Illustrate technical papers with authentic application screenshots
Browse files- .gitattributes +5 -0
- SHA256SUMS.txt +14 -14
- docs/AFFECTOR-Technical-Paper.pdf +2 -2
- docs/TECHNICAL_PAPER.md +24 -14
- docs/TECHNICAL_REPORT.md +24 -14
- paper/AFFECTOR-ACM-source.zip +2 -2
- paper/AFFECTOR-ACM.pdf +2 -2
- paper/AFFECTOR-ACM.source.zip +2 -2
- paper/AFFECTOR-ACM.tex +43 -29
- paper/AFFECTOR-Technical-Report-ACM-source.zip +2 -2
- paper/AFFECTOR-Technical-Report-ACM.pdf +2 -2
- paper/AFFECTOR-Technical-Report-ACM.source.zip +2 -2
- paper/AFFECTOR-Technical-Report-ACM.tex +43 -29
- paper/ILLUSTRATIONS.json +62 -37
- paper/PAPER-SHA256SUMS.txt +11 -11
- paper/SCREENSHOTS.json +57 -0
- paper/TECHNICAL_PAPER.md +22 -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/figures/screenshot-05.png +3 -0
- paper/source.md +22 -12
.gitattributes
CHANGED
|
@@ -38,3 +38,8 @@ docs/acm/AFFECTOR-ACM.pdf filter=lfs diff=lfs merge=lfs -text
|
|
| 38 |
docs/AFFECTOR-Technical-Paper.pdf filter=lfs diff=lfs merge=lfs -text
|
| 39 |
paper/AFFECTOR-ACM.pdf filter=lfs diff=lfs merge=lfs -text
|
| 40 |
paper/AFFECTOR-Technical-Report-ACM.pdf filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
docs/AFFECTOR-Technical-Paper.pdf filter=lfs diff=lfs merge=lfs -text
|
| 39 |
paper/AFFECTOR-ACM.pdf filter=lfs diff=lfs merge=lfs -text
|
| 40 |
paper/AFFECTOR-Technical-Report-ACM.pdf filter=lfs diff=lfs merge=lfs -text
|
| 41 |
+
paper/figures/screenshot-01.png filter=lfs diff=lfs merge=lfs -text
|
| 42 |
+
paper/figures/screenshot-02.png filter=lfs diff=lfs merge=lfs -text
|
| 43 |
+
paper/figures/screenshot-03.png filter=lfs diff=lfs merge=lfs -text
|
| 44 |
+
paper/figures/screenshot-04.png filter=lfs diff=lfs merge=lfs -text
|
| 45 |
+
paper/figures/screenshot-05.png filter=lfs diff=lfs merge=lfs -text
|
SHA256SUMS.txt
CHANGED
|
@@ -1,24 +1,24 @@
|
|
| 1 |
321cd041f8060331e034b2a72578f57cd3e498883e660016cf73f0a8ba69a9a6 README.md
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
ed8fa321af4fc4e4703efcc22037e244fd072c936d11afc566b8eb862f3f0b1b downloads/AFFECTOR-macOS-AppleSilicon.zip
|
| 6 |
d7621338f71fdc31ce9b6e1d53f15d61d6a76344f2111520fb62f9b00293853b downloads/AFFECTOR-source.zip
|
| 7 |
-
|
| 8 |
d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d paper/AFFECTOR-ACM.bbl
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d paper/AFFECTOR-Technical-Report-ACM.bbl
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 paper/LITERATURE_SOURCES.md
|
| 19 |
578a471b301cdf67305c2b997fce39494c2fd0d15eaf56836379943b3d10f90c paper/PAPER-SHA256SUMS.txt
|
| 20 |
22f8f7bf6992778eb1190ee02b6e83dacbd54ff34af7777a18b4e27a195309a3 paper/README.md
|
| 21 |
-
|
| 22 |
4bd1a1b7f6af31b12b9c92bcd978fe116482892b3463f8ec9c922bc1f636c7a0 paper/figure-specifications.json
|
| 23 |
b11798a28d94d60110a7b19f58107d5e60926b68265f64e62f59e9230a8f500d paper/figures/illustration-01.pdf
|
| 24 |
d2ae42a17ff3f16b89a7cb85d21ac1f3c7a19c2bc62d61f1cc581c2ca6d3e596 paper/figures/illustration-01.png
|
|
@@ -32,4 +32,4 @@ a6a497a38f4ba9de8088f410535bee13dd862d1a97f4175d205df768e1d2c1c9 paper/figures/
|
|
| 32 |
7d8f30b829993326e8d764ae18b3974acddbd86b760d4071bdff4edfcb9b299a paper/figures/illustration-05.png
|
| 33 |
9257f38423bd7e496f1a0916c0fbaefbfb32ab544a684ca9a411508201f0535e paper/references.bib
|
| 34 |
eed8b979a5a299b42e925d9027332e7df876400dcef046f2cfe26beef91d1189 paper/render_figures.py
|
| 35 |
-
|
|
|
|
| 1 |
321cd041f8060331e034b2a72578f57cd3e498883e660016cf73f0a8ba69a9a6 README.md
|
| 2 |
+
c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a docs/AFFECTOR-Technical-Paper.pdf
|
| 3 |
+
b354dd4c97e9521760dfe0c4a2e16bc4762d672b3186132ab6717764c86caff6 docs/TECHNICAL_PAPER.md
|
| 4 |
+
b354dd4c97e9521760dfe0c4a2e16bc4762d672b3186132ab6717764c86caff6 docs/TECHNICAL_REPORT.md
|
| 5 |
ed8fa321af4fc4e4703efcc22037e244fd072c936d11afc566b8eb862f3f0b1b downloads/AFFECTOR-macOS-AppleSilicon.zip
|
| 6 |
d7621338f71fdc31ce9b6e1d53f15d61d6a76344f2111520fb62f9b00293853b downloads/AFFECTOR-source.zip
|
| 7 |
+
cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 paper/AFFECTOR-ACM-source.zip
|
| 8 |
d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d paper/AFFECTOR-ACM.bbl
|
| 9 |
+
c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a paper/AFFECTOR-ACM.pdf
|
| 10 |
+
cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 paper/AFFECTOR-ACM.source.zip
|
| 11 |
+
7ec8d6a693c2de155d62d434bef4c387f911c75f0dceadbd1de9213b9c7691e2 paper/AFFECTOR-ACM.tex
|
| 12 |
+
cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 paper/AFFECTOR-Technical-Report-ACM-source.zip
|
| 13 |
d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d paper/AFFECTOR-Technical-Report-ACM.bbl
|
| 14 |
+
c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a paper/AFFECTOR-Technical-Report-ACM.pdf
|
| 15 |
+
cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 paper/AFFECTOR-Technical-Report-ACM.source.zip
|
| 16 |
+
7ec8d6a693c2de155d62d434bef4c387f911c75f0dceadbd1de9213b9c7691e2 paper/AFFECTOR-Technical-Report-ACM.tex
|
| 17 |
+
e2c0590754823491af4657a764ab6682e27c5eaf0173fc9f5f03a62e84ccb308 paper/ILLUSTRATIONS.json
|
| 18 |
d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 paper/LITERATURE_SOURCES.md
|
| 19 |
578a471b301cdf67305c2b997fce39494c2fd0d15eaf56836379943b3d10f90c paper/PAPER-SHA256SUMS.txt
|
| 20 |
22f8f7bf6992778eb1190ee02b6e83dacbd54ff34af7777a18b4e27a195309a3 paper/README.md
|
| 21 |
+
4b62ddd1788644df9f817c2b10fb649239e5263e22fbc381af3b6f8b411e504f paper/TECHNICAL_PAPER.md
|
| 22 |
4bd1a1b7f6af31b12b9c92bcd978fe116482892b3463f8ec9c922bc1f636c7a0 paper/figure-specifications.json
|
| 23 |
b11798a28d94d60110a7b19f58107d5e60926b68265f64e62f59e9230a8f500d paper/figures/illustration-01.pdf
|
| 24 |
d2ae42a17ff3f16b89a7cb85d21ac1f3c7a19c2bc62d61f1cc581c2ca6d3e596 paper/figures/illustration-01.png
|
|
|
|
| 32 |
7d8f30b829993326e8d764ae18b3974acddbd86b760d4071bdff4edfcb9b299a paper/figures/illustration-05.png
|
| 33 |
9257f38423bd7e496f1a0916c0fbaefbfb32ab544a684ca9a411508201f0535e paper/references.bib
|
| 34 |
eed8b979a5a299b42e925d9027332e7df876400dcef046f2cfe26beef91d1189 paper/render_figures.py
|
| 35 |
+
4b62ddd1788644df9f817c2b10fb649239e5263e22fbc381af3b6f8b411e504f paper/source.md
|
docs/AFFECTOR-Technical-Paper.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:c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a
|
| 3 |
+
size 2132249
|
docs/TECHNICAL_PAPER.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
# AFFECTOR: Coordinate-Guided Text Rewriting through Transformer Activation Interventions
|
| 2 |
|
| 3 |
-
**Technical paper:** [PDF](
|
| 4 |
|
| 5 |
**William Taggart**
|
| 6 |
Modern and Classical Languages and Literatures, Virginia Tech
|
|
@@ -11,6 +11,28 @@ November 13, 2026 · Implementation report · Desktop version 0.1.0
|
|
| 11 |
|
| 12 |
AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
|
| 13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
## Literature review: activation steering and emotion
|
| 15 |
|
| 16 |
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.
|
|
@@ -71,9 +93,6 @@ The resulting `B[l]` is a 3 by 1,024 matrix. These are the editable steering-mod
|
|
| 71 |
|
| 72 |
There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
|
| 73 |
|
| 74 |
-

