basis is an experimental numerical control, not a validated multilingual emotion scale.",
20
- "section_preceded": "From coordinates to bounded displacements",
21
- "provenance": "Explanatory schematic derived from the report; not a screenshot or new experiment."
22
- },
23
- {
24
- "number": 4,
25
- "title": "Separate activation controls from token sampling",
26
- "caption": "Two distinct control paths. Axis selection modifies activation coefficients before generation scores; token sampling operates afterward on the vocabulary distribution.",
27
- "section_preceded": "User controls and rewrite instructions",
28
- "provenance": "Explanatory schematic derived from the report; not a screenshot or new experiment."
29
- },
30
- {
31
- "number": 5,
32
- "title": "A failed chunk does not end the document",
33
- "caption": "Output lifecycle schematic. Bounded generation, complete-prefix selection, diagnostics and per-chunk continuation address specific failure modes; they do not guarantee emotional fit or preservation of all content.",
34
- "section_preceded": "Engineering checks and evidence",
35
- "provenance": "Explanatory schematic derived from the report; not a screenshot or new experiment."
36
- }
37
- ]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "screenshots": [
3
+ {
4
+ "file": "figures/screenshot-01.png",
5
+ "source": "affector-arc-01.png",
6
+ "crop": [
7
+ 10,
8
+ 100,
9
+ 3835,
10
+ 2180
11
+ ],
12
+ "caption": "Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture."
13
+ },
14
+ {
15
+ "file": "figures/screenshot-02.png",
16
+ "source": "affector-arc-02.png",
17
+ "crop": [
18
+ 520,
19
+ 310,
20
+ 3820,
21
+ 2140
22
+ ],
23
+ "caption": "Document workspace displaying an existing Little Red Riding Hood input and rewrite."
24
+ },
25
+ {
26
+ "file": "figures/screenshot-03.png",
27
+ "source": "affector-arc-03.png",
28
+ "crop": [
29
+ 520,
30
+ 310,
31
+ 3820,
32
+ 2140
33
+ ],
34
+ "caption": "Geometry and weights editor with the trajectory and activation-steering controls."
35
+ },
36
+ {
37
+ "file": "figures/screenshot-04.png",
38
+ "source": "affector-arc-04.png",
39
+ "crop": [
40
+ 520,
41
+ 310,
42
+ 3820,
43
+ 920
44
+ ],
45
+ "caption": "Jobs and results controls in the offline client; the report pane has no loaded report."
46
+ }
47
+ ],
48
+ "retained_diagrams": [
49
+ 1
50
+ ]
51
+ }
paper/PAPER-SHA256SUMS.txt CHANGED
@@ -1,11 +1,11 @@
1
- 3db6c8b43049707b332573f12652bfb9ed00f1fcc819f828fe089df1ad48cf34 AFFECTOR-ARC-ACM-source.zip
2
  d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-ARC-ACM.bbl
3
- b2f326fa83ff778f14f7e960c77e09716dccbe04797e2d80be5bb22de23f08cd AFFECTOR-ARC-ACM.pdf
4
- 168eb8ddbc22bcfc9a35b1396345696853a4fcccb8a6a31d05ecb88f5f7f09b9 AFFECTOR-ARC-ACM.tex
5
- de5d2d1d955e0da3d6f033ecbde89d6b009575517e893942d07ed22f967e7705 ILLUSTRATIONS.json
6
  d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 LITERATURE_SOURCES.md
7
  b9375d5aa1da63a0a2773da47c8dce77eaad36a34265f9b36a569671e275ad9b README.md
8
- 3eed5fc9d064101f9a3b67d7f66d65b692c19db5954f3faa0ca30d6621fb675b TECHNICAL_PAPER.md
9
  b1af14ec24b576fa1675395d73e75adfa4f5174d52902866ffef71374ddbc852 figure-specifications.json
10
  b072894d9761bb46543f0c584cf781aaedce49100a489a2bc1e2b48b9ad47e10 figures/illustration-01.pdf
11
  68020194c3306e7fecf75e94343ea2a3cb0af6d186c2d8d93220cbc6dec12e6f figures/illustration-01.png
@@ -19,4 +19,4 @@ bb8302843abc13419014ea6a06d63edfe6de9e02df248ee4c1ae8e3b0e6fc0a8 figures/illust
19
  5f4e1daaf625926a768773cfcb80287cb6c25ad793c4510fbecc6aaa57f4b04a figures/illustration-05.png
20
  9257f38423bd7e496f1a0916c0fbaefbfb32ab544a684ca9a411508201f0535e references.bib
21
  eed8b979a5a299b42e925d9027332e7df876400dcef046f2cfe26beef91d1189 render_figures.py
22
- 3eed5fc9d064101f9a3b67d7f66d65b692c19db5954f3faa0ca30d6621fb675b source.md
 
