EvoRiri: Interpretable Evolutionary Search for Diagnosing Reward Design Failures in a Bangla Mental Health Assistant
- 1Department of Computer Science & Engineering, Rajshahi University of Engineering & Technology, Rajshahi 6204, Bangladesh
- 2Transdisciplinary School, University of Technology Sydney, 15 Broadway, Ultimo, NSW 2007, Australia
Abstract
Mental health conversational systems must balance empathy, risk-aware safety, and actionable structure, yet optimising hand-designed proxies for these qualities can expose reward design failures. We present EvoRiri, a template-based Bangla–Banglish–English mental health assistant that uses no language model. Its response policy is an eight-gene mixed continuous–discrete genome (empathy, safety and structure weights, a distress-gating gene, two reserved escalation thresholds, a template selector, and a response-part budget). A genetic algorithm (population 100, 20 generations, tournament selection, and uniform crossover, with 2000 evaluations per run) maximises a scalar fitness that weights rule-based empathy, safety and structure scores (0.40/0.40/0.15) with per-example safety penalties. On 2872 de-identified records rendered as synthetic prompts, the evolved genome raises structural completeness from 0.65 to 1.00 and fitness from 0.556 to 0.703 over a fixed reference genome, with comparable safety. A one-factor ablation attributes this gain to a single structure-weight threshold, and over 30 paired seeds, genetic and equal-budget random search are equivalent on validation prompts. Trace analysis exposed failures that aggregate fitness hid, including question-only collapse, empathy-marker stuffing, and response truncation. The contribution is a diagnostic case study of reward design on a synthetic single-turn corpus, not an optimisation advantage or clinical validation.
Related
- LOONA: An LLM Assistant for Spanish Language User Story Quality Assurance in Agile Development
- Strong failures of club guessing at the successor of a regular cardinal
- The Cauchy–Modified Garima Distribution: Applications to Device Lifetimes and Aircraft Windshield Failures
- Optimal Strategies in a Sequential Contest with Failures
- Prescribed and simultaneous failures of log-concavity in trees with perfect matchings
BibTeX
@article{rahman2026evoriri,
title = {{EvoRiri: Interpretable Evolutionary Search for Diagnosing Reward Design Failures in a Bangla Mental Health Assistant}},
author = {Jahnnobi Rahman and Mir Md. Jahangir Kabir},
journal = {Mathematics},
year = {2026},
month = nov,
doi = {10.3390/math14213982},
url = {https://www.mdpi.com/2227-7390/14/21/3982},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - EvoRiri: Interpretable Evolutionary Search for Diagnosing Reward Design Failures in a Bangla Mental Health Assistant
AU - Jahnnobi Rahman
AU - Mir Md. Jahangir Kabir
PY - 2026
DA - 2026/11/01
JO - Mathematics
DO - 10.3390/math14213982
UR - https://www.mdpi.com/2227-7390/14/21/3982
AB - Mental health conversational systems must balance empathy, risk-aware safety, and actionable structure, yet optimising hand-designed proxies for these qualities can expose reward design failures. We present EvoRiri, a template-based Bangla–Banglish–English mental health assistant that uses no language model. Its response policy is an eight-gene mixed continuous–discrete genome (empathy, safety and structure weights, a distress-gating gene, two reserved escalation thresholds, a template selector, and a response-part budget). A genetic algorithm (population 100, 20 generations, tournament selection, and uniform crossover, with 2000 evaluations per run) maximises a scalar fitness that weights rule-based empathy, safety and structure scores (0.40/0.40/0.15) with per-example safety penalties. On 2872 de-identified records rendered as synthetic prompts, the evolved genome raises structural completeness from 0.65 to 1.00 and fitness from 0.556 to 0.703 over a fixed reference genome, with comparable safety. A one-factor ablation attributes this gain to a single structure-weight threshold, and over 30 paired seeds, genetic and equal-budget random search are equivalent on validation prompts. Trace analysis exposed failures that aggregate fitness hid, including question-only collapse, empathy-marker stuffing, and response truncation. The contribution is a diagnostic case study of reward design on a synthetic single-turn corpus, not an optimisation advantage or clinical validation.
SN - 2227-7390
ER -