Describe native macOS workflow and refine research context
Browse files- TECHNICAL_REPORT.md +27 -10
- paper/AFFECTOR-ARC-ACM-source.zip +2 -2
- paper/AFFECTOR-ARC-ACM.pdf +2 -2
- paper/AFFECTOR-ARC-ACM.tex +29 -29
- paper/PAPER-SHA256SUMS.txt +5 -5
- paper/TECHNICAL_PAPER.md +27 -10
- paper/source.md +27 -10
TECHNICAL_REPORT.md
CHANGED
|
@@ -3,32 +3,42 @@
|
|
| 3 |
**William Taggart**
|
| 4 |
Modern and Classical Languages and Literatures, Virginia Tech
|
| 5 |
Email: wtaggart@vt.edu
|
| 6 |
-
**Technical report · November
|
| 7 |
|
| 8 |
## Abstract
|
| 9 |
|
| 10 |
-
AFFECTOR-ARC is an experimental
|
| 11 |
|
| 12 |
## Purpose and Research Context
|
| 13 |
|
| 14 |
-
AFFECTOR-ARC investigates
|
| 15 |
|
| 16 |
-
The
|
|
|
|
|
|
|
| 17 |
|
| 18 |

|
| 19 |
|
| 20 |
*Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
|
| 21 |
|
| 22 |
-
|
| 23 |
|
| 24 |
-
|
| 25 |
|
| 26 |
-
|
| 27 |
|
| 28 |
-
|
| 29 |
|
| 30 |
## Runtime and data flow
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
The runner loads `Qwen/Qwen3.5-0.8B` at revision `2fc06364715b967f1860aea9cf38778875588b17`. The implementation uses `Qwen3_5ForConditionalGeneration`, its text decoder layers, float32 computation, eager attention, and the first process-visible CUDA device. It checks for 24 decoder layers; the geometry and calibration code assume hidden width 1,024. The backbone is in evaluation mode, its parameters have gradients disabled, and calibration and generation run without gradient tracking.
|
| 33 |
|
| 34 |

|
|
@@ -248,6 +258,8 @@ Successful generation records include stopping reason, token count, elapsed time
|
|
| 248 |
|
| 249 |
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.
|
| 250 |
|
|
|
|
|
|
|
| 251 |
## Engineering checks and evidence
|
| 252 |
|
| 253 |
Smoke mode is the default and operates on one supplied passage. It runs baseline, zero-strength, positive-valence, negative-valence, and restored-baseline conditions. Positive and negative conditions use coefficient vectors $(0.8,0,0)$ and $(-0.8,0,0)$. The third coordinate zero corresponds to intensity 0.5.
|
|
@@ -258,9 +270,13 @@ Smoke mode defaults to the original 128-token allowance and 600-second time thre
|
|
| 258 |
|
| 259 |
For the November 13 revision, seven CPU policy tests passed, Python and shell syntax were checked, and CLI loading was verified. Tests cover such behavior as unchanged requested controls, paragraph allowance accounting, draft retention, sentence cleanup, and continuing after a failed chunk. They do not execute the CUDA model. The current revision has not yet been run on an ARC GPU. A successful GPU run was reported for an earlier revision on November 12; it is historical evidence, not validation of the current code.
|
| 260 |
|
|
|
|
|
|
|
| 261 |
## Reproduction
|
| 262 |
|
| 263 |
-
|
|
|
|
|
|
|
| 264 |
|
| 265 |
```bash
|
| 266 |
bash setup.sh
|
|
@@ -298,8 +314,9 @@ Changing to a larger model requires new architecture validation, basis calibrati
|
|
| 298 |
|
| 299 |
## Implementation sources
|
| 300 |
|
| 301 |
-
This report
|
| 302 |
|
|
|
|
| 303 |
- [CUDA runner and smoke protocol](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/run_arc.py)
|
| 304 |
- [Document transformation and incremental output](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/document_run.py)
|
| 305 |
- [Calibration, displacement math, hooks, and decoding](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/engine/steering.py)
|
|
|
|
| 3 |
**William Taggart**
|
| 4 |
Modern and Classical Languages and Literatures, Virginia Tech
|
| 5 |
Email: wtaggart@vt.edu
|
| 6 |
+
**Technical report · November 14, 2026 · Revision 0.3**
|
| 7 |
|
| 8 |
## Abstract
|
| 9 |
|
| 10 |
+
AFFECTOR-ARC is an experimental text-rewriting system with a native macOS application and a command-line CUDA runner. The desktop application manages documents, emotional geometries, SSH connections, Slurm GPU jobs, and downloaded results; the runner executes a frozen Qwen3.5-0.8B model on an allocated ARC GPU. A trajectory specifies valence, inward/outward emotional orientation, and intensity across normalized document position. Contrasting English passages establish three activation directions at each selected transformer layer. During generation, trajectory coordinates determine a displacement scaled to the hidden-state magnitude, smoothed across decoding steps, and constrained by an RMS budget. Paired hooks apply it to the feed-forward input and residual path. Four editing styles, sentence-aware segmentation, bounded generation, and incremental reports make the intervention available within a document workflow. This report describes both interfaces and derives the steering mechanism, while distinguishing implementation checks from evidence of emotional effectiveness. Earlier desktop GPU runs and CPU policy tests are documented separately; they do not validate every subsequent runner revision or establish emotional accuracy, semantic preservation, or multilingual effectiveness.
|
| 11 |
|
| 12 |
## Purpose and Research Context
|
| 13 |
|
| 14 |
+
AFFECTOR-ARC investigates whether a writer can guide the emotional progression of a document by specifying a trajectory and intervening in a language model's internal computation. The practical aim is to make that experiment usable across a complete rewrite: the writer selects a text, supplies an emotional geometry, adjusts the intervention, and inspects the resulting passages and diagnostics. A native macOS application brings these activities into a workspace, while SSH and Slurm connect that workspace to a GPU allocation on Virginia Tech ARC. A command-line runner exposes the computation separately for scripted experiments and reproducible execution.
|
| 15 |
|
| 16 |
+
The geometry gives this workflow a common representation. It is an ordered path through valence, inward/outward emotional orientation, and intensity, accompanied by layer settings and optional calibration matrices. Sampling the path at successive document positions supplies the coordinates used during rewriting. Those coordinates become displacements of the model's hidden states, so the desired progression acts on the computation that produces tokens rather than prescribing particular words. The model's stored parameters remain fixed throughout.
|
| 17 |
+
|
| 18 |
+
This mechanism builds on contrastive activation steering. Turner and colleagues derive directions from contrasting prompts, and Panickssery and colleagues average differences between positive and negative behavioral examples ([Turner et al., 2023](https://arxiv.org/abs/2308.10248v1); [Panickssery et al., 2024](https://arxiv.org/abs/2312.06681v4)). AFFECTOR-ARC applies the same broad principle to small sets of emotionally contrasting passages, then mixes the resulting directions according to the requested trajectory. The central research question is how reliably this mapping can influence a sustained rewrite, given that a numerical displacement need not produce a unique or consistent emotional interpretation.
|
| 19 |
|
| 20 |

|
| 21 |
|
| 22 |
*Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
|
| 23 |
|
| 24 |
+
That question connects the implementation to research on the relation between model representations and human judgments. Wu and colleagues study concept-based emotional steering across linguistic-cultural settings, while Diallo and colleagues use human ratings to examine perceived emotional intensity and text quality ([Wu et al., 2025](https://arxiv.org/abs/2506.13978v1); [Diallo et al., 2026](https://arxiv.org/abs/2601.21505v1)). These studies motivate coordinate-based experiments whose outcomes are judged through readers' interpretations. They do not, however, establish the validity of this system's inward/outward axis or calibrate its RMS-scaled controls. The requested trajectory, the activation displacement, and the emotion perceived in the output must therefore be evaluated as related but distinct quantities.