|
| 75 |
-
|
| 76 |
-
*Illustration 2. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.*
|
| 77 |
|
| 78 |
## 4. Coordinate-to-activation computation
|
| 79 |
|
|
@@ -99,9 +118,6 @@ Zero strength, budget, or layer weight gives an exact zero perturbation. Each ch
|
|
| 99 |
|
| 100 |
Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
|
| 101 |
|
| 102 |
-

|
| 103 |
-
|
| 104 |
-
*Illustration 3. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.*
|
| 105 |
|
| 106 |
## 5. Intervention within the transformer
|
| 107 |
|
|
@@ -119,9 +135,6 @@ The module does not search for trigger phrases or recognize an idiom before acti
|
|
| 119 |
|
| 120 |
A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
|
| 121 |
|
| 122 |
-

|
| 123 |
-
|
| 124 |
-
*Illustration 4. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.*
|
| 125 |
|
| 126 |
## 6. Prompts, sampling, and bounded output
|
| 127 |
|
|
@@ -149,9 +162,6 @@ The denominator is the maximum distance across the three-dimensional `[-1,1]` cu
|
|
| 149 |
|
| 150 |
The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
|
| 151 |
|
| 152 |
-

|
| 153 |
-
|
| 154 |
-
*Illustration 5. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.*
|
| 155 |
|
| 156 |
## 8. Architecture, persistence, and privacy
|
| 157 |
|
|
@@ -191,7 +201,7 @@ The equations and implementation descriptions above are derived from the include
|
|
| 191 |
|
| 192 |
## Research references
|
| 193 |
|
| 194 |
-
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](
|
| 195 |
|
| 196 |
**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.
|
| 197 |
|
|
|
|
| 1 |
# AFFECTOR: Coordinate-Guided Text Rewriting through Transformer Activation Interventions
|
| 2 |
|
| 3 |
+
**Technical paper:** [PDF](AFFECTOR-ACM.pdf) · [LaTeX source](AFFECTOR-ACM.tex) · [Complete source ZIP](AFFECTOR-ACM-source.zip)
|
| 4 |
|
| 5 |
**William Taggart**
|
| 6 |
Modern and Classical Languages and Literatures, Virginia Tech
|
|
|
|
| 11 |
|
| 12 |
AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
|
| 13 |
|
| 14 |
+
## Application interface
|
| 15 |
+
|
| 16 |
+

|
| 17 |
+
|
| 18 |
+
*Document workspace with an example input, rewrite controls, and an empty output pane before generation.*
|
| 19 |
+
|
| 20 |
+

|
| 21 |
+
|
| 22 |
+
*Geometry editor showing the selected three-point trajectory and its steering controls.*
|
| 23 |
+
|
| 24 |
+

|
| 25 |
+
|
| 26 |
+
*Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.*
|
| 27 |
+
|
| 28 |
+

|
| 29 |
+
|
| 30 |
+
*Expanded numerical weights for the selected geometry, including the local calibration metadata.*
|
| 31 |
+
|
| 32 |
+

