What does a reward model really learn?
Testing the gap between reward-score accuracy, policy agreement, and action correctness in language-conditioned navigation.
1 paper
Machine learning researcher · Hyderabad, India
Anandaswarup Vadapalli
I test whether generative model outputs are actually correct, using controlled benchmarks, statistical analysis, and human review.
I have spent over 15 years researching speech, language, and multimodal AI, from building neural text-to-speech systems to evaluating large language and audio models. I teach the Machine Learning track of the MSIT program at CIHL, IIIT Hyderabad Campus, and mentor applied research at ResearchSpark Hub.
Latest work · Submitted · ERC-LM @ ICASSP 2027
Reference Agreement Is Not Acoustic Correctness: Evaluating Large Audio Language Model Captions on Clotho
Testing whether reference-based metrics agree with listening when ranking large audio language model captions on Clotho. In a blind listening review, the model ranked third or fourth by reference-based metrics was rated accurate on 47 of 50 recordings.
4,180captions from four LALMs
14reference-based and audio-aware metrics
47 / 50GPT-Audio-1.5 captions rated accurate by listening
News
- Submitted
Sole-author paper on evaluating large audio language model captions: reference agreement is not acoustic correctness. ERC-LM Workshop, ICASSP 2027.
- Accepted
Three papers with mentees at ACDSA 2027, on reward functions for robots, LLM image captions, and brain MRI explanations.
- Preprint
The hidden costs of translation accuracy: distillation, quantization, and carbon impact for NLLB-200.
- Published
Phrase break prediction in an end-to-end TTS system, SN Computer Science, 2025.
- Ph.D.
Awarded a Ph.D. in Computer Science and Engineering by IIIT Hyderabad, March 2025. Read the thesis (PDF).
Research
All research threads →The hidden costs of translation accuracy
Understanding the trade-offs between translation quality, inference latency, and estimated emissions.
1 paper
Giving synthetic speech room to breathe
A research thread from word-terminal syllables and learned representations to phrase-aware neural text-to-speech.
4 papers
Selected publications
All publications →SubmittedReference Agreement Is Not Acoustic Correctness: Evaluating Large Audio Language Model Captions on Clotho
ERC-LM Workshop at ICASSP 2027
PreprintThe Hidden Costs of Translation Accuracy: Distillation, Quantization, and Environmental Impact
arXiv preprint · 2025
An Investigation of Phrase Break Prediction in an End-to-End TTS System
SN Computer Science · 2025
An Investigation of Recurrent Neural Network Architectures Using Word Embeddings for Phrase Break Prediction
17th Annual Conference of the International Speech Communication Association, Interspeech 2016, San Francisco, CA, USA, September 8-12, 2016