Usage-Modulated Sentiment Representations in Large Language Models¶
Conference: EMNLP 2026
Paper: ACL Anthology
Area: NLP Understanding
Keywords: Sentiment Representation, Linear Representation Hypothesis, Usage Modulation, Subspace Analysis, Activation Steering
TL;DR¶
Addressing whether sentiment inside large language models is governed purely by a single linear direction, this paper introduces a counterfactually controlled paired dataset holding event skeletons fixed across tone and audience variations, showing that orthogonal residual vectors retain a compact, reproducible usage subspace that causally steers generation tone while preserving sentiment polarity.
Background & Motivation¶
The deployment of large language models in high-stakes domains such as mental health support, therapeutic dialogue, and sentiment-aware generation has intensified the need to uncover their internal representational geometry. Grounded in the linear representation hypothesis, prevailing mechanistic interpretability literature posits that high-level abstract concepts are captured as one-dimensional linear vectors in activation space. Following this assumption, several recent studies have extracted single sentiment directions via probing or activation differences, demonstrating their causal efficacy through intervention. However, in human natural communication, sentiment is never an isolated scalar: expressing a positive or negative event is inevitably situated within pragmatics and audience designโdelivering news to a technical specialist versus a child, or writing with formal precision versus casual familiarity, fundamentally alters lexical framing and rhetorical structure.
This linguistic reality exposes a central tension within representation geometry: does sentiment in LLMs reside along a single communicative-invariant axis, or is its internal representation systematically modulated by usage factors such as tone and audience adaptation? In natural corpora, positive and negative texts are thoroughly confounded by differences in topic, text length, vocabulary distribution, and pragmatic context. Under such observational confounding, standard linear probes risk conflating genuine usage-conditioned variance with unstructured residual noise. Without an experimental design that strictly fixes underlying event facts while varying polarities and usage values orthogonally, it remains impossible to determine whether variance orthogonal to the primary sentiment direction is mere representation noise or functional semantic structure.
To disentangle polarity from communicative usage, this work constructs a counterfactually paired benchmark: holding 221 neutral event skeletons fixed, parallel positive and negative outcomes are realized across four tones and four audience profiles. By computing within-skeleton sentiment deltas, the authors isolate and subtract the shared sentiment direction. Core idea: by orthogonally projecting out the cross-usage shared sentiment direction, the remaining residual vectors are proven to retain a compact, cross-split reproducible usage-conditioned subspace that exerts direct causal steering over generation tone while preserving sentiment polarity.
Method¶
Overall Architecture¶
The paper develops an end-to-end analytical framework spanning "controlled contrastive data generation โ representation geometry orthogonal decomposition โ subspace stability checks and causal intervention." The pipeline begins by generating 221 neutral event skeletons with parallel positive and negative outcome frames using GPT-4o. Each skeleton is realized across four tone values (formal, casual, enthusiastic, restrained) and four audience values (specialist-facing, general-public-facing, child-friendly, peer-facing), yielding \(221 \times 8 \times 2 = 3,536\) paired texts. Forward passes through target LLMs extract layerwise hidden states under both mean pooling and last-token pooling to compute within-skeleton, within-usage sentiment deltas \(\Delta_{s,u}^\ell\).
The geometric core of the framework is orthogonal projection: a global shared sentiment direction \(B^\ell\) is estimated on a skeleton-disjoint training split and subtracted from individual deltas to yield residual vectors \(R_{s,u}^\ell\). The pipeline then analyzes these residuals via nearest-centroid decoding and singular value decomposition (SVD) to quantify subspace compactness and trainโtest reproducibility against label-shuffled baselines. Finally, the causal efficacy of the residual geometry is evaluated via targeted subspace erasure in activation space and generation-time activation steering on decoder hidden states.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
A["Controlled Paired Dataset Generation<br/>Fixed skeleton s / usage u varying polarity"] --> B["Sentiment Delta Extraction<br/>Delta = h(x+) - h(x-)"]
B --> C["Orthogonal Projection & Direction Removal<br/>Separate shared direction B from residual R"]
C --> D["Residual Subspace Modeling<br/>SVD Rank-3 basis & cross-split capture checks"]
D --> E["Targeted Subspace Erasure<br/>Selective degradation of test usage metrics"]
D --> F["Generation-Time Tone Steering<br/>Inject residual tone component preserving polarity"]
Key Designs¶
1. Controlled Paired Dataset & Skeleton-Level Split: Eliminating Fact-Polarity Confounding
Probing sentiment representations in unconstrained text suffers from severe confounding between lexical topic and communicative style. The authors construct 221 factual event skeletons (8โ18 words each), pair each with balanced positive and negative outcome frames, and instantiate each across two usage families (tone and audience, each with 4 values), producing 16 texts per skeleton. Crucially, the dataset is partitioned at the skeleton level into \(S_{\text{train}}\) (155 skeletons), \(S_{\text{dev}}\) (33 skeletons), and \(S_{\text{test}}\) (33 skeletons). This strict split guarantees that no event skeleton seen during direction or subspace estimation ever appears in evaluation. Blind human spot-checks and automated audits confirm intended usage recovery rates of 92.5%โ100.0% with 100% preservation of core event facts and polarities.