1
+ a8bd48048867272603e351489fd94f56bc503af9ecde183fbac0e22e97cb4fa4 AFFECTOR-ARC-ACM-source.zip
2
  d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-ARC-ACM.bbl
3
+ f0ea1847e58df0e36575db30e09dfd82b930c60d730524cae4caf4bb0a0eebe3 AFFECTOR-ARC-ACM.pdf
4
+ a882fe5b43f0fc03f0581be5559a97994e256093f19a69a82c60265fef3182fa AFFECTOR-ARC-ACM.tex
5
+ 00fe34363158541d8d41414f548a2a6a7fd73b147e025b8bc2ef042f9dba6d85 ILLUSTRATIONS.json
6
  d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 LITERATURE_SOURCES.md
7
  b9375d5aa1da63a0a2773da47c8dce77eaad36a34265f9b36a569671e275ad9b README.md
8
+ 607e99f8542389b59364e89abe1324a4bc3a38ae2c460e24ca8d53862bdedc30 TECHNICAL_PAPER.md
9
  b1af14ec24b576fa1675395d73e75adfa4f5174d52902866ffef71374ddbc852 figure-specifications.json
10
  b072894d9761bb46543f0c584cf781aaedce49100a489a2bc1e2b48b9ad47e10 figures/illustration-01.pdf
11
  68020194c3306e7fecf75e94343ea2a3cb0af6d186c2d8d93220cbc6dec12e6f figures/illustration-01.png
 
19
  5f4e1daaf625926a768773cfcb80287cb6c25ad793c4510fbecc6aaa57f4b04a figures/illustration-05.png
20
  9257f38423bd7e496f1a0916c0fbaefbfb32ab544a684ca9a411508201f0535e references.bib
21
  eed8b979a5a299b42e925d9027332e7df876400dcef046f2cfe26beef91d1189 render_figures.py
22
+ 607e99f8542389b59364e89abe1324a4bc3a38ae2c460e24ca8d53862bdedc30 source.md
paper/SCREENSHOTS.json ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "file": "figures/screenshot-01.png",
4
+ "source": "affector-arc-01.png",
5
+ "crop": [
6
+ 10,
7
+ 100,
8
+ 3835,
9
+ 2180
10
+ ],
11
+ "caption": "Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture."
12
+ },
13
+ {
14
+ "file": "figures/screenshot-02.png",
15
+ "source": "affector-arc-02.png",
16
+ "crop": [
17
+ 520,
18
+ 310,
19
+ 3820,
20
+ 2140
21
+ ],
22
+ "caption": "Document workspace displaying an existing Little Red Riding Hood input and rewrite."
23
+ },
24
+ {
25
+ "file": "figures/screenshot-03.png",
26
+ "source": "affector-arc-03.png",
27
+ "crop": [
28
+ 520,
29
+ 310,
30
+ 3820,
31
+ 2140
32
+ ],
33
+ "caption": "Geometry and weights editor with the trajectory and activation-steering controls."
34
+ },
35
+ {
36
+ "file": "figures/screenshot-04.png",
37
+ "source": "affector-arc-04.png",
38
+ "crop": [
39
+ 520,
40
+ 310,
41
+ 3820,
42
+ 920
43
+ ],
44
+ "caption": "Jobs and results controls in the offline client; the report pane has no loaded report."
45
+ }
46
+ ]
paper/TECHNICAL_PAPER.md CHANGED
@@ -9,6 +9,24 @@ Email: wtaggart@vt.edu
9
 