|
| 25 |
|
| 26 |
+
The reliability of that relationship also depends on where and how the intervention is applied. Shu and colleagues find uneven causal organization across emotion categories, and Wang and colleagues show that strong probing accuracy need not identify the most effective intervention layers ([Shu et al., 2026](https://arxiv.org/abs/2604.25866v3); [Wang et al., 2026](https://arxiv.org/abs/2609.22135v1)). Ha and colleagues' finding that supportive-empathy directions are only partially separable further cautions against treating emotional axes as independent controls ([Ha et al., 2026](https://arxiv.org/abs/2609.15654v1)). Automated vector extraction and context-dependent strength offer approaches to these calibration problems ([Cui et al., 2025](https://arxiv.org/abs/2509.22739v3); [Ferrando et al., 2025](https://arxiv.org/abs/2512.03661v2)). AFFECTOR-ARC makes layer weights and axis mixtures explicit so that such questions can be investigated, although its current runner uses user-selected settings rather than an adaptive emotional-feedback controller.
|
| 27 |
|
| 28 |
+
Applying those settings across a document makes generation stability part of the same problem. E-STEER reports non-monotonic effects of emotional interventions, and Jeong observes both emotional changes and degeneration in small models ([Sun et al., 2026](https://arxiv.org/abs/2604.00005v1); [Jeong, 2026](https://arxiv.org/abs/2604.04064v1)). A stronger displacement can thus disrupt a rewrite without bringing it closer to the intended trajectory. The system addresses the operational consequences through bounded generation, repetition checks, passage-level failure records, and a desktop workflow for reviewing jobs and results. These safeguards preserve an inspectable experiment; they do not establish that a completed rewrite meets its emotional target.
|
| 29 |
|
| 30 |
+
Evaluation must consequently consider emotional alignment together with fluency, content change, and failure rates. SteerScope examines efficacy, side effects, generalization, and dependence on steering data, finding no general advantage over prompt steering in its tested trade-offs ([Yang et al., 2026](https://arxiv.org/abs/2610.07722v1)). For AFFECTOR-ARC, this motivates comparisons with prompt-only and unsteered rewriting, blinded emotional and quality ratings, and complete reporting of retained originals, repetition, and truncation. The scope of this report is the desktop-to-cluster workflow and the implemented steering procedure, including their engineering checks. Establishing emotional effectiveness and a reliable operating range remains an empirical task.
|
| 31 |
|
| 32 |
## Runtime and data flow
|
| 33 |
|
| 34 |
+
### Native macOS client and remote execution
|
| 35 |
+
|
| 36 |
+
The native SwiftUI application is distributed on Hugging Face for Apple Silicon Macs running macOS 15 or later. Its connection workspace configures SSH access and the Slurm resource request. The document workspace holds the input and generated text; the shape and weights workspace supports importing, converting, and editing geometries and intervention settings. Jobs and results controls expose remote progress, diagnostics, and output downloads. The application also manages Hugging Face source updates and submits GPU-environment setup through Slurm.
|
| 37 |
+
|
| 38 |
+
The Mac serves as the client: lightweight SSH requests manage files and jobs, while model setup and execution run in requested GPU allocations. The bundled Python remote agent handles requests over SSH without a persistent listening server. The desktop interface and the command-line runner are separately versioned components, so a recorded application version alone does not identify the remote engine revision. Both should be recorded when reporting a run.
|
| 39 |
+
|
| 40 |
+
### CUDA runner
|
| 41 |
+
|
| 42 |
The runner loads `Qwen/Qwen3.5-0.8B` at revision `2fc06364715b967f1860aea9cf38778875588b17`. The implementation uses `Qwen3_5ForConditionalGeneration`, its text decoder layers, float32 computation, eager attention, and the first process-visible CUDA device. It checks for 24 decoder layers; the geometry and calibration code assume hidden width 1,024. The backbone is in evaluation mode, its parameters have gradients disabled, and calibration and generation run without gradient tracking.
|
| 43 |
|
| 44 |

|
|
|
|
| 258 |
|
| 259 |
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.
|
| 260 |
|
| 261 |
+
The desktop document view is a presentation layer over these artifacts. The published client displays generated passages separately from source text and diagnostics; its text exports can therefore differ from the runner's assembled file, which includes retained originals. The downloadable report is needed to distinguish generated, retained, failed, and unprocessed passages. Reconnecting can recover an existing job and its outputs, but this client recovery does not make the runner's incremental report a resumable generation checkpoint.
|
| 262 |
+
|
| 263 |
## Engineering checks and evidence
|
| 264 |
|
| 265 |
Smoke mode is the default and operates on one supplied passage. It runs baseline, zero-strength, positive-valence, negative-valence, and restored-baseline conditions. Positive and negative conditions use coefficient vectors $(0.8,0,0)$ and $(-0.8,0,0)$. The third coordinate zero corresponds to intensity 0.5.
|
|
|
|
| 270 |
|
| 271 |
For the November 13 revision, seven CPU policy tests passed, Python and shell syntax were checked, and CLI loading was verified. Tests cover such behavior as unchanged requested controls, paragraph allowance accounting, draft retention, sentence cleanup, and continuing after a failed chunk. They do not execute the CUDA model. The current revision has not yet been run on an ARC GPU. A successful GPU run was reported for an earlier revision on November 12; it is historical evidence, not validation of the current code.
|
| 272 |
|
| 273 |
+
The desktop release also has its own validation record. Its published version 1.0.4 documentation reports native Swift tests, 14 remote-agent tests, a completed sample rewrite, and an earlier GPU setup on an NVIDIA A100-SXM4-80GB. These are release-specific engineering observations, not measurements of emotional accuracy or validation of later changes to the remote runner. The screenshots show the native application's workspaces, including offline states where indicated.
|
| 274 |
+
|
| 275 |
## Reproduction
|
| 276 |
|
| 277 |
+
For the desktop workflow, download the native macOS application from the repository's macos directory on Hugging Face, unpack it, and open AFFECTOR ARC.app. Configure the connection profile for the user's own ARC login, SSH authentication, and Slurm allocation. After connecting, install or update the remote source and prepare the GPU environment as needed. Import a document and geometry, adjust the settings, and choose Transform on ARC. Monitor the submitted job in Jobs + results and download its input, output, report, geometry, and logs. The published desktop README identifies the application as version 1.0.4 and describes its build as ad-hoc signed.