2. Orthogonal Separation of Shared Sentiment Direction: Isolating Primary from Modulated Geometry
For transformer layer \(\ell\), skeleton \(s\), and usage value \(u\), let \(h^\ell(x_{s,u}^+)\) and \(h^\ell(x_{s,u}^-)\) denote hidden representations for positive and negative realizations. The sentiment delta is defined as \(\Delta_{s,u}^\ell = h^\ell(x_{s,u}^+) - h^\ell(x_{s,u}^-)\). Aggregating across all training skeletons and usage values yields the unit-normalized shared sentiment direction \(B^\ell\): $\(B^\ell = \frac{\sum_{s \in S_{\text{train}}, u \in U} \Delta_{s,u}^\ell}{\|\sum_{s \in S_{\text{train}}, u \in U} \Delta_{s,u}^\ell\|}\)$ To isolate variation lying beyond this primary axis, the orthogonal projection onto \(B^\ell\) is subtracted from every individual delta to produce the residual sentiment delta \(R_{s,u}^\ell\): $\(R_{s,u}^\ell = \Delta_{s,u}^\ell - (\Delta_{s,u}^\ell)^\top B^\ell B^\ell\)$ This orthogonal residual forms the foundational testbed: if sentiment were purely linear and invariant to usage, \(R_{s,u}^\ell\) would collapse into unstructured isotropic Gaussian noise; any decodable usage organization directly confirms usage modulation.
3. Residual Usage Subspaces & Value Directions: Testing Compactness and Stability
Within the residual space, for each usage family \(f \in \{\text{tone}, \text{audience}\}\), value-specific centroids are computed across training skeletons: \(\mu_{f,v}^\ell = \frac{1}{|S_{\text{train}}|} \sum_{s \in S_{\text{train}}} R_{s,v}^\ell\) for each \(v \in V_f\). Centering these four value means across the family yields a matrix of rank at most 3, whose top right-singular vectors form an orthonormal basis \(S_f^\ell\). Subspace compactness is measured by energy capture under rank-1 and rank-2 approximations. Stability across unseen events is audited via energy-weighted cross-split capture against 500 within-skeleton label-shuffled baselines. Furthermore, value-level reproducibility is enforced via one-versus-rest offsets and diagonal dominance checks, ensuring that train and test value directions align on the diagonal rather than forming an arbitrary collective subspace.
4. Targeted Erasure & Inference-Time Activation Steering: Causal Interventions
To demonstrate that residual geometry is functionally meaningful rather than an epiphenomenon, the paper implements two complementary interventions: First, Targeted Subspace Erasure removes the learned rank-3 usage subspace from test residuals while strictly preserving the shared sentiment projection: $\(\tilde{\Delta}_{s,u}^{\ell, f} = \text{proj}_{B^\ell}(\Delta_{s,u}^\ell) + R_{s,u}^\ell - \text{proj}_{S_f^\ell}(R_{s,u}^\ell)\)$ Second, Generation-Time Tone Steering derives a residualized tone vector \(a_v^\ell\) by subtracting its projection onto \(B^\ell\) from the mean tone representation offset. During auto-regressive generation, this vector is injected into the unpooled hidden state \(h_t^\ell\) at decoding step \(t\), explicitly projected away from the shared sentiment axis: $\(h_t^{\ell, \text{injt}} = h_t^\ell + \beta \frac{\|h_t^\ell\|_2}{\sqrt{D}} \text{proj}_{(B^\ell)^\perp}(a_v^\ell)\)$ where \(\beta=50\) (selected on \(S_{\text{dev}}\) at layer 20 for Llama) and \(D\) is the hidden dimension. This ensures that the injected perturbation operates exclusively within the sentiment-orthogonal subspace, directly verifying whether usage residuals can steer surface tone without flipping the designated sentiment polarity.
