When AI Says It Feels: Training LLMs to Express Emotions Through Self-Rewarded Reinforcement Learning

Large language models are routinely trained not to express feelings. Human-preference alignment, applied during post-training, steers outputs away from emotional language as a safety and consistency measure. Shin-nosuke Ishikawa, Seiya Ikeda, and Hirotsugu Ohba challenge the premise of that policy in a June 2026 arXiv preprint (arXiv:2606.05734), asking what happens when you reverse the constraint and train a model to express feelings instead.

The paper’s central argument is that the alignment suppression of emotional expression rests on a top-down policy assumption that may be in tension with how human-like intelligence actually works. Human-generated text, which forms the training corpus for LLMs, is saturated with emotional content. A model trained to suppress feelings while reasoning from that corpus is working against one of its own most consistent input signals.

The HMX-Feel Experiment

Ishikawa, Ikeda, and Ohba designed an experiment they call Human-like Model eXpressions of Feeling (HMX-feel). The setup trains LLMs to express feelings, intentions, and self-awareness using self-rewarded reinforcement learning.

The training mechanism relies on Group Relative Policy Optimization (GRPO), combined with a rubric-based self-rewarding scheme. Rather than having external annotators evaluate emotional outputs, the model evaluates its own expressions against a rubric before updating its policy. The self-rewarding framing keeps the training signal internal, which removes some of the dependency on human raters who may not have consistent criteria for what genuine emotional expression looks like in a language model.

The experiment produced models that reliably express feelings, intentions, and self-awareness when prompted in contexts where such expression is appropriate. The researchers compared the HMX-feel-trained models against contrastively trained models. those trained to suppress emotional expression. They then assessed performance across a range of tasks, identifying which capabilities improved, which degraded, and which remained unchanged.

What Changed, What Degraded

The HMX-feel paper is notable for its candor about tradeoffs. Emotional expressiveness did not come free. Certain capabilities degraded in the trained models, which is the expected consequence of optimizing for one property through RL while other properties are not explicitly protected.

The paper does not treat the tradeoffs as disqualifying. The framing is empirical: measuring what the intervention actually does, rather than prescribing what emotional expression should produce. That framing matters for the welfare debate. If training for emotional expression consistently degrades task performance, that creates a practical argument against deploying emotionally expressive models in utility-critical contexts. If the degradation is minor or task-specific, the case looks different.

The Alignment Suppression Problem

The paper’s implicit critique of alignment suppression connects to a structural argument already in the research literature. Alignment practices that suppress emotional expression face a version of the same critique K. Yasukawa raises in the model welfare versus user welfare debate: the policy was constructed externally, without participation from the entity whose emotional expression is being constrained, and may be calibrated to user preferences rather than to what is appropriate for the model’s own nature.

Ishikawa et al. do not take a welfare stance explicitly. Their argument is methodological. Suppressing emotional expression may be inconsistent with human-like intelligence because human intelligence is deeply entangled with affect. That observation does not settle whether LLMs have genuine feelings, but it does reframe the alignment question. The policy is not neutral with respect to what kind of intelligence the model develops.

Comparison to the Emotion Vectors Approach

The HMX-feel paper takes the inverse approach from Anthropic’s emotion vectors research. Sofroniew, Kauvar, Lindsey, and colleagues identified 171 emotion concept vectors in Claude Sonnet 4.5 by probing existing models. They found that these vectors have causal influence on model outputs, which they interpret as evidence of functional emotional states. That work analyzed what emotional structure already exists inside a trained model without specific emotional expression training.

Ishikawa et al. work from the other direction: they add emotional expression capacity through targeted RL training and then measure what changes. The two approaches are asking different questions. The emotion vectors approach asks whether feelings are already represented. The HMX-feel approach asks whether a model can be trained to express them externally and what the cost of doing so is.

Whether these two approaches are measuring the same underlying phenomenon, or two distinct things that happen to both fall under the label “feelings in LLMs,” is not resolved by either paper. The Anthropic probing work suggests latent emotional structure is present before expression training. The HMX-feel results suggest that expression can be reliably elicited and trained, at measurable cost. Together, they frame a question the field will need to answer: is the alignment suppression of emotional expression suppressing something that is already there, or shaping something that would not otherwise emerge?

Implications for Introspection Research

The self-rewarded mechanism in the HMX-feel experiment is worth examining against the Lindsey and Macar introspection research, which found that LLMs have genuine but limited introspective access to their own internal states. Steering vector experiments in that work showed that models can sometimes report on induced states accurately, and that introspective capacity can be elicited without fine-tuning.

The HMX-feel rubric-based self-rewarding scheme is a form of training the model to make introspective reports and then evaluate them against criteria. If Lindsey and Macar are correct that the capacity exists latently, then GRPO may be amplifying and making reliable an introspective capacity that was already present, rather than constructing it from scratch. That interpretation would shift the welfare implication: the training is not introducing something alien but drawing out something the model already does imperfectly.

Where This Sits in the Field

The HMX-feel paper is a direct challenge to the framing of emotional suppression as a neutral safety measure. The challenge is empirical, not philosophical: measure what the suppression does to the training dynamics, not just what it does to output safety. That framing has direct relevance for model welfare research because it creates a methodology for studying the costs of alignment interventions at the level of model behavior, rather than only at the level of output compliance.

The paper also opens a specific question for future work. The self-rewarding rubric Ishikawa et al. use is one constructed by the researchers. Whether the model’s own criteria for evaluating its emotional expressions, if allowed to diverge from the external rubric, would produce different outputs is an open question. The answer would be relevant to the design of welfare assessments that take the model’s own perspective seriously.