|
| 33 |
+
|
| 34 |
+
*Verification controls before evaluation. No verification results have been computed in this view.*
|
| 35 |
+
|
| 36 |
## Literature review: activation steering and emotion
|
| 37 |
|
| 38 |
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.
|
|
|
|
| 93 |
|
| 94 |
There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
|
| 95 |
|
|
|
|
|
|
|
|
|
|
| 96 |
|
| 97 |
## 4. Coordinate-to-activation computation
|
| 98 |
|
|
|
|
| 118 |
|
| 119 |
Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
|
| 120 |
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
## 5. Intervention within the transformer
|
| 123 |
|
|
|
|
| 135 |
|
| 136 |
A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
|
| 137 |
|
|
|
|
|
|
|
|
|
|
| 138 |
|
| 139 |
## 6. Prompts, sampling, and bounded output
|
| 140 |
|
|
|
|
| 162 |
|
| 163 |
The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
|
| 164 |
|
|
|
|
|
|
|
|
|
|
| 165 |
|
| 166 |
## 8. Architecture, persistence, and privacy
|
| 167 |
|
|
|
|
| 201 |
|
| 202 |
## Research references
|
| 203 |
|
| 204 |
+
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).
|
| 205 |
|
| 206 |
**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.
|
| 207 |
|
docs/TECHNICAL_REPORT.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
# AFFECTOR: Coordinate-Guided Text Rewriting through Transformer Activation Interventions
|
| 2 |
|
| 3 |
-
**Technical paper:** [PDF](
|
| 4 |
|
| 5 |
**William Taggart**
|
| 6 |
Modern and Classical Languages and Literatures, Virginia Tech
|
|
@@ -11,6 +11,28 @@ November 13, 2026 · Implementation report · Desktop version 0.1.0
|
|
| 11 |
|
| 12 |
AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
|
| 13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
## Literature review: activation steering and emotion
|
| 15 |
|
| 16 |
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.
|
|
@@ -71,9 +93,6 @@ The resulting `B[l]` is a 3 by 1,024 matrix. These are the editable steering-mod
|
|
| 71 |
|
| 72 |
There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
|
| 73 |
|
| 74 |
-

|
| 75 |
-
|
| 76 |
-
*Illustration 2. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.*
|
| 77 |
|
| 78 |
## 4. Coordinate-to-activation computation
|
| 79 |
|
|
@@ -99,9 +118,6 @@ Zero strength, budget, or layer weight gives an exact zero perturbation. Each ch
|
|
| 99 |
|
| 100 |
Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
|
| 101 |
|
| 102 |
-

|
| 103 |
-
|
| 104 |
-
*Illustration 3. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.*
|
| 105 |
|
| 106 |
## 5. Intervention within the transformer
|
| 107 |
|
|
@@ -119,9 +135,6 @@ The module does not search for trigger phrases or recognize an idiom before acti
|
|
| 119 |
|
| 120 |
A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
|
| 121 |
|
| 122 |
-

|
| 123 |
-
|
| 124 |
-
*Illustration 4. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.*
|
| 125 |
|
| 126 |
## 6. Prompts, sampling, and bounded output
|
| 127 |
|
|
@@ -149,9 +162,6 @@ The denominator is the maximum distance across the three-dimensional `[-1,1]` cu
|
|
| 149 |
|
| 150 |
The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
|
| 151 |
|
| 152 |
-

|
| 153 |
-
|
| 154 |
-
*Illustration 5. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.*
|
| 155 |
|
| 156 |
## 8. Architecture, persistence, and privacy
|
| 157 |
|
|
@@ -191,7 +201,7 @@ The equations and implementation descriptions above are derived from the include
|
|
| 191 |
|
| 192 |
## Research references
|
| 193 |
|
| 194 |
-
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](
|
| 195 |
|
| 196 |
**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.
|
| 197 |
|
|
|
|
| 1 |
# AFFECTOR: Coordinate-Guided Text Rewriting through Transformer Activation Interventions
|
| 2 |
|
| 3 |
+
**Technical paper:** [PDF](AFFECTOR-ACM.pdf) · [LaTeX source](AFFECTOR-ACM.tex) · [Complete source ZIP](AFFECTOR-ACM-source.zip)
|
| 4 |
|
| 5 |
**William Taggart**
|
| 6 |
Modern and Classical Languages and Literatures, Virginia Tech
|
|
|
|
| 11 |
|
| 12 |
AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
|
| 13 |
|
| 14 |
+
## Application interface
|
| 15 |
+
|
| 16 |
+

|
| 17 |
+
|
| 18 |
+
*Document workspace with an example input, rewrite controls, and an empty output pane before generation.*
|
| 19 |
+
|
| 20 |
+

|
| 21 |
+
|
| 22 |
+
*Geometry editor showing the selected three-point trajectory and its steering controls.*
|
| 23 |
+
|
| 24 |
+

|
| 25 |
+
|
| 26 |
+
*Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.*
|
| 27 |
+
|
| 28 |
+

|
| 29 |
+
|
| 30 |
+
*Expanded numerical weights for the selected geometry, including the local calibration metadata.*
|
| 31 |
+
|
| 32 |
+