|
| 278 |
+
|
| 279 |
+
The following commands provide the separate command-line workflow. They run inside a suitable ARC GPU allocation from a fresh repository copy:
|
| 280 |
|
| 281 |
```bash
|
| 282 |
bash setup.sh
|
|
|
|
| 314 |
|
| 315 |
## Implementation sources
|
| 316 |
|
| 317 |
+
This report covers the native macOS client, its SSH request agent, and the CUDA runner. The desktop release and its source are available alongside the remote engine:
|
| 318 |
|
| 319 |
+
- [Native macOS application, Swift source, and release documentation](https://huggingface.co/will-taggart/AFFECTOR-ARC/tree/main/macos)
|
| 320 |
- [CUDA runner and smoke protocol](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/run_arc.py)
|
| 321 |
- [Document transformation and incremental output](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/document_run.py)
|
| 322 |
- [Calibration, displacement math, hooks, and decoding](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/engine/steering.py)
|
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 sha256:761494d17669ce0ca24001000bfa75f1376c1a3a8007c8ef87c6d1df1ad9f0e9
|
| 3 |
+
size 5213478
|
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:faf072babff6b3f33eb7160e0c03b6b7bc4316ec8844aa0230785c0f9e8f4343
|
| 3 |
+
size 2481206
|
paper/AFFECTOR-ARC-ACM.tex
CHANGED
|
@@ -35,34 +35,19 @@ Activation Steering}
|
|
| 35 |
\affiliation{\department{Modern and Classical Languages and Literatures}
|
| 36 |
\institution{Virginia Tech}\city{Blacksburg}\state{Virginia}\country{USA}}
|
| 37 |
\email{wtaggart@vt.edu}
|
| 38 |
-
\subtitle{Technical Report: November
|
| 39 |
\begin{abstract}
|
| 40 |
-
AFFECTOR-ARC is an experimental command-line
|
| 41 |
-
under a user-defined emotional trajectory. It runs a frozen Qwen3.5-0.8B
|
| 42 |
-
model on an NVIDIA GPU and changes selected internal activations during
|
| 43 |
-
generation. A trajectory specifies valence, inward/outward emotional
|
| 44 |
-
orientation, and intensity across normalized document position. Small
|
| 45 |
-
sets of contrasting English passages establish three activation
|
| 46 |
-
directions at each selected transformer layer. At inference time, the
|
| 47 |
-
requested coordinates determine a weighted displacement, scaled relative
|
| 48 |
-
to the current hidden-state magnitude, smoothed across decoding steps,
|
| 49 |
-
and constrained by an RMS budget. Paired hooks apply this displacement
|
| 50 |
-
to both the feed-forward input and its residual path. Document rewriting
|
| 51 |
-
combines these interventions with four editing styles, sentence-aware
|
| 52 |
-
segmentation, bounded generation, and incremental reporting. This report
|
| 53 |
-
derives the mechanism from its implemented operations and distinguishes
|
| 54 |
-
numerical controls from emotional outcomes. CPU policy tests pass; the
|
| 55 |
-
current revision still requires CUDA validation. No measured claim of
|
| 56 |
-
emotional accuracy, semantic preservation, or multilingual effectiveness
|
| 57 |
-
is made.
|
| 58 |
\end{abstract}
|
| 59 |
\keywords{activation steering, emotional text generation, transformer models, controllable generation}
|
| 60 |
\maketitle
|
| 61 |
\section{Purpose and Research Context}\label{literature-review-activation-steering-and-emotion}\label{problem-and-operating-principle}
|
| 62 |
|
| 63 |
-
AFFECTOR-ARC investigates
|
| 64 |
|
| 65 |
-
The
|
|
|
|
|
|
|
| 66 |
|
| 67 |
\begin{figure}[!htb]
|
| 68 |
\centering
|
|
@@ -72,16 +57,24 @@ The intervention acts on the numerical hidden states used to predict successive
|
|
| 72 |
\label{fig:app-screen-2}
|
| 73 |
\end{figure}
|
| 74 |
|
| 75 |
-
|
| 76 |
|
| 77 |
-
|
| 78 |
|
| 79 |
-
|
| 80 |
|
| 81 |
-
|
| 82 |
|
| 83 |
\section{Runtime and data flow}\label{runtime-and-data-flow}
|
| 84 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 85 |
|
| 86 |
\begin{figure}[!htb]
|
| 87 |
\centering
|
|
@@ -527,6 +520,8 @@ distinct output filename for each experiment. Reports contain the input
|
|
| 527 |
text and should be treated as document-bearing artifacts.
|
| 528 |
|
| 529 |
|
|
|
|
|
|
|
| 530 |
\section{Engineering checks and
|
| 531 |
evidence}\label{engineering-checks-and-evidence}
|
| 532 |
|
|
@@ -564,9 +559,13 @@ has not yet been run on an ARC GPU. A successful GPU run was reported
|
|
| 564 |
for an earlier revision on November 12; it is historical evidence, not
|
| 565 |
validation of the current code.
|
| 566 |
|
|
|
|
|
|
|
| 567 |
\section{Reproduction}\label{reproduction}
|
| 568 |
|
| 569 |
-
|
|
|
|
|
|
|
| 570 |
|
| 571 |
\begin{verbatim}
|
| 572 |
bash setup.sh
|
|
@@ -641,11 +640,12 @@ prompt wording alone does not establish it.
|
|
| 641 |
|
| 642 |
\section{Implementation sources}\label{implementation-sources}
|
| 643 |
|
| 644 |
-
This report
|
| 645 |
-
relying on another application's interface or documentation.
|
| 646 |
|
| 647 |
\begin{itemize}
|
| 648 |
\tightlist
|
|
|
|
|
|
|
| 649 |
\item
|
| 650 |
\href{https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/run_arc.py}{CUDA
|
| 651 |
runner and smoke protocol}
|
|
@@ -692,7 +692,7 @@ and limitations are stated separately in this report.
|
|
| 692 |
\section*{Suggested citation}
|
| 693 |
William Taggart. 2026. AFFECTOR-ARC: Emotional Trajectory Control
|
| 694 |
through Transformer Activation Steering. Technical report. Modern and
|
| 695 |
-
Classical Languages and Literatures, Virginia Tech. November
|
| 696 |
\url{https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/TECHNICAL_REPORT.md}
|
| 697 |
|
| 698 |
\end{document}
|
|
|
|
| 35 |
\affiliation{\department{Modern and Classical Languages and Literatures}
|
| 36 |
\institution{Virginia Tech}\city{Blacksburg}\state{Virginia}\country{USA}}
|
| 37 |
\email{wtaggart@vt.edu}
|
| 38 |
+
\subtitle{Technical Report: November 14, 2026}
|
| 39 |
\begin{abstract}
|
| 40 |
+
AFFECTOR-ARC is an experimental text-rewriting system with a native macOS application and a command-line CUDA runner. The desktop application manages documents, emotional geometries, SSH connections, Slurm GPU jobs, and downloaded results; the runner executes a frozen Qwen3.5-0.8B model on an allocated ARC GPU. A trajectory specifies valence, inward/outward emotional orientation, and intensity across normalized document position. Contrasting English passages establish three activation directions at each selected transformer layer. During generation, trajectory coordinates determine a displacement scaled to the hidden-state magnitude, smoothed across decoding steps, and constrained by an RMS budget. Paired hooks apply it to the feed-forward input and residual path. Four editing styles, sentence-aware segmentation, bounded generation, and incremental reports make the intervention available within a document workflow. This report describes both interfaces and derives the steering mechanism, while distinguishing implementation checks from evidence of emotional effectiveness. Earlier desktop GPU runs and CPU policy tests are documented separately; they do not validate every subsequent runner revision or establish emotional accuracy, semantic preservation, or multilingual effectiveness.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
\end{abstract}
|
| 42 |
\keywords{activation steering, emotional text generation, transformer models, controllable generation}
|
| 43 |
\maketitle
|
| 44 |
\section{Purpose and Research Context}\label{literature-review-activation-steering-and-emotion}\label{problem-and-operating-principle}
|
| 45 |
|
| 46 |
+
AFFECTOR-ARC investigates whether a writer can guide the emotional progression of a document by specifying a trajectory and intervening in a language model's internal computation. The practical aim is to make that experiment usable across a complete rewrite: the writer selects a text, supplies an emotional geometry, adjusts the intervention, and inspects the resulting passages and diagnostics. A native macOS application brings these activities into a workspace, while SSH and Slurm connect that workspace to a GPU allocation on Virginia Tech ARC. A command-line runner exposes the computation separately for scripted experiments and reproducible execution.
|
| 47 |
|
| 48 |
+
The geometry gives this workflow a common representation. It is an ordered path through valence, inward/outward emotional orientation, and intensity, accompanied by layer settings and optional calibration matrices. Sampling the path at successive document positions supplies the coordinates used during rewriting. Those coordinates become displacements of the model's hidden states, so the desired progression acts on the computation that produces tokens rather than prescribing particular words. The model's stored parameters remain fixed throughout.
