
#MachineTranslation Research Unit @FBK_research. #nlproc #deeplearning #ai
Trento, Trentino-South Tyrol
Joined October 2015
- Tweets3.3K
- Following450
- Followers1.4K
- Likes3K
Late update but we had two great seminars in October by:
🎤 Alice Ross, a PhD canditate at University of Edinburgh on "Perceived characteristics of synthesised speech in social context"
1
4
100
Our #PickOfTheWeek 📚: “Training Dynamics Impact Post-Training Quantization Robustness” by @actatjer, @n_ajroldi, @jonasgeiping
Pick of the week @fbk_mt
Cool paper on how robustness to quantization in LLMs is also impacted by training conditions, instead of just the number of training tokens!
arxiv.org/abs/2510.06213
1
4
171
🛠️ By analyzing optimization hyperparameters and checkpoint evolution, they demonstrate how smoother training dynamics act as an implicit regularizer, yielding models that are naturally more resilient to aggressive weight and activation compression.
1
24
📄 Paper: arxiv.org/abs/2510.06213
20
🧩 The authors derive word-level “contextual alignments” from an MT model’s perplexity and enforce them by inserting silences 🤫 or resynthesizing targets with an alignment-aware TTS.
1
76
📝 The model also predicts aligned text tokens as scaffolding for speech, and 🗣️ conditional training plus classifier-free guidance improve voice transfer.
📄 Paper: arxiv.org/abs/2502.03382
57
Can LLM Agents Learn how to behave According to your preferences?
PersonaAgent is a framework that combines personalized memory and tool use to help LLM agents adapt their actions to individual users without fine-tuning the underlying model.
1
3
132
In this case, personalization goes beyond remembering user information: agents can use past interactions to adapt their reasoning and tool usage. This solution outperforms RAG-based personalization and general-purpose agents across four personalized decision-making tasks.
1
1
52
🔗 Paper link: PersonaAgent: Bridging Memory and Action for Personalized LLM Agents: aclanthology.org/2026.findin…
#PickOfTheWeek #Personalization #LLM #AIAgents #MTUnit #FBK
30
Our #PickOfTheWeek, selected by @mgaido91: "Alignment Is All You Need: Instruction-Free Training for General Audio-Language Models" by Xuanru Zhou, Yiwen Shao, Jiahong Li, and Dong Yu 🎧
Can instruction-free training power the next generation of audio-language models?
Cool to see we are not the only ones working on building SpeechLLMs with instruction-following capabilities without requiring LLM adaptation and instruction-following data that is expensive to collect!
arxiv.org/abs/2608.18132
@fbk_mt
1
5
128
This paper shows how alignment alone can streamline training and improve real-world evaluation beyond lab benchmarks.
#AudioAI #VoiceAI #MachineLearning #LLM #AIResearch
1
37
Our #PickOfTheWeek, selected by @BeatriceSavoldi:
"The AI Observatory: A Public Measure of Real-World AI Use" by @ShayneRedford, @AnkaReuel, @zoeykii et al.
A crucial framework tracking how AI models are deployed, adopted, and evaluated beyond lab benchmarks.
Pick of the week @fbk_mt
📚"The AI Observatory: A Public Measure of Real-World AI Use"
A study and taxonomy to explore real conversations across multiple datasets.
dataprovenance.org/ai_observ…
1
8
231
Can algorithmic gender prediction ever be valid?
Check out this week's top pick by @lina_conti: "Algorithmic Gender Prediction Is Illegitimate, But Gender Imputation Can Yield Valid Measurements" by @evandongyx & @ang3linawang.
Pick of the week by @evandongyx & @ang3linawang:
arxiv.org/pdf/2608.13444
Predicting gender from images or names can reveal discrimination. But the practice itself harms trans people. This paper works through when that tradeoff might be justified and how to do it responsibly.
1
2
175