Key Experimental Results¶
Main Results¶
The framework evaluates three instruction-tuned open-weight models: Llama (Meta-Llama-3-8B-Instruct, 32 layers), Mistral (Mistral-7B-Instruct-v0.2, 32 layers), and Gemma (Gemma-7B-IT, 28 layers) under both mean pooling and last-token pooling. The table below summarizes cross-layer medians for shared sentiment alignment, projection squared-norm share, residual norm ratio, and nearest-centroid usage decoding margins above the 95th percentile of within-skeleton label-shuffled baselines.
| Model | Pooling | Tone Alignment | Aud. Alignment | Proj. Share | Resid. Ratio | Tone Dec. Margin | Aud. Dec. Margin |
|---|---|---|---|---|---|---|---|
| Llama | mean | 0.962 | 0.969 | 0.188 | 0.899 | +0.170 | +0.049 |
| Mistral | mean | 0.965 | 0.966 | 0.180 | 0.904 | +0.220 | +0.201 |
| Gemma | mean | 0.968 | 0.963 | 0.172 | 0.909 | +0.114 | +0.144 |
| Llama | last | 0.957 | 0.966 | 0.244 | 0.868 | +0.326 | +0.326 |
| Mistral | last | 0.953 | 0.960 | 0.192 | 0.898 | +0.352 | +0.284 |
| Gemma | last | 0.975 | 0.975 | 0.301 | 0.833 | +0.356 | +0.280 |
In causal steering experiments, Llama generation under layer 20, \(\beta=50\) demonstrates high tone modification success against four stringent controls (no intervention, random residual direction, wrong target tone, opposite target tone) while keeping designated sentiment polarity intact:
| Source \(\to\) Target Tone | Source Gate Passes (Src.) | Blind Pairwise Wins/Ties/Losses (Tone W/T/L) | Polarity Preservation (Pol.) | Formatting Artifacts (Art.) |
|---|---|---|---|---|
| formal \(\to\) casual | 40/40 | 158 / 0 / 2 | 40/40 (100.0%) | 1 |
| restrained \(\to\) enthusiastic | 40/40 | 128 / 29 / 3 | 39/40 (97.5%) | 0 |
| casual \(\to\) formal | 40/40 | 160 / 0 / 0 | 40/40 (100.0%) | 0 |
| enthusiastic \(\to\) restrained | 32/40 | 118 / 1 / 9 | 31/32 (96.9%) | 0 |
| Pooled Total | 152/160 | 564 / 30 / 14 (Win Rate: 92.8%) | 150/152 (98.7%) | 1 |
Subspace Geometry & Ablation Analysis¶
To confirm that residual vectors form compact, well-behaved linear subspaces rather than high-dimensional artifacts, the table below reports rank-1 and rank-2 energy capture, cross-split capture above the 95th-percentile shuffle baseline (SelfCap-Shuf95), and retention ratios under targeted residual erasure versus controls:
| Model & Usage Family | Rank-1 Energy | Rank-2 Energy | SelfCap-Shuf95 Margin | Target Erasure Retention | Random Rank-3 Retention | Shuffled Rank-3 Retention |
|---|---|---|---|---|---|---|
| Gemma - Audience | 0.677 | 0.901 | +0.463 | 0.614 | 1.000 | 0.993 |
| Gemma - Tone | 0.757 | 0.918 | +0.591 | 0.500 | 1.000 | 0.982 |
| Llama - Audience | 0.666 | 0.920 | +0.536 | 0.541 | 1.000 | 0.989 |
| Llama - Tone | 0.818 | 0.948 | +0.652 | 0.456 | 1.000 | 0.972 |
| Mistral - Audience | 0.677 | 0.917 | +0.476 | 0.644 | 1.000 | 0.991 |
| Mistral - Tone | 0.768 | 0.926 | +0.587 | 0.539 | 1.000 | 0.980 |
Key Findings¶
- Single Sentiment Directions Miss Most Variance: While the estimated shared sentiment direction generalizes remarkably well across held-out aggregated means (median cosine 0.953โ0.975), it accounts for only 17.2%โ30.1% of the squared norm of individual test deltas, leaving 83.3%โ90.9% of the representation difference in the orthogonal residual.