|
| 49 |
+
|
| 50 |
+
This mechanism builds on contrastive activation steering. Turner and colleagues derive directions from contrasting prompts, and Panickssery and colleagues average differences between positive and negative behavioral examples \cite{turner2023actadd,panickssery2024caa}. AFFECTOR-ARC applies the same broad principle to small sets of emotionally contrasting passages, then mixes the resulting directions according to the requested trajectory. The central research question is how reliably this mapping can influence a sustained rewrite, given that a numerical displacement need not produce a unique or consistent emotional interpretation.
|
| 51 |
|
| 52 |
\begin{figure}[!htb]
|
| 53 |
\centering
|
|
|
|
| 57 |
\label{fig:app-screen-2}
|
| 58 |
\end{figure}
|
| 59 |
|
| 60 |
+
That question connects the implementation to research on the relation between model representations and human judgments. Wu and colleagues study concept-based emotional steering across linguistic-cultural settings, while Diallo and colleagues use human ratings to examine perceived emotional intensity and text quality \cite{wu2025emotion,diallo2026style}. These studies motivate coordinate-based experiments whose outcomes are judged through readers' interpretations. They do not, however, establish the validity of this system's inward/outward axis or calibrate its RMS-scaled controls. The requested trajectory, the activation displacement, and the emotion perceived in the output must therefore be evaluated as related but distinct quantities.
|
| 61 |
|
| 62 |
+
The reliability of that relationship also depends on where and how the intervention is applied. Shu and colleagues find uneven causal organization across emotion categories, and Wang and colleagues show that strong probing accuracy need not identify the most effective intervention layers \cite{shu2026emotion,wang2026readbest}. Ha and colleagues' finding that supportive-empathy directions are only partially separable further cautions against treating emotional axes as independent controls \cite{ha2026empathy}. Automated vector extraction and context-dependent strength offer approaches to these calibration problems \cite{cui2025pas,ferrando2025dsas}. AFFECTOR-ARC makes layer weights and axis mixtures explicit so that such questions can be investigated, although its current runner uses user-selected settings rather than an adaptive emotional-feedback controller.
|
| 63 |
|
| 64 |
+
Applying those settings across a document makes generation stability part of the same problem. E-STEER reports non-monotonic effects of emotional interventions, and Jeong observes both emotional changes and degeneration in small models \cite{sun2026emotion,jeong2026emotion}. A stronger displacement can thus disrupt a rewrite without bringing it closer to the intended trajectory. The system addresses the operational consequences through bounded generation, repetition checks, passage-level failure records, and a desktop workflow for reviewing jobs and results. These safeguards preserve an inspectable experiment; they do not establish that a completed rewrite meets its emotional target.
|
| 65 |
|
| 66 |
+
Evaluation must consequently consider emotional alignment together with fluency, content change, and failure rates. SteerScope examines efficacy, side effects, generalization, and dependence on steering data, finding no general advantage over prompt steering in its tested trade-offs \cite{yang2026steerscope}. For AFFECTOR-ARC, this motivates comparisons with prompt-only and unsteered rewriting, blinded emotional and quality ratings, and complete reporting of retained originals, repetition, and truncation. The scope of this report is the desktop-to-cluster workflow and the implemented steering procedure, including their engineering checks. Establishing emotional effectiveness and a reliable operating range remains an empirical task.
|
| 67 |
|
| 68 |
\section{Runtime and data flow}\label{runtime-and-data-flow}
|
| 69 |
|
| 70 |
+
\subsection{Native macOS client and remote execution}
|
| 71 |
+
|
| 72 |
+
The native SwiftUI application is distributed on Hugging Face for Apple Silicon Macs running macOS 15 or later. Its connection workspace configures SSH access and the Slurm resource request. The document workspace holds the input and generated text; the shape and weights workspace supports importing, converting, and editing geometries and intervention settings. Jobs and results controls expose remote progress, diagnostics, and output downloads. The application also manages Hugging Face source updates and submits GPU-environment setup through Slurm.
|
| 73 |
+
|
| 74 |
+
The Mac serves as the client: lightweight SSH requests manage files and jobs, while model setup and execution run in requested GPU allocations. The bundled Python remote agent handles requests over SSH without a persistent listening server. The desktop interface and the command-line runner are separately versioned components, so a recorded application version alone does not identify the remote engine revision. Both should be recorded when reporting a run.
|
| 75 |
+
|
| 76 |
+
\subsection{CUDA runner}
|
| 77 |
+
|
| 78 |
|
| 79 |
\begin{figure}[!htb]
|
| 80 |
\centering
|
|
|
|
| 520 |
text and should be treated as document-bearing artifacts.
|
| 521 |
|
| 522 |
|
| 523 |
+
The desktop document view is a presentation layer over these artifacts. The published client displays generated passages separately from source text and diagnostics; its text exports can therefore differ from the runner's assembled file, which includes retained originals. The downloadable report is needed to distinguish generated, retained, failed, and unprocessed passages. Reconnecting can recover an existing job and its outputs, but this client recovery does not make the runner's incremental report a resumable generation checkpoint.
|
| 524 |
+
|
| 525 |
\section{Engineering checks and
|
| 526 |
evidence}\label{engineering-checks-and-evidence}
|
| 527 |
|
|
|
|
| 559 |
for an earlier revision on November 12; it is historical evidence, not
|
| 560 |
validation of the current code.
|
| 561 |
|
| 562 |
+
The desktop release also has its own validation record. Its published version 1.0.4 documentation reports native Swift tests, 14 remote-agent tests, a completed sample rewrite, and an earlier GPU setup on an NVIDIA A100-SXM4-80GB. These are release-specific engineering observations, not measurements of emotional accuracy or validation of later changes to the remote runner. The screenshots show the native application's workspaces, including offline states where indicated.
|
| 563 |
+
|
| 564 |
\section{Reproduction}\label{reproduction}
|
| 565 |
|
| 566 |
+
For the desktop workflow, download the native macOS application from the repository's macos directory on Hugging Face, unpack it, and open AFFECTOR ARC.app. Configure the connection profile for the user's own ARC login, SSH authentication, and Slurm allocation. After connecting, install or update the remote source and prepare the GPU environment as needed. Import a document and geometry, adjust the settings, and choose Transform on ARC. Monitor the submitted job in Jobs + results and download its input, output, report, geometry, and logs. The published desktop README identifies the application as version 1.0.4 and describes its build as ad-hoc signed.
|
| 567 |
+
|
| 568 |
+
The following commands provide the separate command-line workflow. They run inside a suitable ARC GPU allocation from a fresh repository copy:
|
| 569 |
|
| 570 |
\begin{verbatim}
|
| 571 |
bash setup.sh
|
|
|
|
| 640 |
|
| 641 |
\section{Implementation sources}\label{implementation-sources}
|
| 642 |
|
| 643 |
+
This report covers the native macOS client, its SSH request agent, and the CUDA runner. The desktop release and its source are available alongside the remote engine:
|
|
|
|
| 644 |
|
| 645 |
\begin{itemize}
|
| 646 |
\tightlist
|
| 647 |
+
\item
|
| 648 |
+
\href{https://huggingface.co/will-taggart/AFFECTOR-ARC/tree/main/macos}{Native macOS application, Swift source, and release documentation}
|
| 649 |
\item
|
| 650 |
\href{https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/run_arc.py}{CUDA
|
| 651 |
runner and smoke protocol}
|
|
|
|
| 692 |
\section*{Suggested citation}
|
| 693 |
William Taggart. 2026. AFFECTOR-ARC: Emotional Trajectory Control
|
| 694 |
through Transformer Activation Steering. Technical report. Modern and
|
| 695 |
+
Classical Languages and Literatures, Virginia Tech. November 14, 2026.
