Universidad de San Andrés · Buenos Aires Department of Engineering

Efficient
Language
Intelligence.

The ELIAS Lab builds efficient language-model methods, systems, and orchestrations — smaller, faster, cheaper, greener — to make advanced AI more accessible and easier to deploy. We study the full stack: synthetic data, distillation, adaptive inference, cost-aware algorithms, and agentic retrieval.

ACL 2026 · 2 papers accepted Agentic AI executive program · 2nd edition starts Sep 24 AI en la Trinchera · next session Sep 22 ACL 2026 · 2 papers accepted Agentic AI executive program · 2nd edition starts Sep 24 AI en la Trinchera · next session Sep 22

What the lab works on.

Orchestration efficiency · Accepted at ACL 2026

Layer-selective hidden-state probes

Production LLM stacks pay for a full forward pass to generate, then pay again for separate guard and classifier models. We reuse the serving model's hidden states: lightweight probes (as few as 100K parameters) aggregate signals across tokens and layers to match guard-model accuracy on ToxicChat and WildGuardMix — inside a single forward pass. Validated across Llama-3.2, GPT-OSS-20B, and Qwen3-30B.

Agentic retrieval · Accepted at ACL 2026

ARK — adaptive retriever of knowledge

A training-free agentic retriever that gives a language model control over the breadth–depth tradeoff on knowledge graphs, using only two tools: global lexical search and one-hop neighborhood expansion. Reaches 59.1% avg Hit@1 on STaRK, improving over prior retrieval-based and agentic baselines by up to 31.4 points. Distilled into an 8B model via label-free imitation. Joint work with Harvard Medical School and Oxford.

Runtime monitoring

Entropy Sentinel

A cheap inference-time signal for production reliability: entropy trajectories from top-k log-probabilities, summarized into a compact feature vector, feed a lightweight classifier that estimates slice-level accuracy under domain shift. Evaluated across ten STEM reasoning benchmarks and nine LLMs (3B–20B). Tracks held-out accuracy well enough to prioritize data acquisition toward the worst-performing slices.

Cost-aware ranking

Active learners as PRP rerankers

Pairwise ranking prompting pairs an LLM judge with a classical sorting algorithm — but LLM comparisons are noisy, order-sensitive, and sometimes intransitive, so sorting wastes budget. We reframe reranking as active learning from noisy pairwise comparisons, and introduce a one-call randomized-direction oracle that converts position bias into zero-mean noise. 3–9× fewer LLM calls than sorting PRP at matched NDCG@10.

Selected work from lab members.

2026
Autonomous Knowledge Graph Exploration with Adaptive Breadth-Depth Retrieval Polonuer, Vittor, Arango, Noori, Clifton, Del Corro, Zitnik · Harvard Medical School / Oxford / UdeSA
ACL 2026
2026
Active Learners as Efficient PRP Rerankers Figueiredo Paschmann, Kaplan, Nattero, Barron Bucolo, Wisznia · Del Corro, L.
arXiv
2025
Are Optimal Algorithms Still Optimal? Rethinking Sorting in LLM-Based Pairwise Ranking with Batching and Caching Wisznia, Bolaños, Tollo, Marraffini, Gianolini, Hsueh · Del Corro, L.
ACL 2025
2024
The Greatest Good Benchmark: Measuring LLMs' Alignment with Utilitarian Moral Dilemmas Marraffini, Cotton, Hsueh, Fridman, Wisznia · Del Corro, L.
EMNLP 2024
2024
AgentInstruct: Toward Generative Teaching with Agentic Flows Mitra, Del Corro, Zheng, Mahajan, Rouhana, Codas, et al. · Microsoft Research
Microsoft
2023
Orca 2: Teaching Small Language Models How to Reason Mitra, Del Corro, Mahajan, Codas, Simoes, Agarwal, et al. · Microsoft Research
Microsoft
2023
SkipDecode: Autoregressive Skip Decoding with Batching and Caching for Efficient LLM Inference Del Corro, Del Giorno, Agarwal, Yu, Awadallah, Mukherjee · Microsoft Research
Microsoft

The people behind the lab.

Andrés Arpi

NLP Co-Instructor
LinkedIn ↗

Juan Wisznia

Research Assistant
LinkedIn ↗

Agustín Gianolini

Teaching Assistant
LinkedIn ↗

Teo Gutter

Teaching Assistant
LinkedIn ↗

Ezequiel Ponce

Teaching Assistant

Gonzalo A. Meyoyan

Undergrad student member · Hidden-State Probes

Jeremías Figueiredo Paschmann

Undergrad student member · Active Learners
LinkedIn ↗

Juan Kaplan

Undergrad student member · Active Learners

Francisco Nattero

Undergrad student member · Active Learners

Santiago Barron

Undergrad student member · Active Learners
Undergraduate thesis students

Matias Antenaza · Alex Bodner · Santiago Carrillo · Josefina Dehan · Nataly Sol Hofkamp · Segundo Santos · Ignacio Shuemer · Antonio Tepsich · Agustina Videla Rivero · Mora Vigo Malusardi · Facundo Vulcano

What the lab is up to.

2026

Two papers accepted at ACL 2026

ARK (adaptive breadth-depth retrieval on knowledge graphs, with Harvard Medical School and Oxford) and Hidden-State Probes (single-pass classification, Meyoyan & Del Corro) — both main conference, long papers.

September 24, 2026

Agentic AI executive program returns

The first edition sold out. A second edition of the UdeSA executive program starts September 24.

Sep – Nov 2026

AI en la Trinchera announces new sessions

Guillermo Gaete (Tiendanube) on September 22, Eugenio Leiguarda (Vexta) on October 20, and Javier Cardoso on November 3 — all at Campus UdeSA, Victoria.

AI en la Trinchera.
Practitioners, on the record.

A talk series at UdeSA where people running AI systems in production tell it like it is. Organized by Franco Zan. Sessions at Campus UdeSA, Vito Dumas 284, Victoria · 14:40–16:10.

Upcoming
Guillermo GaeteTiendanube
Sep 22, 2026
Upcoming
Oct 20, 2026
Upcoming
Javier Cardoso 
Nov 3, 2026

Courses & programs.

Executive · UdeSA Educación Ejecutiva · Online + Presencial

Agentic AI — de la teoría a la producción

Programa para profesionales que quieren construir sistemas agénticos en producción. Cubre prompt engineering, RAG basado en agentes, function calling, el protocolo MCP, sistemas multi-agente y despliegue, con un hackatón final sobre un caso de negocio real. Coordinación académica: Luciano Del Corro. Profesor: Juan Wisznia.

Primera edición completa (sold out) · Segunda edición comienza el 24 de septiembre de 2026 · Inscripción ↗
Academic courses
Procesamiento de Lenguaje Natural
Universidad de San AndrésIng. en Inteligencia Artificial
Grado
Procesamiento de Lenguaje Natural
Universidad de San AndrésMaestría en Inteligencia Artificial
Maestría

Final-year projects (tesinas) from Luciano's courses have repeatedly grown into papers at venues like ACL and EMNLP — most of the lab's publications are first-authored by students.

Start your research with us.

We're a small group of grad researchers and final-year project students working on agentic AI and LLM efficiency. If you're a student looking to start a research project — a thesis, an internship, your first paper — this is a good place to do it. Students here have published their first paper at NLP venues like ACL, EMNLP and NAACL, and most of our papers are first-authored by students. Prior ML experience helps, but curiosity matters more.