|
| 33 |
+
|
| 34 |
+
*Verification controls before evaluation. No verification results have been computed in this view.*
|
| 35 |
+
|
| 36 |
## Literature review: activation steering and emotion
|
| 37 |
|
| 38 |
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.
|
|
|
|
| 93 |
|
| 94 |
There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
|
| 95 |
|
|
|
|
|
|
|
|
|
|
| 96 |
|
| 97 |
## 4. Coordinate-to-activation computation
|
| 98 |
|
|
|
|
| 118 |
|
| 119 |
Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
|
| 120 |
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
## 5. Intervention within the transformer
|
| 123 |
|
|
|
|
| 135 |
|
| 136 |
A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
|
| 137 |
|
|
|
|
|
|
|
|
|
|
| 138 |
|
| 139 |
## 6. Prompts, sampling, and bounded output
|
| 140 |
|
|
|
|
| 162 |
|
| 163 |
The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
|
| 164 |
|
|
|
|
|
|
|
|
|
|
| 165 |
|
| 166 |
## 8. Architecture, persistence, and privacy
|
| 167 |
|
|
|
|
| 201 |
|
| 202 |
## Research references
|
| 203 |
|
| 204 |
+
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).
|
| 205 |
|
| 206 |
**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.
|
| 207 |
|
paper/AFFECTOR-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:cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256
|
| 3 |
+
size 6604509
|
paper/AFFECTOR-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:c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a
|
| 3 |
+
size 2132249
|
paper/AFFECTOR-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:cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256
|
| 3 |
+
size 6604509
|
paper/AFFECTOR-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: Coordinate-Guided Text Rewriting through Transformer
|
| 19 |
Activation Interventions}
|
|
@@ -114,6 +114,44 @@ the rewrite style, and complete accounting of repetition, truncation,
|
|
| 114 |
and original-text fallbacks. This is an evaluation agenda, not a
|
| 115 |
completed benchmark. \citep{yang2026steerscope}.
|
| 116 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 117 |
\section{Problem formulation}\label{problem-formulation}
|
| 118 |
|
| 119 |
A document conveys more than propositions: lexical choice, focus,
|
|
@@ -228,13 +266,7 @@ genre, topic, lexical content, and emotional categories. Calibration is
|
|
| 228 |
exploratory and English-based; multilingual effectiveness has not been
|
| 229 |
established.
|
| 230 |
|
| 231 |
-
|
| 232 |
-
\centering
|
| 233 |
-
\includegraphics[width=\textwidth]{figures/illustration-02.pdf}
|
| 234 |
-
\caption{Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.}
|
| 235 |
-
\Description{Contrastive calibration supplies three directions. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.}
|
| 236 |
-
\label{fig:illustration-2}
|
| 237 |
-
\end{figure}
|
| 238 |
|
| 239 |
\section{Coordinate-to-activation
|
| 240 |
computation}\label{coordinate-to-activation-computation}
|
|
@@ -276,13 +308,7 @@ state at that site. Large values can destabilize generation. Raising the
|
|
| 276 |
budget only matters when the proposed perturbation would otherwise be
|
| 277 |
clipped; it cannot correct an inaccurate direction.
|
| 278 |
|
| 279 |
-
|
| 280 |
-
\centering
|
| 281 |
-
\includegraphics[width=\textwidth]{figures/illustration-03.pdf}
|
| 282 |
-
\caption{Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.}
|
| 283 |
-
\Description{Controls shape the activation displacement. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.}
|
| 284 |
-
\label{fig:illustration-3}
|
| 285 |
-
\end{figure}
|
| 286 |
|
| 287 |
\section{Intervention within the
|
| 288 |
transformer}\label{intervention-within-the-transformer}
|
|
@@ -317,13 +343,7 @@ no direct valence-axis perturbation; it does not actively neutralize an
|
|
| 317 |
already negative passage. Correlations between basis vectors further
|
| 318 |
complicate interpretation.
|
| 319 |
|
| 320 |
-
|
| 321 |
-
\centering
|
| 322 |
-
\includegraphics[width=\textwidth]{figures/illustration-04.pdf}
|
| 323 |
-
\caption{Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.}
|
| 324 |
-
\Description{A change to both feed-forward branches. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.}
|
| 325 |
-
\label{fig:illustration-4}
|
| 326 |
-
\end{figure}
|
| 327 |
|
| 328 |
\section{Prompts, sampling, and bounded
|
| 329 |
output}\label{prompts-sampling-and-bounded-output}
|
|
@@ -398,13 +418,7 @@ with incoherence, factual errors, or undesirable content. The
|
|
| 398 |
implementation does not optimize this score during generation and does
|
| 399 |
not include a formal semantic-fidelity metric.
|
| 400 |
|
| 401 |
-
|
| 402 |
-
\centering
|
| 403 |
-
\includegraphics[width=\textwidth]{figures/illustration-05.pdf}
|
| 404 |
-
\caption{Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.}
|
| 405 |
-
\Description{Verification is a separate assessment. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.}
|
| 406 |
-
\label{fig:illustration-5}
|
| 407 |
-
\end{figure}
|
| 408 |
|
| 409 |
\section{Architecture, persistence, and
|
| 410 |
privacy}\label{architecture-persistence-and-privacy}
|
|
|
|
| 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: Coordinate-Guided Text Rewriting through Transformer
|
| 19 |
Activation Interventions}
|
|
|
|
| 114 |
and original-text fallbacks. This is an evaluation agenda, not a
|
| 115 |
completed benchmark. \citep{yang2026steerscope}.
|
| 116 |
|
| 117 |
+
\section{Application interface}
|
| 118 |
+
The application exposes its principal operations through the following workspaces and controls.
|
| 119 |
+
\begin{figure}[H]
|
| 120 |
+
\centering
|
| 121 |
+
\includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-01.png}
|
| 122 |
+
\caption{Document workspace with an example input, rewrite controls, and an empty output pane before generation.}
|
| 123 |
+
\Description{Document workspace with an example input, rewrite controls, and an empty output pane before generation.}
|
| 124 |
+
\label{fig:app-screen-1}
|
| 125 |
+
\end{figure}
