- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper presents RLE-Reasoner, a three-stage framework for emotion recognition from EEG and EOG with a large language model.
In the first stage, separate EEG and EOG encoders are combined by a gated fusion module (CARF), and the result is passed to a LoRA-tuned Qwen model (4B, 8B, 14B) as a short sequence of input tokens. The model outputs an emotion label, five EEG states, two EOG states, and a short text that the authors call an explanation. The seven states are produced by a fixed rule-based extractor applied to the raw signals and discretized into three levels using statistics from the training subjects. Qwen-Max turns the label and the states into the target texts.
In the second stage, the EEG and/or EOG of a test trial is replaced with a recording from another subject. The generated states are then compared with those computed on the modified input.
In the third stage, the language model is further trained with GRPO, using a reward based on the accuracy of the generated states. Responses whose emotion differs from the first-stage prediction are penalized.
On a private dataset of twelve subjects and seven classes, the 14B model reaches 66.52% emotion accuracy. Post-training raises state accuracy from 63.84% to 74.15% without changing emotion accuracy, and three blinded raters judge the texts of the post-trained model more consistent with the reference states. A further experiment on DEAP evaluates classification only.
Critical strengths
A concrete test for whether the generated evidence reflects the actual signals. A common failure of emotion-reasoning models is producing fluent, physiological-sounding text ("elevated frontal theta, reduced alpha...") that follows from the predicted label rather than from the recorded signals. RLE-Eval addresses this directly: it replaces the EEG, the EOG, or both with a recording from a different subject viewing a different stimulus, chosen so that at least one reference attribute changes, and then checks whether the generated states match the attributes of the signals actually presented. Because the reference is recomputed from the input rather than assumed from the label, agreement becomes meaningful evidence that the model is reading the signal instead of collapsing .
Critical weaknesses
-The main results rely on a single split into eight training subjects, one validation subject, and three test subjects. LOSO evaluation is the standard in cross-subject EEG emotion recognition, and several of the cited baselines use it. The six additional runs in Table 3 cover only the first stage, and the paper does not describe how training and validation subjects were assigned in them.
-In post-training, Qwen-Max assigns a score of 0, 0.5 or 1 for how well the text organizes the evidence, avoids causal overstatement, and handles conflicting evidence. The prompt, the criteria, and how causal overstatement is detected are not given, and the judge is not compared with human ratings. Because the targets are written by Qwen-Max, the judge is Qwen-Max, and the trained model is also a Qwen model, the score may reward similarity of style rather than correctness, since language models tend to favor text resembling their own (Panickssery et al., 2024).
-The thresholds at +-0.75 robust standard deviations imply that, for a roughly Gaussian attribute, about 55% of trials fall in the "typical" state. A model that always predicts "typical" could therefore match or exceed the 4B model (44.83%) and the 14B accuracy on posterior alpha (46.52%). No such baseline is reported.
-The authors should train a simple classifier on the frozen fused features to predict the seven states. Without it, one cannot tell whether the language model is limited by the information it receives or by its own decoding. The encoders are trained only to classify emotions, they receive inputs normalized per trial, and they are frozen afterwards. Meanwhile, amplitude based attributes, such as the EOG and the evoked responses, are computed before normalization. Part of the information to be predicted may simply be absent from the model's input.
-Replacement recordings always come from another person, and several attributes, such as EOG amplitude or posterior alpha, vary strongly between individuals. A model could "follow" the replacement by recognizing the donor's typical level rather than by reading the single trial. A baseline that predicts each state from the subject's average would clarify this. It also remains unexplained why accuracy on modified inputs is higher than on original ones for all model sizes, and how often replacing only the EEG also changes the generated EOG states.
-The paper calls the generated text an explanation of the emotion judgment. However, the text verbalizes the extractor's states and the label, following targets written by Qwen-Max, and nothing connects it to how the model reaches its decision. An explanation, in the usual sense, should reflect the reasons for the prediction. What is evaluated here is closer to a description of the input. Two design choices reinforce this gap. The emotion is generated before the states, so the states are produced already knowing the label. Second, the reward ties the emotion to the first-stage prediction. This makes the preservation of emotion accuracy largely a consequence of the design rather than an empirical result, and it prevents post-training from making the decision more consistent with the evidence. No experiment tests whether the reported evidence affects the prediction, or whether the seven single-trial attributes are related to the emotion categories at all.
-Finally, the reward measures agreement with states the authors already know. The first stage never saw modified inputs, so the gain could come from simply training on this new kind of input. A comparison with standard supervised training on the same 13440 modified instances is needed. Without it, the specific contribution of the third stage cannot be assessed.
Minor
Two citations are missing: one in the Introduction, after "(Ma et al., 2026; Jiang et al., 2024; ?)", and one in Related Work, after "emotion recognition (?)".
The language model is described only as "Qwen 4B/8B/14B", without version or citation. The available sizes and Appendix H suggest Qwen3, but this is never stated for the main experiments. The paper also does not say which layers LoRA adapts or where the input tokens are placed in the prompt.
Qwen-Max, a proprietary model available only through an online service, is used both to write the training targets and to score texts during post-training. Neither its version, the date of use, nor the prompts are reported.
The main text and the appendix disagree on the number of post-training instances (13440 versus 400 in Table 4) and on how the reward is computed when no attribute is left unchanged.
Questions for authors
1.Can the second and third stages be evaluated with LOSO? For the six runs in Table 3, how were training and validation subjects assigned?
-
What prompt and criteria does Qwen-Max use for the soft reward, and how is causal overstatement detected? Was the judge compared with human ratings, or with a judge from a different model family?
-
What accuracy and Macro-F1 does a model that always predicts "typical" achieve, overall and for each attribute? What does a simple classifier trained on the frozen fused features achieve on the seven states?
-
What accuracy is obtained by predicting each state from the subject's average? Why is accuracy higher on modified inputs than on original ones, and how often do the generated EOG states change when only the EEG is replaced?
-
Does the predicted emotion change when the generated evidence changes? Are the seven single-trial attributes predictive of the emotion categories?
The autors could generate the evidence before the emotion, and letting the emotion vary during post-training. This would allow a direct test of whether the decision depends on the evidence.