|
| 696 |
\url{https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/TECHNICAL_REPORT.md}
|
| 697 |
|
| 698 |
\end{document}
|
paper/PAPER-SHA256SUMS.txt
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
-
|
| 2 |
d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-ARC-ACM.bbl
|
| 3 |
-
|
| 4 |
-
|
| 5 |
00fe34363158541d8d41414f548a2a6a7fd73b147e025b8bc2ef042f9dba6d85 ILLUSTRATIONS.json
|
| 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 |
+
761494d17669ce0ca24001000bfa75f1376c1a3a8007c8ef87c6d1df1ad9f0e9 AFFECTOR-ARC-ACM-source.zip
|
| 2 |
d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-ARC-ACM.bbl
|
| 3 |
+
faf072babff6b3f33eb7160e0c03b6b7bc4316ec8844aa0230785c0f9e8f4343 AFFECTOR-ARC-ACM.pdf
|
| 4 |
+
b39b322e1e0ef985284751bc7be4cd9da3994c575b1986b4cd1f5c5f6c64af49 AFFECTOR-ARC-ACM.tex
|
| 5 |
00fe34363158541d8d41414f548a2a6a7fd73b147e025b8bc2ef042f9dba6d85 ILLUSTRATIONS.json
|
| 6 |
d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 LITERATURE_SOURCES.md
|
| 7 |
b9375d5aa1da63a0a2773da47c8dce77eaad36a34265f9b36a569671e275ad9b README.md
|
| 8 |
+
2ebd4f309183a1f91c471c92c92addb179c8c550ea465daddfc9c5502a0d1527 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 |
+
2ebd4f309183a1f91c471c92c92addb179c8c550ea465daddfc9c5502a0d1527 source.md
|
paper/TECHNICAL_PAPER.md
CHANGED
|
@@ -3,32 +3,42 @@
|
|
| 3 |
**William Taggart**
|
| 4 |
Modern and Classical Languages and Literatures, Virginia Tech
|
| 5 |
Email: wtaggart@vt.edu
|
| 6 |
-
**Technical report · November
|
| 7 |
|
| 8 |
## Abstract
|
| 9 |
|
| 10 |
-
AFFECTOR-ARC is an experimental
|
| 11 |
|
| 12 |
## Purpose and Research Context
|
| 13 |
|
| 14 |
-
AFFECTOR-ARC investigates
|
| 15 |
|
| 16 |
-
The
|
|
|
|
|
|
|
| 17 |
|
| 18 |

|
| 19 |
|
| 20 |
*Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
|
| 21 |
|
| 22 |
-
|
| 23 |
|
| 24 |
-
|
| 25 |
|
| 26 |
-
|
| 27 |
|
| 28 |
-
|
| 29 |
|
| 30 |
## Runtime and data flow
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
The runner loads `Qwen/Qwen3.5-0.8B` at revision `2fc06364715b967f1860aea9cf38778875588b17`. The implementation uses `Qwen3_5ForConditionalGeneration`, its text decoder layers, float32 computation, eager attention, and the first process-visible CUDA device. It checks for 24 decoder layers; the geometry and calibration code assume hidden width 1,024. The backbone is in evaluation mode, its parameters have gradients disabled, and calibration and generation run without gradient tracking.
|
| 33 |
|
| 34 |

|
|
@@ -248,6 +258,8 @@ Successful generation records include stopping reason, token count, elapsed time
|
|
| 248 |
|
| 249 |
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.
|
| 250 |
|
|
|
|
|
|
|
| 251 |
## Engineering checks and evidence
|
| 252 |
|
| 253 |
Smoke mode is the default and operates on one supplied passage. It runs baseline, zero-strength, positive-valence, negative-valence, and restored-baseline conditions. Positive and negative conditions use coefficient vectors $(0.8,0,0)$ and $(-0.8,0,0)$. The third coordinate zero corresponds to intensity 0.5.
|
|
@@ -258,9 +270,13 @@ Smoke mode defaults to the original 128-token allowance and 600-second time thre
|
|
| 258 |
|
| 259 |
For the November 13 revision, seven CPU policy tests passed, Python and shell syntax were checked, and CLI loading was verified. Tests cover such behavior as unchanged requested controls, paragraph allowance accounting, draft retention, sentence cleanup, and continuing after a failed chunk. They do not execute the CUDA model. The current revision has not yet been run on an ARC GPU. A successful GPU run was reported for an earlier revision on November 12; it is historical evidence, not validation of the current code.
|
| 260 |
|
|
|
|
|
|
|
| 261 |
## Reproduction
|
| 262 |
|
| 263 |
-
|
|
|
|
|
|
|
| 264 |
|
| 265 |
```bash
|
| 266 |
bash setup.sh
|
|
@@ -298,8 +314,9 @@ Changing to a larger model requires new architecture validation, basis calibrati
|
|
| 298 |
|
| 299 |
## Implementation sources
|
| 300 |
|
| 301 |
-
This report
|
| 302 |
|
|
|
|
| 303 |
- [CUDA runner and smoke protocol](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/run_arc.py)
|
| 304 |
- [Document transformation and incremental output](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/document_run.py)
|
| 305 |
- [Calibration, displacement math, hooks, and decoding](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/engine/steering.py)
|
|
|
|
| 3 |
**William Taggart**
|
| 4 |
Modern and Classical Languages and Literatures, Virginia Tech
|
| 5 |
Email: wtaggart@vt.edu
|
| 6 |
+
**Technical report · November 14, 2026 · Revision 0.3**
|
| 7 |
|
| 8 |
## Abstract
|
| 9 |
|
| 10 |
+
AFFECTOR-ARC is an experimental text-rewriting system with a native macOS application and a command-line CUDA runner. The desktop application manages documents, emotional geometries, SSH connections, Slurm GPU jobs, and downloaded results; the runner executes a frozen Qwen3.5-0.8B model on an allocated ARC GPU. A trajectory specifies valence, inward/outward emotional orientation, and intensity across normalized document position. Contrasting English passages establish three activation directions at each selected transformer layer. During generation, trajectory coordinates determine a displacement scaled to the hidden-state magnitude, smoothed across decoding steps, and constrained by an RMS budget. Paired hooks apply it to the feed-forward input and residual path. Four editing styles, sentence-aware segmentation, bounded generation, and incremental reports make the intervention available within a document workflow. This report describes both interfaces and derives the steering mechanism, while distinguishing implementation checks from evidence of emotional effectiveness. Earlier desktop GPU runs and CPU policy tests are documented separately; they do not validate every subsequent runner revision or establish emotional accuracy, semantic preservation, or multilingual effectiveness.
|
| 11 |
|
| 12 |
## Purpose and Research Context
|
| 13 |
|
| 14 |
+
AFFECTOR-ARC investigates whether a writer can guide the emotional progression of a document by specifying a trajectory and intervening in a language model's internal computation. The practical aim is to make that experiment usable across a complete rewrite: the writer selects a text, supplies an emotional geometry, adjusts the intervention, and inspects the resulting passages and diagnostics. A native macOS application brings these activities into a workspace, while SSH and Slurm connect that workspace to a GPU allocation on Virginia Tech ARC. A command-line runner exposes the computation separately for scripted experiments and reproducible execution.
|
| 15 |
|
| 16 |
+
The geometry gives this workflow a common representation. It is an ordered path through valence, inward/outward emotional orientation, and intensity, accompanied by layer settings and optional calibration matrices. Sampling the path at successive document positions supplies the coordinates used during rewriting. Those coordinates become displacements of the model's hidden states, so the desired progression acts on the computation that produces tokens rather than prescribing particular words. The model's stored parameters remain fixed throughout.