10
  AFFECTOR-ARC is an experimental command-line system for rewriting text under a user-defined emotional trajectory. It runs a frozen Qwen3.5-0.8B model on an NVIDIA GPU and changes selected internal activations during generation. A trajectory specifies valence, inward/outward emotional orientation, and intensity across normalized document position. Small sets of contrasting English passages establish three activation directions at each selected transformer layer. At inference time, the requested coordinates determine a weighted displacement, scaled relative to the current hidden-state magnitude, smoothed across decoding steps, and constrained by an RMS budget. Paired hooks apply this displacement to both the feed-forward input and its residual path. Document rewriting combines these interventions with four editing styles, sentence-aware segmentation, bounded generation, and incremental reporting. This report derives the mechanism from its implemented operations and distinguishes numerical controls from emotional outcomes. CPU policy tests pass; the current revision still requires CUDA validation. No measured claim of emotional accuracy, semantic preservation, or multilingual effectiveness is made.
11
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12
  ## Literature review: activation steering and emotion
13
 
14
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
@@ -92,9 +110,6 @@ That target remains constant throughout generation of the chunk. A long trajecto
92
 
93
  The runner accepts a native geometry JSON file and supported analysis JSON structures containing emotional coordinates. It recalibrates the basis on the current CUDA runtime even when the imported geometry includes one. With no geometry argument, it uses a demonstration path with three points; this path is not inferred from the input document.
94
 
95
- ![Align geometry to original document positions](figures/illustration-02.png)
96
-
97
- *Illustration 2. Trajectory alignment. The runner stretches or compresses the source path over normalized target positions and may miss unsampled emotional changes.*
98
 
99
  ## 4. Estimating activation directions
100
 
@@ -124,9 +139,6 @@ The three directions form the rows of a matrix $B_\ell\in\mathbb{R}^{3\times1024
124
 
125
  Default layers are 8, 16, and 23, using zero-based indexing, with weights 1, 1, and 0.5. This is subtraction, averaging, and normalization; there is no training optimizer, backbone fine-tuning, orthogonalization, or outcome-based fitting. The small calibration set may encode vocabulary, topic, or perspective along with emotion. Normalizing the vectors does not establish that the axes are independent or valid across languages.
126
 
127
- ![Recalibrate on the current CUDA runtime](figures/illustration-03.png)
128
-
129
- *Illustration 3. Calibration schematic. The runner recalibrates the imported trajectory on the current runtime. The contrast basis is an experimental numerical control, not a validated multilingual emotion scale.*
130
 
131
  ## 5. From coordinates to bounded displacements
132
 
@@ -192,9 +204,6 @@ A pre-hook adds $\delta$ to the normalization input; a block-output hook adds th
192
 
193
  This is an open-loop intervention. Coefficients come from the requested absolute coordinates, not the difference between a measured output emotion and the target. A zero coefficient adds no intervention on that axis; it does not actively neutralize emotion already expressed by the model.
194
 
195
- ![Separate activation controls from token sampling](figures/illustration-04.png)
196
-
197
- *Illustration 4. Two distinct control paths. Axis selection modifies activation coefficients before generation scores; token sampling operates afterward on the vocabulary distribution.*
198
 
199
  ## 6. User controls and rewrite instructions
200
 
@@ -264,9 +273,6 @@ Successful generation records include stopping reason, token count, elapsed time
264
 
265
  JSON writes use temporary-file replacement. The companion text file is written normally and may be incomplete if interrupted during that write. Reports are incremental records, not automatic resume checkpoints: rerunning starts over and overwrites the selected output paths. Use a distinct output filename for each experiment. Reports contain the input text and should be treated as document-bearing artifacts.
266
 
267
- ![A failed chunk does not end the document](figures/illustration-05.png)
268
-
269
- *Illustration 5. Output lifecycle schematic. Bounded generation, complete-prefix selection, diagnostics and per-chunk continuation address specific failure modes; they do not guarantee emotional fit or preservation of all content.*
270
 