|
| 126 |
+
\begin{figure}[H]
|
| 127 |
+
\centering
|
| 128 |
+
\includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-02.png}
|
| 129 |
+
\caption{Geometry editor showing the selected three-point trajectory and its steering controls.}
|
| 130 |
+
\Description{Geometry editor showing the selected three-point trajectory and its steering controls.}
|
| 131 |
+
\label{fig:app-screen-2}
|
| 132 |
+
\end{figure}
|
| 133 |
+
\begin{figure}[H]
|
| 134 |
+
\centering
|
| 135 |
+
\includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-03.png}
|
| 136 |
+
\caption{Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.}
|
| 137 |
+
\Description{Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.}
|
| 138 |
+
\label{fig:app-screen-3}
|
| 139 |
+
\end{figure}
|
| 140 |
+
\begin{figure}[H]
|
| 141 |
+
\centering
|
| 142 |
+
\includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-04.png}
|
| 143 |
+
\caption{Expanded numerical weights for the selected geometry, including the local calibration metadata.}
|
| 144 |
+
\Description{Expanded numerical weights for the selected geometry, including the local calibration metadata.}
|
| 145 |
+
\label{fig:app-screen-4}
|
| 146 |
+
\end{figure}
|
| 147 |
+
\begin{figure}[H]
|
| 148 |
+
\centering
|
| 149 |
+
\includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-05.png}
|
| 150 |
+
\caption{Verification controls before evaluation. No verification results have been computed in this view.}
|
| 151 |
+
\Description{Verification controls before evaluation. No verification results have been computed in this view.}
|
| 152 |
+
\label{fig:app-screen-5}
|
| 153 |
+
\end{figure}
|
| 154 |
+
|
| 155 |
\section{Problem formulation}\label{problem-formulation}
|
| 156 |
|
| 157 |
A document conveys more than propositions: lexical choice, focus,
|
|
|
|
| 266 |
exploratory and English-based; multilingual effectiveness has not been
|
| 267 |
established.
|
| 268 |
|
| 269 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 270 |
|
| 271 |
\section{Coordinate-to-activation
|
| 272 |
computation}\label{coordinate-to-activation-computation}
|
|
|
|
| 308 |
budget only matters when the proposed perturbation would otherwise be
|
| 309 |
clipped; it cannot correct an inaccurate direction.
|
| 310 |
|
| 311 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 312 |
|
| 313 |
\section{Intervention within the
|
| 314 |
transformer}\label{intervention-within-the-transformer}
|
|
|
|
| 343 |
already negative passage. Correlations between basis vectors further
|
| 344 |
complicate interpretation.
|
| 345 |
|
| 346 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 347 |
|
| 348 |
\section{Prompts, sampling, and bounded
|
| 349 |
output}\label{prompts-sampling-and-bounded-output}
|
|
|
|
| 418 |
implementation does not optimize this score during generation and does
|
| 419 |
not include a formal semantic-fidelity metric.
|
| 420 |
|
| 421 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 422 |
|
| 423 |
\section{Architecture, persistence, and
|
| 424 |
privacy}\label{architecture-persistence-and-privacy}
|
paper/AFFECTOR-Technical-Report-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:cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256
|
| 3 |
+
size 6604509
|
paper/AFFECTOR-Technical-Report-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:c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a
|
| 3 |
+
size 2132249
|
paper/AFFECTOR-Technical-Report-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:cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256
|
| 3 |
+
size 6604509
|
paper/AFFECTOR-Technical-Report-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: Coordinate-Guided Text Rewriting through Transformer
|
| 19 |
Activation Interventions}
|
|
@@ -114,6 +114,44 @@ the rewrite style, and complete accounting of repetition, truncation,
|
|
| 114 |
and original-text fallbacks. This is an evaluation agenda, not a
|
| 115 |
completed benchmark. \citep{yang2026steerscope}.
|
| 116 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 117 |
\section{Problem formulation}\label{problem-formulation}
|
| 118 |
|
| 119 |
A document conveys more than propositions: lexical choice, focus,
|
|
@@ -228,13 +266,7 @@ genre, topic, lexical content, and emotional categories. Calibration is
|
|
| 228 |
exploratory and English-based; multilingual effectiveness has not been
|
| 229 |
established.
|
| 230 |
|
| 231 |
-
|
| 232 |
-
\centering
|
| 233 |
-
\includegraphics[width=\textwidth]{figures/illustration-02.pdf}
|
| 234 |
-
\caption{Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.}
|
| 235 |
-
\Description{Contrastive calibration supplies three directions. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.}
|
| 236 |
-
\label{fig:illustration-2}
|
| 237 |
-
\end{figure}
|
| 238 |
|
| 239 |
\section{Coordinate-to-activation
|
| 240 |
computation}\label{coordinate-to-activation-computation}
|
|
@@ -276,13 +308,7 @@ state at that site. Large values can destabilize generation. Raising the
|
|
| 276 |
budget only matters when the proposed perturbation would otherwise be
|
| 277 |
clipped; it cannot correct an inaccurate direction.
|
| 278 |
|
| 279 |
-
|
| 280 |
-
\centering
|
| 281 |
-
\includegraphics[width=\textwidth]{figures/illustration-03.pdf}
|
| 282 |
-
\caption{Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.}
|
| 283 |
-
\Description{Controls shape the activation displacement. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.}
|
| 284 |
-
\label{fig:illustration-3}
|
| 285 |
-
\end{figure}
|
| 286 |
|
| 287 |
\section{Intervention within the
|
| 288 |
transformer}\label{intervention-within-the-transformer}
|
|
@@ -317,13 +343,7 @@ no direct valence-axis perturbation; it does not actively neutralize an
|
|
| 317 |
already negative passage. Correlations between basis vectors further
|
| 318 |
complicate interpretation.
|
| 319 |
|
| 320 |
-
|
| 321 |
-
\centering
|
| 322 |
-
\includegraphics[width=\textwidth]{figures/illustration-04.pdf}
|
| 323 |
-
\caption{Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.}