|
| 17 |
+
|
| 18 |
+
This mechanism builds on contrastive activation steering. Turner and colleagues derive directions from contrasting prompts, and Panickssery and colleagues average differences between positive and negative behavioral examples ([Turner et al., 2023](https://arxiv.org/abs/2308.10248v1); [Panickssery et al., 2024](https://arxiv.org/abs/2312.06681v4)). AFFECTOR-ARC applies the same broad principle to small sets of emotionally contrasting passages, then mixes the resulting directions according to the requested trajectory. The central research question is how reliably this mapping can influence a sustained rewrite, given that a numerical displacement need not produce a unique or consistent emotional interpretation.
|
| 19 |
|
| 20 |

|
| 21 |
|
| 22 |
*Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
|
| 23 |
|
| 24 |
+
That question connects the implementation to research on the relation between model representations and human judgments. Wu and colleagues study concept-based emotional steering across linguistic-cultural settings, while Diallo and colleagues use human ratings to examine perceived emotional intensity and text quality ([Wu et al., 2025](https://arxiv.org/abs/2506.13978v1); [Diallo et al., 2026](https://arxiv.org/abs/2601.21505v1)). These studies motivate coordinate-based experiments whose outcomes are judged through readers' interpretations. They do not, however, establish the validity of this system's inward/outward axis or calibrate its RMS-scaled controls. The requested trajectory, the activation displacement, and the emotion perceived in the output must therefore be evaluated as related but distinct quantities.
|
| 25 |
|
| 26 |
+
The reliability of that relationship also depends on where and how the intervention is applied. Shu and colleagues find uneven causal organization across emotion categories, and Wang and colleagues show that strong probing accuracy need not identify the most effective intervention layers ([Shu et al., 2026](https://arxiv.org/abs/2604.25866v3); [Wang et al., 2026](https://arxiv.org/abs/2609.22135v1)). Ha and colleagues' finding that supportive-empathy directions are only partially separable further cautions against treating emotional axes as independent controls ([Ha et al., 2026](https://arxiv.org/abs/2609.15654v1)). Automated vector extraction and context-dependent strength offer approaches to these calibration problems ([Cui et al., 2025](https://arxiv.org/abs/2509.22739v3); [Ferrando et al., 2025](https://arxiv.org/abs/2512.03661v2)). AFFECTOR-ARC makes layer weights and axis mixtures explicit so that such questions can be investigated, although its current runner uses user-selected settings rather than an adaptive emotional-feedback controller.
|
| 27 |
|
| 28 |
+
Applying those settings across a document makes generation stability part of the same problem. E-STEER reports non-monotonic effects of emotional interventions, and Jeong observes both emotional changes and degeneration in small models ([Sun et al., 2026](https://arxiv.org/abs/2604.00005v1); [Jeong, 2026](https://arxiv.org/abs/2604.04064v1)). A stronger displacement can thus disrupt a rewrite without bringing it closer to the intended trajectory. The system addresses the operational consequences through bounded generation, repetition checks, passage-level failure records, and a desktop workflow for reviewing jobs and results. These safeguards preserve an inspectable experiment; they do not establish that a completed rewrite meets its emotional target.
|
| 29 |
|
| 30 |
+
Evaluation must consequently consider emotional alignment together with fluency, content change, and failure rates. SteerScope examines efficacy, side effects, generalization, and dependence on steering data, finding no general advantage over prompt steering in its tested trade-offs ([Yang et al., 2026](https://arxiv.org/abs/2610.07722v1)). For AFFECTOR-ARC, this motivates comparisons with prompt-only and unsteered rewriting, blinded emotional and quality ratings, and complete reporting of retained originals, repetition, and truncation. The scope of this report is the desktop-to-cluster workflow and the implemented steering procedure, including their engineering checks. Establishing emotional effectiveness and a reliable operating range remains an empirical task.
|
| 31 |
|
| 32 |
## Runtime and data flow
|
| 33 |
|
| 34 |
+
### Native macOS client and remote execution
|
| 35 |
+
|
| 36 |
+
The native SwiftUI application is distributed on Hugging Face for Apple Silicon Macs running macOS 15 or later. Its connection workspace configures SSH access and the Slurm resource request. The document workspace holds the input and generated text; the shape and weights workspace supports importing, converting, and editing geometries and intervention settings. Jobs and results controls expose remote progress, diagnostics, and output downloads. The application also manages Hugging Face source updates and submits GPU-environment setup through Slurm.
|
| 37 |
+
|
| 38 |
+
The Mac serves as the client: lightweight SSH requests manage files and jobs, while model setup and execution run in requested GPU allocations. The bundled Python remote agent handles requests over SSH without a persistent listening server. The desktop interface and the command-line runner are separately versioned components, so a recorded application version alone does not identify the remote engine revision. Both should be recorded when reporting a run.
|
| 39 |
+
|
| 40 |
+
### CUDA runner
|
| 41 |
+
|
| 42 |
The runner loads `Qwen/Qwen3.5-0.8B` at revision `2fc06364715b967f1860aea9cf38778875588b17`. The implementation uses `Qwen3_5ForConditionalGeneration`, its text decoder layers, float32 computation, eager attention, and the first process-visible CUDA device. It checks for 24 decoder layers; the geometry and calibration code assume hidden width 1,024. The backbone is in evaluation mode, its parameters have gradients disabled, and calibration and generation run without gradient tracking.
|
| 43 |
|
| 44 |

|
|
|
|
| 258 |
|
| 259 |
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.
|
| 260 |
|
| 261 |
+
The desktop document view is a presentation layer over these artifacts. The published client displays generated passages separately from source text and diagnostics; its text exports can therefore differ from the runner's assembled file, which includes retained originals. The downloadable report is needed to distinguish generated, retained, failed, and unprocessed passages. Reconnecting can recover an existing job and its outputs, but this client recovery does not make the runner's incremental report a resumable generation checkpoint.
|
| 262 |
+
|
| 263 |
## Engineering checks and evidence
|
| 264 |
|
| 265 |
Smoke mode is the default and operates on one supplied passage. It runs baseline, zero-strength, positive-valence, negative-valence, and restored-baseline conditions. Positive and negative conditions use coefficient vectors $(0.8,0,0)$ and $(-0.8,0,0)$. The third coordinate zero corresponds to intensity 0.5.
|
|
|
|
| 270 |
|
| 271 |
For the November 13 revision, seven CPU policy tests passed, Python and shell syntax were checked, and CLI loading was verified. Tests cover such behavior as unchanged requested controls, paragraph allowance accounting, draft retention, sentence cleanup, and continuing after a failed chunk. They do not execute the CUDA model. The current revision has not yet been run on an ARC GPU. A successful GPU run was reported for an earlier revision on November 12; it is historical evidence, not validation of the current code.
|
| 272 |
|
| 273 |
+
The desktop release also has its own validation record. Its published version 1.0.4 documentation reports native Swift tests, 14 remote-agent tests, a completed sample rewrite, and an earlier GPU setup on an NVIDIA A100-SXM4-80GB. These are release-specific engineering observations, not measurements of emotional accuracy or validation of later changes to the remote runner. The screenshots show the native application's workspaces, including offline states where indicated.
|
| 274 |
+
|
| 275 |
## Reproduction
|
| 276 |
|
| 277 |
+
For the desktop workflow, download the native macOS application from the repository's macos directory on Hugging Face, unpack it, and open AFFECTOR ARC.app. Configure the connection profile for the user's own ARC login, SSH authentication, and Slurm allocation. After connecting, install or update the remote source and prepare the GPU environment as needed. Import a document and geometry, adjust the settings, and choose Transform on ARC. Monitor the submitted job in Jobs + results and download its input, output, report, geometry, and logs. The published desktop README identifies the application as version 1.0.4 and describes its build as ad-hoc signed.
