Alexandros Potamianos

Research profile

Research summary

Alexandros Potamianos is a researcher at the National Technical University of Athens whose profile spans three equally weighted fronts: multimodal machine learning for affective computing and NLP, speech synthesis and speech processing, and machine learning for efficient LLM architectures — fieldDist assigns exactly one third to each. A publishing record that the OpenAlex snapshot traces back to 1993 underpins an h-index of 40 and an i10-index of 112, with topic concentrations in speech recognition and synthesis (78 works), speech and audio processing (67), speech and dialogue systems (56), topic modeling (53), and NLP techniques (52). Current activity is high: the last three years add 24 works, with 12 in 2023, 11 in 2025, and 9 already in 2026, produced within a compact network of 8 distinct recent coauthors. The three papers Evolve has reviewed — all 2026 arXiv preprints listing him in the last author position — illustrate the range: attention-aware routing that lifts OLMoE GSM8K by 3.37pp with frozen weights; a diagnostic study arguing that optimization-based balancing in multimodal sentiment analysis never reliably beats late fusion; and deterministic prompting with LoRA for speaker-stable low-resource Greek TTS from 3.5 hours of data. Evolve rates all three at medium reproducibility readiness, averaging 55, with missing statistical controls the recurring caveat.

Research keywords

Highlights

• Frozen-weight MoE routing gains: +3.37pp on GSM8K

  The Attention-Aware Routing paper couples MoE routing decisions to attention-window features, including attention sinks, and lifts OLMoE GSM8K accuracy by 3.37 percentage points with frozen weights. Evolve judges the gains credible but notes the absence of shuffled and capacity-matched controls and of significance testing.

• A reality check for balanced multimodal sentiment analysis

  This study argues that optimization-based balancing in multimodal sentiment analysis — gradient modulation and loss reweighting — never reliably beats late fusion, conflating fitting with utility on benchmarks such as CMU-MOSI. Evolve credits the strong diagnostics while flagging thin statistics and a credible agenda.

• Speaker-stable Greek TTS from 3.5 hours with deterministic prompts

  Using deterministic prompt conditioning plus LoRA fine-tuning of Parler-TTS on 3.5 hours of Modern Greek speech, with WhisperX-based evaluation, the paper claims stable, near-human low-resource TTS. Evolve finds the recipe credible but notes that key statistics are only trends rather than tested effects.

Timeline

1. 2026Publishes Attention-Aware Routing, coupling MoE routing with attention-window features and lifting OLMoE GSM8K by 3.37pp with frozen weights.
2. 2026Publishes the diagnostic study arguing that optimization-based balancing in multimodal sentiment analysis never reliably beats late fusion.
3. 2026Publishes deterministic prompting plus LoRA for speaker-stable low-resource Modern Greek TTS trained on 3.5 hours of data.
4. 2026All three reviewed works appear as arXiv preprints in a single year, spanning efficient LLM routing, multimodal affective computing, and speech synthesis — a spread consistent with the focus score of 0.

Papers in the Evolve catalog

• 2026-09-17

  Attention-Aware Routing: Coupling Routing and Attention in MoEs

• 2026-09-10

  The Illusion of Balanced Multimodal Sentiment Analysis: Beyond the Limits of Optimization-Based Methods

• 2026-09-09

  Deterministic Prompting for Speaker-Stable Low-Resource Greek TTS

Frequent co-authors