|
| 324 |
-
\Description{A change to both feed-forward branches. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.}
|
| 325 |
-
\label{fig:illustration-4}
|
| 326 |
-
\end{figure}
|
| 327 |
|
| 328 |
\section{Prompts, sampling, and bounded
|
| 329 |
output}\label{prompts-sampling-and-bounded-output}
|
|
@@ -398,13 +418,7 @@ with incoherence, factual errors, or undesirable content. The
|
|
| 398 |
implementation does not optimize this score during generation and does
|
| 399 |
not include a formal semantic-fidelity metric.
|
| 400 |
|
| 401 |
-
|
| 402 |
-
\centering
|
| 403 |
-
\includegraphics[width=\textwidth]{figures/illustration-05.pdf}
|
| 404 |
-
\caption{Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.}
|
| 405 |
-
\Description{Verification is a separate assessment. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.}
|
| 406 |
-
\label{fig:illustration-5}
|
| 407 |
-
\end{figure}
|
| 408 |
|
| 409 |
\section{Architecture, persistence, and
|
| 410 |
privacy}\label{architecture-persistence-and-privacy}
|
|
|
|
| 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: Coordinate-Guided Text Rewriting through Transformer
|
| 19 |
Activation Interventions}
|
|
|
|
| 114 |
and original-text fallbacks. This is an evaluation agenda, not a
|
| 115 |
completed benchmark. \citep{yang2026steerscope}.
|
| 116 |
|
| 117 |
+
\section{Application interface}
|
| 118 |
+
The application exposes its principal operations through the following workspaces and controls.
|
| 119 |
+
\begin{figure}[H]
|
| 120 |
+
\centering
|
| 121 |
+
\includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-01.png}
|
| 122 |
+
\caption{Document workspace with an example input, rewrite controls, and an empty output pane before generation.}
|
| 123 |
+
\Description{Document workspace with an example input, rewrite controls, and an empty output pane before generation.}
|
| 124 |
+
\label{fig:app-screen-1}
|
| 125 |
+
\end{figure}
|
| 126 |
+
\begin{figure}[H]
|
| 127 |
+
\centering
|
| 128 |
+
\includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-02.png}
|
| 129 |
+
\caption{Geometry editor showing the selected three-point trajectory and its steering controls.}
|
| 130 |
+
\Description{Geometry editor showing the selected three-point trajectory and its steering controls.}
|
| 131 |
+
\label{fig:app-screen-2}
|
| 132 |
+
\end{figure}
|
| 133 |
+
\begin{figure}[H]
|
| 134 |
+
\centering
|
| 135 |
+
\includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-03.png}
|
| 136 |
+
\caption{Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.}
|
| 137 |
+
\Description{Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.}
|
| 138 |
+
\label{fig:app-screen-3}
|
| 139 |
+
\end{figure}
|
| 140 |
+
\begin{figure}[H]
|
| 141 |
+
\centering
|
| 142 |
+
\includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-04.png}
|
| 143 |
+
\caption{Expanded numerical weights for the selected geometry, including the local calibration metadata.}
|
| 144 |
+
\Description{Expanded numerical weights for the selected geometry, including the local calibration metadata.}
|
| 145 |
+
\label{fig:app-screen-4}
|
| 146 |
+
\end{figure}
|
| 147 |
+
\begin{figure}[H]
|
| 148 |
+
\centering
|
| 149 |
+
\includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-05.png}
|
| 150 |
+
\caption{Verification controls before evaluation. No verification results have been computed in this view.}
|
| 151 |
+
\Description{Verification controls before evaluation. No verification results have been computed in this view.}
|
| 152 |
+
\label{fig:app-screen-5}
|
| 153 |
+
\end{figure}
|
| 154 |
+
|
| 155 |
\section{Problem formulation}\label{problem-formulation}
|
| 156 |
|
| 157 |
A document conveys more than propositions: lexical choice, focus,
|
|
|
|
| 266 |
exploratory and English-based; multilingual effectiveness has not been
|
| 267 |
established.
|
| 268 |
|
| 269 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 270 |
|
| 271 |
\section{Coordinate-to-activation
|
| 272 |
computation}\label{coordinate-to-activation-computation}
|
|
|
|
| 308 |
budget only matters when the proposed perturbation would otherwise be
|
| 309 |
clipped; it cannot correct an inaccurate direction.
|
| 310 |
|
| 311 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 312 |
|
| 313 |
\section{Intervention within the
|
| 314 |
transformer}\label{intervention-within-the-transformer}
|
|
|
|
| 343 |
already negative passage. Correlations between basis vectors further
|
| 344 |
complicate interpretation.
|
| 345 |
|
| 346 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 347 |
|
| 348 |
\section{Prompts, sampling, and bounded
|
| 349 |
output}\label{prompts-sampling-and-bounded-output}
|
|
|
|
| 418 |
implementation does not optimize this score during generation and does
|
| 419 |
not include a formal semantic-fidelity metric.
|
| 420 |
|
| 421 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 422 |
|
| 423 |
\section{Architecture, persistence, and
|
| 424 |
privacy}\label{architecture-persistence-and-privacy}
|
paper/ILLUSTRATIONS.json
CHANGED
|
@@ -1,37 +1,62 @@
|
|
| 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-01.png",
|
| 6 |
+
"crop": [
|
| 7 |
+
10,
|
| 8 |
+
90,
|
| 9 |
+
2690,
|
| 10 |
+
1660
|
| 11 |
+
],
|
| 12 |
+
"caption": "Document workspace with an example input, rewrite controls, and an empty output pane before generation."
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"file": "figures/screenshot-02.png",
|
| 16 |
+
"source": "affector-02.png",
|
| 17 |
+
"crop": [
|
| 18 |
+
450,
|
| 19 |
+
260,
|
| 20 |
+
2660,
|
| 21 |
+
1590
|
| 22 |
+
],
|
| 23 |
+
"caption": "Geometry editor showing the selected three-point trajectory and its steering controls."
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"file": "figures/screenshot-03.png",
|
| 27 |
+
"source": "affector-03.png",
|
| 28 |
+
"crop": [
|
| 29 |
+
22,
|
| 30 |
+
24,
|
| 31 |
+
1498,
|
| 32 |
+
1138
|
| 33 |
+
],
|
| 34 |
+
"caption": "Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis."
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"file": "figures/screenshot-04.png",