271
  ## 9. Engineering checks and evidence
272
 
 
9
 
10
  AFFECTOR-ARC is an experimental command-line system for rewriting text under a user-defined emotional trajectory. It runs a frozen Qwen3.5-0.8B model on an NVIDIA GPU and changes selected internal activations during generation. A trajectory specifies valence, inward/outward emotional orientation, and intensity across normalized document position. Small sets of contrasting English passages establish three activation directions at each selected transformer layer. At inference time, the requested coordinates determine a weighted displacement, scaled relative to the current hidden-state magnitude, smoothed across decoding steps, and constrained by an RMS budget. Paired hooks apply this displacement to both the feed-forward input and its residual path. Document rewriting combines these interventions with four editing styles, sentence-aware segmentation, bounded generation, and incremental reporting. This report derives the mechanism from its implemented operations and distinguishes numerical controls from emotional outcomes. CPU policy tests pass; the current revision still requires CUDA validation. No measured claim of emotional accuracy, semantic preservation, or multilingual effectiveness is made.
11
 
12
+ ## Application interface
13
+
14
+ ![Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture.](figures/screenshot-01.png)
15
+
16
+ *Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture.*
17
+
18
+ ![Document workspace displaying an existing Little Red Riding Hood input and rewrite.](figures/screenshot-02.png)
19
+
20
+ *Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
21
+
22
+ ![Geometry and weights editor with the trajectory and activation-steering controls.](figures/screenshot-03.png)
23
+
24
+ *Geometry and weights editor with the trajectory and activation-steering controls.*
25
+
26
+ ![Jobs and results controls in the offline client; the report pane has no loaded report.](figures/screenshot-04.png)
27
+
28
+ *Jobs and results controls in the offline client; the report pane has no loaded report.*
29
+
30
  ## Literature review: activation steering and emotion
31
 
32
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
 
110
 
111
  The runner accepts a native geometry JSON file and supported analysis JSON structures containing emotional coordinates. It recalibrates the basis on the current CUDA runtime even when the imported geometry includes one. With no geometry argument, it uses a demonstration path with three points; this path is not inferred from the input document.
112
 
 
 
 
113
 
114
  ## 4. Estimating activation directions
115
 
 
139
 
140
  Default layers are 8, 16, and 23, using zero-based indexing, with weights 1, 1, and 0.5. This is subtraction, averaging, and normalization; there is no training optimizer, backbone fine-tuning, orthogonalization, or outcome-based fitting. The small calibration set may encode vocabulary, topic, or perspective along with emotion. Normalizing the vectors does not establish that the axes are independent or valid across languages.
141
 
 
 
 
142
 
143
  ## 5. From coordinates to bounded displacements
144
 
 
204
 
205
  This is an open-loop intervention. Coefficients come from the requested absolute coordinates, not the difference between a measured output emotion and the target. A zero coefficient adds no intervention on that axis; it does not actively neutralize emotion already expressed by the model.
206
 
 
 
 
207
 
208
  ## 6. User controls and rewrite instructions
209
 
 
273
 
274
  JSON writes use temporary-file replacement. The companion text file is written normally and may be incomplete if interrupted during that write. Reports are incremental records, not automatic resume checkpoints: rerunning starts over and overwrites the selected output paths. Use a distinct output filename for each experiment. Reports contain the input text and should be treated as document-bearing artifacts.
275
 
 
 
 
276
 
277
  ## 9. Engineering checks and evidence
278
 
paper/figures/screenshot-01.png ADDED

Git LFS Details

  • SHA256: f1ce9f9c869bfaeaefa746d9055582c85abfafb3f31ee4283812c8a912286ea8
  • Pointer size: 131 Bytes
  • Size of remote file: 841 kB
paper/figures/screenshot-02.png ADDED

Git LFS Details

  • SHA256: c619eb0ac1918c2fc4ab7cb296c468f6bad718d7500364c3cd8f45498cc17d88
  • Pointer size: 132 Bytes
  • Size of remote file: 1.01 MB
paper/figures/screenshot-03.png ADDED