|
| 278 |
+
|
| 279 |
+
The following commands provide the separate command-line workflow. They run inside a suitable ARC GPU allocation from a fresh repository copy:
|
| 280 |
|
| 281 |
```bash
|
| 282 |
bash setup.sh
|
|
|
|
| 314 |
|
| 315 |
## Implementation sources
|
| 316 |
|
| 317 |
+
This report covers the native macOS client, its SSH request agent, and the CUDA runner. The desktop release and its source are available alongside the remote engine:
|
| 318 |
|
| 319 |
+
- [Native macOS application, Swift source, and release documentation](https://huggingface.co/will-taggart/AFFECTOR-ARC/tree/main/macos)
|
| 320 |
- [CUDA runner and smoke protocol](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/run_arc.py)
|
| 321 |
- [Document transformation and incremental output](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/document_run.py)
|
| 322 |
- [Calibration, displacement math, hooks, and decoding](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/engine/steering.py)
|
paper/source.md
CHANGED
|
@@ -3,32 +3,42 @@
|
|
| 3 |
**William Taggart**
|
| 4 |
Modern and Classical Languages and Literatures, Virginia Tech
|
| 5 |
Email: wtaggart@vt.edu
|
| 6 |
-
**Technical report · November
|
| 7 |
|
| 8 |
## Abstract
|
| 9 |
|
| 10 |
-
AFFECTOR-ARC is an experimental
|
| 11 |
|
| 12 |
## Purpose and Research Context
|
| 13 |
|
| 14 |
-
AFFECTOR-ARC investigates
|
| 15 |
|
| 16 |
-
The
|
|
|
|
|
|
|
| 17 |
|
| 18 |

|
| 19 |
|
| 20 |
*Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
|
| 21 |
|
| 22 |
-
|
| 23 |
|
| 24 |
-
|
| 25 |
|
| 26 |
-
|
| 27 |
|
| 28 |
-
|
| 29 |
|
| 30 |
## Runtime and data flow
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
The runner loads `Qwen/Qwen3.5-0.8B` at revision `2fc06364715b967f1860aea9cf38778875588b17`. The implementation uses `Qwen3_5ForConditionalGeneration`, its text decoder layers, float32 computation, eager attention, and the first process-visible CUDA device. It checks for 24 decoder layers; the geometry and calibration code assume hidden width 1,024. The backbone is in evaluation mode, its parameters have gradients disabled, and calibration and generation run without gradient tracking.
|
| 33 |
|
| 34 |

|
|
@@ -248,6 +258,8 @@ Successful generation records include stopping reason, token count, elapsed time
|
|
| 248 |
|
| 249 |
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.
|
| 250 |
|
|
|
|
|
|
|
| 251 |
## Engineering checks and evidence
|
| 252 |
|
| 253 |
Smoke mode is the default and operates on one supplied passage. It runs baseline, zero-strength, positive-valence, negative-valence, and restored-baseline conditions. Positive and negative conditions use coefficient vectors $(0.8,0,0)$ and $(-0.8,0,0)$. The third coordinate zero corresponds to intensity 0.5.
|
|
@@ -258,9 +270,13 @@ Smoke mode defaults to the original 128-token allowance and 600-second time thre
|
|
| 258 |
|
| 259 |
For the November 13 revision, seven CPU policy tests passed, Python and shell syntax were checked, and CLI loading was verified. Tests cover such behavior as unchanged requested controls, paragraph allowance accounting, draft retention, sentence cleanup, and continuing after a failed chunk. They do not execute the CUDA model. The current revision has not yet been run on an ARC GPU. A successful GPU run was reported for an earlier revision on November 12; it is historical evidence, not validation of the current code.
|
| 260 |
|
|
|
|
|
|
|
| 261 |
## Reproduction
|
| 262 |
|
| 263 |
-
|
|
|
|
|
|
|
| 264 |
|
| 265 |
```bash
|
| 266 |
bash setup.sh
|
|
@@ -298,8 +314,9 @@ Changing to a larger model requires new architecture validation, basis calibrati
|
|
| 298 |
|
| 299 |
## Implementation sources
|
| 300 |
|
| 301 |
-
This report
|
| 302 |
|
|
|
|
| 303 |
- [CUDA runner and smoke protocol](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/run_arc.py)
|
| 304 |
- [Document transformation and incremental output](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/document_run.py)
|
| 305 |
- [Calibration, displacement math, hooks, and decoding](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/engine/steering.py)
|
|
|
|
| 3 |
**William Taggart**
|
| 4 |
Modern and Classical Languages and Literatures, Virginia Tech
|
| 5 |
Email: wtaggart@vt.edu
|
| 6 |
+
**Technical report · November 14, 2026 · Revision 0.3**
|
| 7 |
|
| 8 |
## Abstract
|
| 9 |
|
| 10 |
+
AFFECTOR-ARC is an experimental text-rewriting system with a native macOS application and a command-line CUDA runner. The desktop application manages documents, emotional geometries, SSH connections, Slurm GPU jobs, and downloaded results; the runner executes a frozen Qwen3.5-0.8B model on an allocated ARC GPU. A trajectory specifies valence, inward/outward emotional orientation, and intensity across normalized document position. Contrasting English passages establish three activation directions at each selected transformer layer. During generation, trajectory coordinates determine a displacement scaled to the hidden-state magnitude, smoothed across decoding steps, and constrained by an RMS budget. Paired hooks apply it to the feed-forward input and residual path. Four editing styles, sentence-aware segmentation, bounded generation, and incremental reports make the intervention available within a document workflow. This report describes both interfaces and derives the steering mechanism, while distinguishing implementation checks from evidence of emotional effectiveness. Earlier desktop GPU runs and CPU policy tests are documented separately; they do not validate every subsequent runner revision or establish emotional accuracy, semantic preservation, or multilingual effectiveness.
|
| 11 |
|
| 12 |
## Purpose and Research Context
|
| 13 |
|
| 14 |
+
AFFECTOR-ARC investigates whether a writer can guide the emotional progression of a document by specifying a trajectory and intervening in a language model's internal computation. The practical aim is to make that experiment usable across a complete rewrite: the writer selects a text, supplies an emotional geometry, adjusts the intervention, and inspects the resulting passages and diagnostics. A native macOS application brings these activities into a workspace, while SSH and Slurm connect that workspace to a GPU allocation on Virginia Tech ARC. A command-line runner exposes the computation separately for scripted experiments and reproducible execution.
|
| 15 |
|
| 16 |
+
The geometry gives this workflow a common representation. It is an ordered path through valence, inward/outward emotional orientation, and intensity, accompanied by layer settings and optional calibration matrices. Sampling the path at successive document positions supplies the coordinates used during rewriting. Those coordinates become displacements of the model's hidden states, so the desired progression acts on the computation that produces tokens rather than prescribing particular words. The model's stored parameters remain fixed throughout.
|
| 17 |
+
|
| 18 |
+
This mechanism builds on contrastive activation steering. Turner and colleagues derive directions from contrasting prompts, and Panickssery and colleagues average differences between positive and negative behavioral examples ([Turner et al., 2023](https://arxiv.org/abs/2308.10248v1); [Panickssery et al., 2024](https://arxiv.org/abs/2312.06681v4)). AFFECTOR-ARC applies the same broad principle to small sets of emotionally contrasting passages, then mixes the resulting directions according to the requested trajectory. The central research question is how reliably this mapping can influence a sustained rewrite, given that a numerical displacement need not produce a unique or consistent emotional interpretation.
|
| 19 |
|
| 20 |

|
| 21 |
|
| 22 |
*Document workspace displaying an existing Little Red Riding Hood input and rewrite.*
|
| 23 |
|
| 24 |
+
That question connects the implementation to research on the relation between model representations and human judgments. Wu and colleagues study concept-based emotional steering across linguistic-cultural settings, while Diallo and colleagues use human ratings to examine perceived emotional intensity and text quality ([Wu et al., 2025](https://arxiv.org/abs/2506.13978v1); [Diallo et al., 2026](https://arxiv.org/abs/2601.21505v1)). These studies motivate coordinate-based experiments whose outcomes are judged through readers' interpretations. They do not, however, establish the validity of this system's inward/outward axis or calibrate its RMS-scaled controls. The requested trajectory, the activation displacement, and the emotion perceived in the output must therefore be evaluated as related but distinct quantities.