|
| 38 |
+
"source": "affector-04.png",
|
| 39 |
+
"crop": [
|
| 40 |
+
450,
|
| 41 |
+
260,
|
| 42 |
+
2660,
|
| 43 |
+
1590
|
| 44 |
+
],
|
| 45 |
+
"caption": "Expanded numerical weights for the selected geometry, including the local calibration metadata."
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"file": "figures/screenshot-05.png",
|
| 49 |
+
"source": "affector-05.png",
|
| 50 |
+
"crop": [
|
| 51 |
+
450,
|
| 52 |
+
270,
|
| 53 |
+
2660,
|
| 54 |
+
850
|
| 55 |
+
],
|
| 56 |
+
"caption": "Verification controls before evaluation. No verification results have been computed in this view."
|
| 57 |
+
}
|
| 58 |
+
],
|
| 59 |
+
"retained_diagrams": [
|
| 60 |
+
1
|
| 61 |
+
]
|
| 62 |
+
}
|
paper/PAPER-SHA256SUMS.txt
CHANGED
|
@@ -1,17 +1,17 @@
|
|
| 1 |
-
|
| 2 |
d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-ACM.bbl
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-Technical-Report-ACM.bbl
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 LITERATURE_SOURCES.md
|
| 13 |
22f8f7bf6992778eb1190ee02b6e83dacbd54ff34af7777a18b4e27a195309a3 README.md
|
| 14 |
-
|
| 15 |
4bd1a1b7f6af31b12b9c92bcd978fe116482892b3463f8ec9c922bc1f636c7a0 figure-specifications.json
|
| 16 |
b11798a28d94d60110a7b19f58107d5e60926b68265f64e62f59e9230a8f500d figures/illustration-01.pdf
|
| 17 |
d2ae42a17ff3f16b89a7cb85d21ac1f3c7a19c2bc62d61f1cc581c2ca6d3e596 figures/illustration-01.png
|
|
@@ -25,4 +25,4 @@ a6a497a38f4ba9de8088f410535bee13dd862d1a97f4175d205df768e1d2c1c9 figures/illust
|
|
| 25 |
7d8f30b829993326e8d764ae18b3974acddbd86b760d4071bdff4edfcb9b299a figures/illustration-05.png
|
| 26 |
9257f38423bd7e496f1a0916c0fbaefbfb32ab544a684ca9a411508201f0535e references.bib
|
| 27 |
eed8b979a5a299b42e925d9027332e7df876400dcef046f2cfe26beef91d1189 render_figures.py
|
| 28 |
-
|
|
|
|
| 1 |
+
cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 AFFECTOR-ACM-source.zip
|
| 2 |
d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-ACM.bbl
|
| 3 |
+
c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a AFFECTOR-ACM.pdf
|
| 4 |
+
cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 AFFECTOR-ACM.source.zip
|
| 5 |
+
7ec8d6a693c2de155d62d434bef4c387f911c75f0dceadbd1de9213b9c7691e2 AFFECTOR-ACM.tex
|
| 6 |
+
cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 AFFECTOR-Technical-Report-ACM-source.zip
|
| 7 |
d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-Technical-Report-ACM.bbl
|
| 8 |
+
c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a AFFECTOR-Technical-Report-ACM.pdf
|
| 9 |
+
cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 AFFECTOR-Technical-Report-ACM.source.zip
|
| 10 |
+
7ec8d6a693c2de155d62d434bef4c387f911c75f0dceadbd1de9213b9c7691e2 AFFECTOR-Technical-Report-ACM.tex
|
| 11 |
+
e2c0590754823491af4657a764ab6682e27c5eaf0173fc9f5f03a62e84ccb308 ILLUSTRATIONS.json
|
| 12 |
d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 LITERATURE_SOURCES.md
|
| 13 |
22f8f7bf6992778eb1190ee02b6e83dacbd54ff34af7777a18b4e27a195309a3 README.md
|
| 14 |
+
4b62ddd1788644df9f817c2b10fb649239e5263e22fbc381af3b6f8b411e504f TECHNICAL_PAPER.md
|
| 15 |
4bd1a1b7f6af31b12b9c92bcd978fe116482892b3463f8ec9c922bc1f636c7a0 figure-specifications.json
|
| 16 |
b11798a28d94d60110a7b19f58107d5e60926b68265f64e62f59e9230a8f500d figures/illustration-01.pdf
|
| 17 |
d2ae42a17ff3f16b89a7cb85d21ac1f3c7a19c2bc62d61f1cc581c2ca6d3e596 figures/illustration-01.png
|
|
|
|
| 25 |
7d8f30b829993326e8d764ae18b3974acddbd86b760d4071bdff4edfcb9b299a figures/illustration-05.png
|
| 26 |
9257f38423bd7e496f1a0916c0fbaefbfb32ab544a684ca9a411508201f0535e references.bib
|
| 27 |
eed8b979a5a299b42e925d9027332e7df876400dcef046f2cfe26beef91d1189 render_figures.py
|
| 28 |
+
4b62ddd1788644df9f817c2b10fb649239e5263e22fbc381af3b6f8b411e504f source.md
|
paper/SCREENSHOTS.json
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"file": "figures/screenshot-01.png",
|
| 4 |
+
"source": "affector-01.png",
|
| 5 |
+
"crop": [
|
| 6 |
+
10,
|
| 7 |
+
90,
|
| 8 |
+
2690,
|
| 9 |
+
1660
|
| 10 |
+
],
|
| 11 |
+
"caption": "Document workspace with an example input, rewrite controls, and an empty output pane before generation."
|
| 12 |
+
},
|
| 13 |
+
{
|
| 14 |
+
"file": "figures/screenshot-02.png",
|
| 15 |
+
"source": "affector-02.png",
|
| 16 |
+
"crop": [
|
| 17 |
+
450,
|
| 18 |
+
260,
|
| 19 |
+
2660,
|
| 20 |
+
1590
|
| 21 |
+
],
|
| 22 |
+
"caption": "Geometry editor showing the selected three-point trajectory and its steering controls."
|
| 23 |
+
},
|
| 24 |
+
{
|
| 25 |
+
"file": "figures/screenshot-03.png",
|
| 26 |
+
"source": "affector-03.png",
|
| 27 |
+
"crop": [
|
| 28 |
+
22,
|
| 29 |
+
24,
|
| 30 |
+
1498,
|
| 31 |
+
1138
|
| 32 |
+
],
|
| 33 |
+
"caption": "Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis."
|
| 34 |
+
},
|
| 35 |
+
{
|
| 36 |
+
"file": "figures/screenshot-04.png",
|
| 37 |
+
"source": "affector-04.png",
|
| 38 |
+
"crop": [
|
| 39 |
+
450,
|
| 40 |
+
260,
|
| 41 |
+
2660,
|
| 42 |
+
1590
|
| 43 |
+
],
|
| 44 |
+
"caption": "Expanded numerical weights for the selected geometry, including the local calibration metadata."
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"file": "figures/screenshot-05.png",
|
| 48 |
+
"source": "affector-05.png",
|
| 49 |
+
"crop": [
|
| 50 |
+
450,
|
| 51 |
+
270,
|
| 52 |
+
2660,
|
| 53 |
+
850
|
| 54 |
+
],
|
| 55 |
+
"caption": "Verification controls before evaluation. No verification results have been computed in this view."
|
| 56 |
+
}
|
| 57 |
+
]
|
paper/TECHNICAL_PAPER.md
CHANGED
|
@@ -11,6 +11,28 @@ November 13, 2026 · Implementation report · Desktop version 0.1.0
|
|
| 11 |
|
| 12 |
AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
|
| 13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
## Literature review: activation steering and emotion
|
| 15 |
|
| 16 |
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.
|
|
@@ -71,9 +93,6 @@ The resulting `B[l]` is a 3 by 1,024 matrix. These are the editable steering-mod
|
|
| 71 |
|
| 72 |
There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
|
| 73 |
|
| 74 |
-