Git LFS Details

  • SHA256: d9c6a611bd77fbe6a350b35f9f3302be08691cde563adbf84aab0e4d67817ea8
  • Pointer size: 131 Bytes
  • Size of remote file: 565 kB
paper/figures/screenshot-04.png ADDED

Git LFS Details

  • SHA256: 8f7b3e46a8fd299cd46a93b84aff8c693a61e9a632af44c9e91c7a495dee224f
  • Pointer size: 131 Bytes
  • Size of remote file: 146 kB
paper/source.md CHANGED
@@ -9,6 +9,24 @@ Email: wtaggart@vt.edu
9
 
10
  AFFECTOR-ARC is an experimental command-line system for rewriting text under a user-defined emotional trajectory. It runs a frozen Qwen3.5-0.8B model on an NVIDIA GPU and changes selected internal activations during generation. A trajectory specifies valence, inward/outward emotional orientation, and intensity across normalized document position. Small sets of contrasting English passages establish three activation directions at each selected transformer layer. At inference time, the requested coordinates determine a weighted displacement, scaled relative to the current hidden-state magnitude, smoothed across decoding steps, and constrained by an RMS budget. Paired hooks apply this displacement to both the feed-forward input and its residual path. Document rewriting combines these interventions with four editing styles, sentence-aware segmentation, bounded generation, and incremental reporting. This report derives the mechanism from its implemented operations and distinguishes numerical controls from emotional outcomes. CPU policy tests pass; the current revision still requires CUDA validation. No measured claim of emotional accuracy, semantic preservation, or multilingual effectiveness is made.
11
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
12
  ## Literature review: activation steering and emotion
13
 
14
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
@@ -92,9 +110,6 @@ That target remains constant throughout generation of the chunk. A long trajecto
92
 
93
  The runner accepts a native geometry JSON file and supported analysis JSON structures containing emotional coordinates. It recalibrates the basis on the current CUDA runtime even when the imported geometry includes one. With no geometry argument, it uses a demonstration path with three points; this path is not inferred from the input document.
94
 
95
- ![Align geometry to original document positions](figures/illustration-02.png)
96
-
97
- *Illustration 2. Trajectory alignment. The runner stretches or compresses the source path over normalized target positions and may miss unsampled emotional changes.*
98
 
99
  ## 4. Estimating activation directions
100
 
@@ -124,9 +139,6 @@ The three directions form the rows of a matrix $B_\ell\in\mathbb{R}^{3\times1024
124
 
125
  Default layers are 8, 16, and 23, using zero-based indexing, with weights 1, 1, and 0.5. This is subtraction, averaging, and normalization; there is no training optimizer, backbone fine-tuning, orthogonalization, or outcome-based fitting. The small calibration set may encode vocabulary, topic, or perspective along with emotion. Normalizing the vectors does not establish that the axes are independent or valid across languages.
126
 
127
- ![Recalibrate on the current CUDA runtime](figures/illustration-03.png)
128
-
129
- *Illustration 3. Calibration schematic. The runner recalibrates the imported trajectory on the current runtime. The contrast basis is an experimental numerical control, not a validated multilingual emotion scale.*
130
 
131
  ## 5. From coordinates to bounded displacements
132
 
@@ -192,9 +204,6 @@ A pre-hook adds $\delta$ to the normalization input; a block-output hook adds th
192
 
193
  This is an open-loop intervention. Coefficients come from the requested absolute coordinates, not the difference between a measured output emotion and the target. A zero coefficient adds no intervention on that axis; it does not actively neutralize emotion already expressed by the model.
194
 
195
- ![Separate activation controls from token sampling](figures/illustration-04.png)
196
-
197
- *Illustration 4. Two distinct control paths. Axis selection modifies activation coefficients before generation scores; token sampling operates afterward on the vocabulary distribution.*
198
 
199
  ## 6. User controls and rewrite instructions
200
 
@@ -264,9 +273,6 @@ Successful generation records include stopping reason, token count, elapsed time
264
 