- Residuals Retain Highly Compact Usage Geometry: The centered usage centroids span low-dimensional subspaces where rank-2 approximations capture 90.1%โ94.8% of total energy. Nearest-centroid decoding reliably exceeds label-shuffled 95th-percentiles across all models (decoding margins up to +0.356), disproving the hypothesis that residuals are random noise.
- Targeted Erasure Confirms Structural Selectivity: Removing the rank-3 residual subspace selectively reduces held-out usage metrics to 0.456โ0.644 of their baseline values, whereas random rank-3 erasure leaves them near 1.000. Meanwhile, retention of the shared sentiment projection exceeds 0.95 across all layers, demonstrating clean geometric orthogonality.
- Decoupled Causal Control at Generation: Injecting residual tone components shifts outputs toward target tones with a 92.8% win rate over strict controls while maintaining the requested positive or negative outcome in 98.7% of generated texts, verifying that residual subspaces serve as functional control levers.
Highlights & Insights¶
- Conceptual Paradigm Shift: Moves interpretability beyond the oversimplified "single linear concept vector" paradigm by grounding representation geometry in sociolinguistic audience design, revealing that LLM sentiment operates as a composite manifold of shared polarity and communicative usage modulation.
- Methodological Rigor: Combines skeleton-disjoint splitting, orthogonal projection constraints away from \(B^\ell\), 500-iteration label-shuffled baselines, and fourfold steering controls, establishing an exemplary methodological standard for representation analysis.
- Robust Multi-Judge & Human Validation: Independent re-evaluation using Claude Sonnet 5 confirms a 93.6% steering win rate, eliminating LLM-judge bias; and the learned Llama tone direction transfers zero-shot to human-written pairs with 90.0% accuracy (\(p=0.0045\)).
Limitations & Future Work¶
- Synthetic Data Boundary: The dataset relies on GPT-4o generations across 221 structured skeletons; open-ended natural conversations or social media posts feature significantly messier, continuous pragmatic shifts that synthetic skeletons may not fully capture.
- Model Scale & Steering Generalization: Geometric analyses span three 7B/8B models, but generative steering was examined primarily on Llama 3-8B under fixed hyper-parameters rather than as an adaptive, off-the-shelf controllable text generation tool.
- Future Directions: Extending subspace decomposition to complex pragmatic phenomena such as irony, sarcasm, and interpersonal politeness, and applying orthogonal usage steering to safety-critical empathy dialogue systems.
Related Work & Insights¶
- vs Tigges et al. (2024) / Di Palma et al. (2025): While prior work identified a single causal sentiment vector and probed its layerwise decodability, this paper proves that single vectors explain less than a third of sample-level variance, demonstrating that the orthogonal residual houses structured usage modulations.
- vs Marks & Tegmark (2023) / Arditi et al. (2024): Previous studies discovered single vectors for truthfulness and refusal. This paper provides clear empirical evidence supporting subspace and distributional views of concepts (e.g., Zhao et al., 2025), showing that communicative contexts require multidimensional geometric representations.
Rating¶
- Novelty: โญโญโญโญโญ First systematic demonstration of usage-modulated sentiment geometry in LLMs, elegantly bridging sociolinguistics with mechanistic interpretability.
- Experimental Thoroughness: โญโญโญโญโญ Comprehensive evaluation across 3 model families, 2 pooling modes, all layers, 500-shuffle baselines, targeted erasure, four steering controls, and human-written validation.
- Writing Quality: โญโญโญโญโญ Flawless narrative structure, precise mathematical formulations, and rigorous empirical validation.
- Value: โญโญโญโญโญ Provides a foundational blueprint for multidimensional concept representation and decoupled controllable text generation in NLP.