|
| 75 |
-
|
| 76 |
-
*Illustration 2. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.*
|
| 77 |
|
| 78 |
## 4. Coordinate-to-activation computation
|
| 79 |
|
|
@@ -99,9 +118,6 @@ Zero strength, budget, or layer weight gives an exact zero perturbation. Each ch
|
|
| 99 |
|
| 100 |
Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
|
| 101 |
|
| 102 |
-

|
| 103 |
-
|
| 104 |
-
*Illustration 3. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.*
|
| 105 |
|
| 106 |
## 5. Intervention within the transformer
|
| 107 |
|
|
@@ -119,9 +135,6 @@ The module does not search for trigger phrases or recognize an idiom before acti
|
|
| 119 |
|
| 120 |
A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
|
| 121 |
|
| 122 |
-

|
| 123 |
-
|
| 124 |
-
*Illustration 4. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.*
|
| 125 |
|
| 126 |
## 6. Prompts, sampling, and bounded output
|
| 127 |
|
|
@@ -149,9 +162,6 @@ The denominator is the maximum distance across the three-dimensional `[-1,1]` cu
|
|
| 149 |
|
| 150 |
The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
|
| 151 |
|
| 152 |
-

|
| 153 |
-
|
| 154 |
-
*Illustration 5. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.*
|
| 155 |
|
| 156 |
## 8. Architecture, persistence, and privacy
|
| 157 |
|
|
|
|
| 11 |
|
| 12 |
AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
|
| 13 |
|
| 14 |
+
## Application interface
|
| 15 |
+
|
| 16 |
+

|
| 17 |
+
|
| 18 |
+
*Document workspace with an example input, rewrite controls, and an empty output pane before generation.*
|
| 19 |
+
|
| 20 |
+

|
| 21 |
+
|
| 22 |
+
*Geometry editor showing the selected three-point trajectory and its steering controls.*
|
| 23 |
+
|
| 24 |
+

|
| 25 |
+
|
| 26 |
+
*Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.*
|
| 27 |
+
|
| 28 |
+

|
| 29 |
+
|
| 30 |
+
*Expanded numerical weights for the selected geometry, including the local calibration metadata.*
|
| 31 |
+
|
| 32 |
+

|
| 33 |
+
|
| 34 |
+
*Verification controls before evaluation. No verification results have been computed in this view.*
|
| 35 |
+
|
| 36 |
## Literature review: activation steering and emotion
|
| 37 |
|
| 38 |
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.
|
|
|
|
| 93 |
|
| 94 |
There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
|
| 95 |
|
|
|
|
|
|
|
|
|
|
| 96 |
|
| 97 |
## 4. Coordinate-to-activation computation
|
| 98 |
|
|
|
|
| 118 |
|
| 119 |
Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
|
| 120 |
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
## 5. Intervention within the transformer
|
| 123 |
|
|
|
|
| 135 |
|
| 136 |
A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
|
| 137 |
|
|
|
|
|
|
|
|
|
|
| 138 |
|
| 139 |
## 6. Prompts, sampling, and bounded output
|
| 140 |
|
|
|
|
| 162 |
|
| 163 |
The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
|
| 164 |
|
|
|
|
|
|
|
|
|
|
| 165 |
|
| 166 |
## 8. Architecture, persistence, and privacy
|
| 167 |
|
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/figures/screenshot-05.png
ADDED
|
Git LFS Details
|
paper/source.md
CHANGED
|
@@ -11,6 +11,28 @@ November 13, 2026 · Implementation report · Desktop version 0.1.0
|
|
| 11 |
|
| 12 |
AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
|
| 13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
## Literature review: activation steering and emotion
|
| 15 |
|
| 16 |
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.
|
|
@@ -71,9 +93,6 @@ The resulting `B[l]` is a 3 by 1,024 matrix. These are the editable steering-mod
|
|
| 71 |
|
| 72 |
There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
|
| 73 |
|
| 74 |
-

|
| 75 |
-
|
| 76 |
-
*Illustration 2. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.*
|
| 77 |
|
| 78 |
## 4. Coordinate-to-activation computation
|
| 79 |
|
|
@@ -99,9 +118,6 @@ Zero strength, budget, or layer weight gives an exact zero perturbation. Each ch
|
|
| 99 |
|
| 100 |
Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
|
| 101 |
|
| 102 |
-

|
| 103 |
-
|
| 104 |
-
*Illustration 3. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.*
|
| 105 |
|
| 106 |
## 5. Intervention within the transformer
|
| 107 |
|
|
@@ -119,9 +135,6 @@ The module does not search for trigger phrases or recognize an idiom before acti
|
|
| 119 |
|
| 120 |
A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
|
| 121 |
|
| 122 |
-

|
| 123 |
-
|
| 124 |
-
*Illustration 4. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.*
|
| 125 |
|
| 126 |
## 6. Prompts, sampling, and bounded output
|
| 127 |
|
|
@@ -149,9 +162,6 @@ The denominator is the maximum distance across the three-dimensional `[-1,1]` cu
|
|
| 149 |
|
| 150 |
The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
|
| 151 |
|
| 152 |
-

|
| 153 |
-
|
| 154 |
-
*Illustration 5. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.*
|
| 155 |
|
| 156 |
## 8. Architecture, persistence, and privacy
|
| 157 |
|
|
|
|
| 11 |
|
| 12 |
AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
|
| 13 |
|
| 14 |
+
## Application interface
|
| 15 |
+
|
| 16 |
+

|
| 17 |
+
|
| 18 |
+
*Document workspace with an example input, rewrite controls, and an empty output pane before generation.*
|
| 19 |
+
|
| 20 |
+

|
| 21 |
+
|
| 22 |
+
*Geometry editor showing the selected three-point trajectory and its steering controls.*
|
| 23 |
+
|
| 24 |
+

|
| 25 |
+
|
| 26 |
+
*Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.*
|
| 27 |
+
|
| 28 |
+

|
| 29 |
+
|
| 30 |
+
*Expanded numerical weights for the selected geometry, including the local calibration metadata.*
|
| 31 |
+
|
| 32 |
+

|
| 33 |
+
|
| 34 |
+
*Verification controls before evaluation. No verification results have been computed in this view.*
|
| 35 |
+
|
| 36 |
## Literature review: activation steering and emotion
|
| 37 |
|
| 38 |
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.
|
|
|
|
| 93 |
|
| 94 |
There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
|
| 95 |
|
|
|
|
|
|
|
|
|
|
| 96 |
|
| 97 |
## 4. Coordinate-to-activation computation
|
| 98 |
|
|
|
|
| 118 |
|
| 119 |
Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
|
| 120 |
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
## 5. Intervention within the transformer
|
| 123 |
|
|
|
|
| 135 |
|
| 136 |
A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
|
| 137 |
|
|
|
|
|
|
|
|
|
|
| 138 |
|
| 139 |
## 6. Prompts, sampling, and bounded output
|
| 140 |
|
|
|
|
| 162 |
|
| 163 |
The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
|
| 164 |
|
|
|
|
|
|
|
|
|
|
| 165 |
|
| 166 |
## 8. Architecture, persistence, and privacy
|
| 167 |
|