|
| 25 |
|
| 26 |
+
The reliability of that relationship also depends on where and how the intervention is applied. Shu and colleagues find uneven causal organization across emotion categories, and Wang and colleagues show that strong probing accuracy need not identify the most effective intervention layers ([Shu et al., 2026](https://arxiv.org/abs/2604.25866v3); [Wang et al., 2026](https://arxiv.org/abs/2609.22135v1)). Ha and colleagues' finding that supportive-empathy directions are only partially separable further cautions against treating emotional axes as independent controls ([Ha et al., 2026](https://arxiv.org/abs/2609.15654v1)). Automated vector extraction and context-dependent strength offer approaches to these calibration problems ([Cui et al., 2025](https://arxiv.org/abs/2509.22739v3); [Ferrando et al., 2025](https://arxiv.org/abs/2512.03661v2)). AFFECTOR-ARC makes layer weights and axis mixtures explicit so that such questions can be investigated, although its current runner uses user-selected settings rather than an adaptive emotional-feedback controller.
|
| 27 |
|
| 28 |
+
Applying those settings across a document makes generation stability part of the same problem. E-STEER reports non-monotonic effects of emotional interventions, and Jeong observes both emotional changes and degeneration in small models ([Sun et al., 2026](https://arxiv.org/abs/2604.00005v1); [Jeong, 2026](https://arxiv.org/abs/2604.04064v1)). A stronger displacement can thus disrupt a rewrite without bringing it closer to the intended trajectory. The system addresses the operational consequences through bounded generation, repetition checks, passage-level failure records, and a desktop workflow for reviewing jobs and results. These safeguards preserve an inspectable experiment; they do not establish that a completed rewrite meets its emotional target.
|
| 29 |
|
| 30 |
+
Evaluation must consequently consider emotional alignment together with fluency, content change, and failure rates. SteerScope examines efficacy, side effects, generalization, and dependence on steering data, finding no general advantage over prompt steering in its tested trade-offs ([Yang et al., 2026](https://arxiv.org/abs/2610.07722v1)). For AFFECTOR-ARC, this motivates comparisons with prompt-only and unsteered rewriting, blinded emotional and quality ratings, and complete reporting of retained originals, repetition, and truncation. The scope of this report is the desktop-to-cluster workflow and the implemented steering procedure, including their engineering checks. Establishing emotional effectiveness and a reliable operating range remains an empirical task.
|
| 31 |
|
| 32 |
## Runtime and data flow
|
| 33 |
|
| 34 |
+
### Native macOS client and remote execution
|
| 35 |
+
|
| 36 |
+
The native SwiftUI application is distributed on Hugging Face for Apple Silicon Macs running macOS 15 or later. Its connection workspace configures SSH access and the Slurm resource request. The document workspace holds the input and generated text; the shape and weights workspace supports importing, converting, and editing geometries and intervention settings. Jobs and results controls expose remote progress, diagnostics, and output downloads. The application also manages Hugging Face source updates and submits GPU-environment setup through Slurm.
|
| 37 |
+
|
| 38 |
+
The Mac serves as the client: lightweight SSH requests manage files and jobs, while model setup and execution run in requested GPU allocations. The bundled Python remote agent handles requests over SSH without a persistent listening server. The desktop interface and the command-line runner are separately versioned components, so a recorded application version alone does not identify the remote engine revision. Both should be recorded when reporting a run.
|
| 39 |
+
|
| 40 |
+
### CUDA runner
|
| 41 |
+
|
| 42 |
The runner loads `Qwen/Qwen3.5-0.8B` at revision `2fc06364715b967f1860aea9cf38778875588b17`. The implementation uses `Qwen3_5ForConditionalGeneration`, its text decoder layers, float32 computation, eager attention, and the first process-visible CUDA device. It checks for 24 decoder layers; the geometry and calibration code assume hidden width 1,024. The backbone is in evaluation mode, its parameters have gradients disabled, and calibration and generation run without gradient tracking.
|
| 43 |
|
| 44 |

|
|
|
|
| 258 |
|
| 259 |
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.
|
| 260 |
|
| 261 |
+
The desktop document view is a presentation layer over these artifacts. The published client displays generated passages separately from source text and diagnostics; its text exports can therefore differ from the runner's assembled file, which includes retained originals. The downloadable report is needed to distinguish generated, retained, failed, and unprocessed passages. Reconnecting can recover an existing job and its outputs, but this client recovery does not make the runner's incremental report a resumable generation checkpoint.
|
| 262 |
+
|
| 263 |
## Engineering checks and evidence
|
| 264 |
|
| 265 |
Smoke mode is the default and operates on one supplied passage. It runs baseline, zero-strength, positive-valence, negative-valence, and restored-baseline conditions. Positive and negative conditions use coefficient vectors $(0.8,0,0)$ and $(-0.8,0,0)$. The third coordinate zero corresponds to intensity 0.5.
|
|
|
|
| 270 |
|
| 271 |
For the November 13 revision, seven CPU policy tests passed, Python and shell syntax were checked, and CLI loading was verified. Tests cover such behavior as unchanged requested controls, paragraph allowance accounting, draft retention, sentence cleanup, and continuing after a failed chunk. They do not execute the CUDA model. The current revision has not yet been run on an ARC GPU. A successful GPU run was reported for an earlier revision on November 12; it is historical evidence, not validation of the current code.
|
| 272 |
|
| 273 |
+
The desktop release also has its own validation record. Its published version 1.0.4 documentation reports native Swift tests, 14 remote-agent tests, a completed sample rewrite, and an earlier GPU setup on an NVIDIA A100-SXM4-80GB. These are release-specific engineering observations, not measurements of emotional accuracy or validation of later changes to the remote runner. The screenshots show the native application's workspaces, including offline states where indicated.
|
| 274 |
+
|
| 275 |
## Reproduction
|
| 276 |
|
| 277 |
+
For the desktop workflow, download the native macOS application from the repository's macos directory on Hugging Face, unpack it, and open AFFECTOR ARC.app. Configure the connection profile for the user's own ARC login, SSH authentication, and Slurm allocation. After connecting, install or update the remote source and prepare the GPU environment as needed. Import a document and geometry, adjust the settings, and choose Transform on ARC. Monitor the submitted job in Jobs + results and download its input, output, report, geometry, and logs. The published desktop README identifies the application as version 1.0.4 and describes its build as ad-hoc signed.
|
| 278 |
+
|
| 279 |
+
The following commands provide the separate command-line workflow. They run inside a suitable ARC GPU allocation from a fresh repository copy:
|
| 280 |
|
| 281 |
```bash
|
| 282 |
bash setup.sh
|
|
|
|
| 314 |
|
| 315 |
## Implementation sources
|
| 316 |
|
| 317 |
+
This report covers the native macOS client, its SSH request agent, and the CUDA runner. The desktop release and its source are available alongside the remote engine:
|
| 318 |
|
| 319 |
+
- [Native macOS application, Swift source, and release documentation](https://huggingface.co/will-taggart/AFFECTOR-ARC/tree/main/macos)
|
| 320 |
- [CUDA runner and smoke protocol](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/run_arc.py)
|
| 321 |
- [Document transformation and incremental output](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/document_run.py)
|
| 322 |
- [Calibration, displacement math, hooks, and decoding](https://huggingface.co/will-taggart/AFFECTOR-ARC/blob/main/engine/steering.py)
|