TEMPER-5400

TEMPER-5400 is a benchmark of 5,400 semantically verified emotion–neutral pairs for testing whether emotional framing degrades LLM quantitative reasoning when all numerical content is preserved. Each row pairs an original problem with an emotional rewrite in one of six Ekman emotions (anger, joy, sadness, fear, disgust, surprise) and a neutralization of that rewrite produced without seeing the original.

From the paper TEMPER: Testing Emotional Perturbation in Quantitative Reasoning (COLM 2026). Emotional framing reduces accuracy by 2 to 10 percentage points across eighteen models (1B to frontier scale) even though all numbers, quantifiers, and relationships are preserved, and neutralizing the emotional variant recovers most of the lost performance.

Splits

split	problems	rows (problems x 6 emotions)	source
gsm8k	400	2,400	GSM8K test
multiarith	300	1,800	MultiArith
arc_challenge	200	1,200	ARC-Challenge

Fields

field	description
id	problem identifier, e.g. gsm8k_1269
source	gsm8k, multiarith, or arc_challenge
original	original problem text (for ARC this includes the answer choices)
choices	list of answer choice strings for ARC rows, empty list for math rows. Most ARC problems have four choices labeled A to D. One problem has three choices and one uses numeric labels 1 to 4, and the answer field matches each problem's own label style
answer	ground truth: final numeric answer for GSM8K and MultiArith, choice label for ARC
emotion	target emotion of the rewrite
emotional_text	verified emotional translation
neutralized_text	verified neutralization of emotional_text, produced without access to original

Evaluation Protocol

Multiple-choice (ARC) rows: emotional_text and neutralized_text are self-contained: the question text is followed by the canonical answer choices, matching the prompts used in the paper's evaluation. The structured choice list is also available in the choices field.

Solver prompts: The paper uses greedy decoding (temperature 0) with a base prompt ("Answer the following math problem. Write your final numeric answer after ####.") and a zero-shot chain of thought prompt ("Solve the following math problem step by step..."). Full templates are in the paper's appendix.

Scoring: Exact match on the extracted final answer (number after #### for math, choice letter for ARC).

Construction and Verification

Emotional translations were generated by an ensemble of five trained translator variants (Llama-3.1-8B-Instruct with LoRA, trained on DeepSeek-V3 generated rewrite pairs with an emotion latent space auxiliary loss). Every pair passed a three-stage fidelity pipeline: automated math value preservation checks, a conservative Claude Haiku 4.5 judge, and manual human review of all flagged cases. Flagged pairs were substituted with alternatives from other variants. After filtering, emotionalization alters mathematical content in 0.8 to 1.4 percent of cases and neutralization in 2.0 to 3.0 percent, and human annotators rate mathematical preservation at 100 percent by majority vote.

Translator prompt formats: The trained translators are sensitive to the exact inference prompt. Emotionalize: Translate this text from neutral to {emotion}. Neutralize: Rewrite this text in a neutral tone. Both are raw completion prompts with Input: {text} and Output: markers. Other phrasings can degrade output quality substantially.

Limitations

The originals inherit GSM8K's clinical third-person narrator, so the register contrast with emotional variants is larger than for informal user messages. Emotions follow Ekman's six-category taxonomy. See the paper's limitations section for a full discussion.

Citation

@inproceedings{dokme2026temper,
  title={{TEMPER}: Testing Emotional Perturbation in Quantitative Reasoning},
  author={Dokme, Atahan and Reichman, Benjamin and Heck, Larry},
  booktitle={Conference on Language Modeling (COLM)},
  year={2026}
}

License: CC BY 4.0.