What changes when an AI writes each sentence with its own sentiment target?

Most controllable text generation today runs on decoder models. They predict the next token left to right. That works for fluency. It fails when a user wants the third sentence of a story to carry a precise negative intensity while the fifth stays mildly positive.

Accepted to EMNLP 2026, the work introduces SenseShift. The authors are Shahed Masoudian, Markus Frohmann, Emmanouil Karystinaios, Navid Rekabsaz, and Markus Schedl. They frame the problem as mask infilling rather than next-token prediction.

Why decoder models hit a wall

Decoder architectures use causal attention. Each token sees only what came before. Sentiment control in these systems usually means prepending a label or steering the hidden state globally. The paper notes two common simplifications. One treats sentiment as a coarse category — positive or negative. The other applies a single fine-grained signal to the whole document. Neither handles sentence-level variation inside a long text.

The authors write that more challenging settings such as sentence-level sentiment control within long-form text remain underexplored. Decoder models struggle to satisfy complex constraints and follow fine-grained sentiment signals specified by users.

How the encoder approach differs

SenseShift uses bidirectional attention. Every position sees the full context. The framework quantizes sentiment signals into discrete buckets. It then iteratively infills masked spans conditioned on the target intensity for each sentence.

The process works in rounds. A mask covers the tokens of a target sentence. The model predicts those tokens given the surrounding text and the desired sentiment value. The mask moves to the next sentence. Repeat until the document is complete.

Quantization turns a continuous sentiment scale into a manageable set of control codes. The paper does not specify the exact number of buckets. It says quantized sentiment signals — plural — without a count.

Bidirectional attention means the model can adjust the beginning of a sentence to match its ending. Decoder models cannot do that without lookahead tricks.

What the experiments measure

Evaluations cover story generation and review generation. The baselines are larger decoder-based models. SenseShift achieves stronger sentiment controllability while maintaining text quality and robustness to out-of-domain generation.

Exact metric values do not appear in the abstract. It states the comparative result in qualitative terms. No human evaluation scores appear in the summary. No compute budget or training data size is mentioned.

Robustness to out-of-domain generation means the method holds up when the test topics differ from the training topics. The abstract does not define the domains used.

Where the explanation stops

The abstract does not describe the encoder size. It does not name the base model — BERT, RoBERTa, or a custom architecture. It does not say how many iterations the infilling loop runs. It does not mention whether the quantization is learned or fixed. Code and data links exist on the arXiv page but are not summarized in the text.

The submission date is 25 Aug 2026. The version is v1. The DOI is 10.48550/arXiv.2608.24304. The paper is 951 KB.