Illustrate technical papers with authentic application screenshots
Browse files- .gitattributes +4 -0
- TECHNICAL_REPORT.md +19 -13
- paper/AFFECTOR-ARC-ACM-source.zip +2 -2
- paper/AFFECTOR-ARC-ACM.pdf +2 -2
- paper/AFFECTOR-ARC-ACM.tex +36 -29
- paper/ILLUSTRATIONS.json +51 -37
- paper/PAPER-SHA256SUMS.txt +6 -6
- paper/SCREENSHOTS.json +46 -0
- paper/TECHNICAL_PAPER.md +18 -12
- paper/figures/screenshot-01.png +3 -0
- paper/figures/screenshot-02.png +3 -0
- paper/figures/screenshot-03.png +3 -0
- paper/figures/screenshot-04.png +3 -0
- paper/source.md +18 -12
.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 |
-

|
| 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 |
-

|
| 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 |
-

|
| 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 |
-

|
| 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](
|
| 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 |
+

|
| 15 |
+
|
| 16 |
+
*Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture.*
|
| 17 |
+
|
| 18 |
+

|
| 19 |
+
|
| 20 |
+
*Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
|
| 21 |
+
|
| 22 |
+

|
| 23 |
+
|
| 24 |
+
*Geometry and weights editor with the trajectory and activation-steering controls.*
|
| 25 |
+
|
| 26 |
+

|
| 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:
|
| 3 |
-
size
|
|
|
|
| 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:
|
| 3 |
-
size
|
|
|
|
| 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 |
-
|
| 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 |
-
|
| 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 |
-
|
| 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 |
-
|
| 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 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 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 |
-
|
| 2 |
d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-ARC-ACM.bbl
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 LITERATURE_SOURCES.md
|
| 7 |
b9375d5aa1da63a0a2773da47c8dce77eaad36a34265f9b36a569671e275ad9b README.md
|
| 8 |
-
|
| 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 |
-
|
|
|
|
| 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 |
-

|
| 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 |
-

|
| 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 |
-

|
| 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 |
-

|
| 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 |
+

|
| 15 |
+
|
| 16 |
+
*Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture.*
|
| 17 |
+
|
| 18 |
+

|
| 19 |
+
|
| 20 |
+
*Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
|
| 21 |
+
|
| 22 |
+

|
| 23 |
+
|
| 24 |
+
*Geometry and weights editor with the trajectory and activation-steering controls.*
|
| 25 |
+
|
| 26 |
+

|
| 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
|
paper/figures/screenshot-02.png
ADDED
|
Git LFS Details
|
paper/figures/screenshot-03.png
ADDED
|
Git LFS Details
|
paper/figures/screenshot-04.png
ADDED
|
Git LFS Details
|
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 |
-

|
| 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 |
-

|
| 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 |
-

|
| 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 |
-

|
| 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 |
+

|
| 15 |
+
|
| 16 |
+
*Connection workspace with SSH settings and the Slurm GPU resource request. The client is offline in this capture.*
|
| 17 |
+
|
| 18 |
+

|
| 19 |
+
|
| 20 |
+
*Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
|
| 21 |
+
|
| 22 |
+

|
| 23 |
+
|
| 24 |
+
*Geometry and weights editor with the trajectory and activation-steering controls.*
|
| 25 |
+
|
| 26 |
+

|
| 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 |
|