265
  JSON writes use temporary-file replacement. The companion text file is written normally and may be incomplete if interrupted during that write. Reports are incremental records, not automatic resume checkpoints: rerunning starts over and overwrites the selected output paths. Use a distinct output filename for each experiment. Reports contain the input text and should be treated as document-bearing artifacts.
266
 
267
- ![A failed chunk does not end the document](figures/illustration-05.png)
268
-
269
- *Illustration 5. Output lifecycle schematic. Bounded generation, complete-prefix selection, diagnostics and per-chunk continuation address specific failure modes; they do not guarantee emotional fit or preservation of all content.*
270
 
271
  ## 9. Engineering checks and evidence
272
 
 
9
 
10
  AFFECTOR-ARC is an experimental command-line system for rewriting text under a user-defined emotional trajectory. It runs a frozen Qwen3.5-0.8B model on an NVIDIA GPU and changes selected internal activations during generation. A trajectory specifies valence, inward/outward emotional orientation, and intensity across normalized document position. Small sets of contrasting English passages establish three activation directions at each selected transformer layer. At inference time, the requested coordinates determine a weighted displacement, scaled relative to the current hidden-state magnitude, smoothed across decoding steps, and constrained by an RMS budget. Paired hooks apply this displacement to both the feed-forward input and its residual path. Document rewriting combines these interventions with four editing styles, sentence-aware segmentation, bounded generation, and incremental reporting. This report derives the mechanism from its implemented operations and distinguishes numerical controls from emotional outcomes. CPU policy tests pass; the current revision still requires CUDA validation. No measured claim of emotional accuracy, semantic preservation, or multilingual effectiveness is made.
11
 
12
+ ## Application interface
13
+
14
+ ![Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture.](figures/screenshot-01.png)
15
+
16
+ *Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture.*
17
+
18
+ ![Document workspace displaying an existing Little Red Riding Hood input and rewrite.](figures/screenshot-02.png)
19
+
20
+ *Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
21
+
22
+ ![Geometry and weights editor with the trajectory and activation-steering controls.](figures/screenshot-03.png)
23
+
24
+ *Geometry and weights editor with the trajectory and activation-steering controls.*
25
+
26
+ ![Jobs and results controls in the offline client; the report pane has no loaded report.](figures/screenshot-04.png)
27
+
28
+ *Jobs and results controls in the offline client; the report pane has no loaded report.*
29
+
30
  ## Literature review: activation steering and emotion
31
 
32
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
 
110
 
111
  The runner accepts a native geometry JSON file and supported analysis JSON structures containing emotional coordinates. It recalibrates the basis on the current CUDA runtime even when the imported geometry includes one. With no geometry argument, it uses a demonstration path with three points; this path is not inferred from the input document.
112
 
 
 
 
113
 
114
  ## 4. Estimating activation directions
115
 
 
139
 
140
  Default layers are 8, 16, and 23, using zero-based indexing, with weights 1, 1, and 0.5. This is subtraction, averaging, and normalization; there is no training optimizer, backbone fine-tuning, orthogonalization, or outcome-based fitting. The small calibration set may encode vocabulary, topic, or perspective along with emotion. Normalizing the vectors does not establish that the axes are independent or valid across languages.
141
 
 
 
 
142
 
143
  ## 5. From coordinates to bounded displacements
144
 
 
204
 
205
  This is an open-loop intervention. Coefficients come from the requested absolute coordinates, not the difference between a measured output emotion and the target. A zero coefficient adds no intervention on that axis; it does not actively neutralize emotion already expressed by the model.
206
 
 
 
 
207
 
208
  ## 6. User controls and rewrite instructions
209
 
 
273
 
274
  JSON writes use temporary-file replacement. The companion text file is written normally and may be incomplete if interrupted during that write. Reports are incremental records, not automatic resume checkpoints: rerunning starts over and overwrites the selected output paths. Use a distinct output filename for each experiment. Reports contain the input text and should be treated as document-bearing artifacts.
275
 
 
 
 
276
 
277
  ## 9. Engineering checks and evidence
278