Sara Candussio

Reading Between the Tokens: Uncovering the Semantic Minima of AI Monologues

Fri, 24 Apr 2026 12:05:00 +0000

Invited talk at the CLCG Linguistics Lunch at the University of Groningen.

Chain-of-Thought prompting asks models to reason step by step — but most of what they write is filler. This talk presents work on identifying the semantic minima of AI reasoning: the tiny subset of tokens that actually carry predictive weight, detectable in real time from the model’s internal states. Erasing up to 95% of the output leaves a sparse set of words that still perfectly predicts the correct answer.

Distilling Formal Logic into Neural Spaces

Thu, 05 Mar 2026 00:00:00 +0000

Continuous neural representations of STL specifications via kernel distillation. Accepted at NeSy 2026.

Distilling Formal Logic into Neural Spaces: A Kernel Alignment Approach for Signal Temporal Logic

Thu, 05 Mar 2026 00:00:00 +0000

Accepted at NeSy 2026 as an oral presentation 🎉

The Idea

Typical knowledge distillation compresses a big neural model into a smaller one. We do something different: our “expert” isn’t a neural network at all — it’s a mathematical kernel built on top of formal logic.

This kernel is provably correct: it captures the true meaning of logical formulas with mathematical guarantees. The problem? It’s expensive to compute and doesn’t scale.

So instead of distilling a big neural model into a smaller neural model, we distill the geometric structure (i.e. relative positions) of a symbolic kernel into a Transformer encoder. We’re not compressing parameters — we’re transferring mathematical meaning into neural space.

Using a teacher-student setup with a kernel-weighted geometric alignment objective, we train the encoder to mirror the semantic distances defined by the symbolic kernel. Errors are penalized proportionally to their semantic discrepancy — not just their magnitude.

Why It Matters

The result is a model that:

runs in a single forward pass
produces semantically faithful embeddings of logical formulas
can reconstruct the original formula from its embedding
all without sacrificing the logical guarantees of the original kernel

This opens up scalable, trustworthy neuro-symbolic reasoning for domains where correctness and efficiency aren’t optional:

🚗 Autonomous driving & robotics (real-time monitoring of safety specs)
🏥 Healthcare (checking patient signals against clinical guidelines)
⚙️ Cyber-physical systems (fast verification & control synthesis)

And beyond STL, this approach generalizes to any domain with a meaningful but expensive similarity function — think genomic sequences, molecular graphs, or structured data where semantics matter more than surface form.

Co-authored with Gabriele Sarti, Gaia Saveri, and Luca Bortolussi.

If you want to explore how to transfer a slow, expensive similarity function into Transformers — let’s connect!

A Dialectic Pipeline for Improving LLM Robustness

Wed, 28 Jan 2026 00:00:00 +0000

Can LLMs improve their accuracy without further training, just through a dialectic way of questioning themselves — as Hegel suggested?

This was the core question behind my Master’s thesis. The short answer: yes, and by a lot.

The Idea

Inspired by Hegelian dialectics, the pipeline structures reasoning into three stages:

The thesis–antithesis–synthesis pipeline.

Thesis — the model produces an initial answer given the question, context, and options.
Antithesis — the model challenges its own answer, now also seeing the thesis.
Synthesis — the model produces a final answer, having seen both thesis and antithesis.

No fine-tuning. No domain-specific verifiers. Just structured self-dialogue.

Results

The pipeline was tested on multi-hop QA benchmarks (HotpotQA, WikiHop) across five open-source models under 20B parameters (Phi-mini, Phi-medium, Gemma-2B, Gemma-9B, LLaMA-8B).

Accuracy improvements across models on HotpotQA.

From 53.4% to 80.7% on HotpotQA with Phi-mini (+27.3%).

Improvements of up to 30% on complex multi-hop questions — beating standard Chain-of-Thought prompting.

CoT vs. pipeline on WikiHop across all models.

Key Takeaways

Self-debating is the main driver: letting models reflect on and contrast their own reasoning significantly boosts performance, especially as question complexity increases.
Instruction following matters: models that strictly follow instructions (Llama, Phi) benefit more than those that get “too creative” (Gemma-2).
Smart filtering > summarization: when dealing with long contexts, filtering for relevant information beats summarization, which can hurt deductive reasoning.
Avoid overthinking: for simpler tasks, too much deliberation can introduce errors. A touch of “impulsivity” sometimes helps.

Original vs. summarized vs. filtered context on WikiHop.

This work also received an Honorable Mention at the Emanuele Pianta Award (AILC) for the best Italian NLP Master’s thesis at CLiC-it 2025. 🏆

CLiC-it 2025, Cagliari.

If you’re interested in agentic reasoning, small language models, or multi-hop QA — feel free to reach out!

Large Language Models: Potenzialità, Limiti e Sistemi Multi-Agent

Mon, 15 Dec 2025 00:00:00 +0000

Workshop at Novalia, Trieste — December 2025.

Part of a two-session seminar series on digital transformation, co-organized with IP4FVG and the University of Trieste.

This session covered the inner workings of Large Language Models, multi-agent architectures, and fine-tuning strategies — with an eye toward practical business applications and a frank discussion of current limitations.

The core message: the challenge isn’t just adopting AI, but integrating it strategically to augment rather than replace human potential.

Bridging Logic and Learning: Decoding Temporal Logic Embeddings via Transformers

Mon, 24 Nov 2025 00:00:00 +0000

Poster presentation of our ECML-PKDD 2025 paper at the RTG Symposium, held at the legendary Schloss Dagstuhl — a week-long gathering of PhD researchers from Max Planck Institute and Saarland University, with guest talks by Mor Geva Pipek and Luca Bortolussi.

Bridging Logic and Learning: Decoding Temporal Logic Embeddings via Transformers

Thu, 10 Jul 2025 00:00:00 +0000

This work introduces a Transformer-based decoder that inverts embeddings of Signal Temporal Logic (STL) formulae. By constructing a small STL vocabulary, the model can generate valid formulae quickly, generalize across semantic structures, and simplify formulas while preserving their meaning. Our methodology is evaluated across varying formula complexity and applied to requirement mining tasks, performing optimization directly in the semantic space.

Create your slides in Markdown - click the Slides button to check out the example.

Add the publication’s full text or supplementary notes here. You can use rich formatting such as including code, math, and images.

If you find overlap with your work or interests, I would be glad to connect and explore possible collaborations.

STLDec: Decoding Temporal Logic Embeddings via Transformers

Thu, 10 Jul 2025 00:00:00 +0000

Transformer-based decoder for inverting semantic embeddings of Signal Temporal Logic (STL) formulae. Published at ECML-PKDD 2025.

Probabilistic Machine Learning

Sat, 01 Mar 2025 00:00:00 +0000

Teaching assistant for the Probabilistic Machine Learning MSc course at the University of Trieste (Spring 2025).

Topics covered: ERM, PAC learnability, Probabilistic Graphical Models, Hidden Markov Models, Bayesian Classification and Regression, Sampling‑based Inference, Expectation‑Maximization, Variational Inference, Generative Modeling (VAEs, Diffusion Models), Gaussian Processes.

OverRef: Studying Over-Refusal in Large Language Models

Wed, 01 Jan 2025 00:00:00 +0000

Ongoing project on over-refusal in LLMs: studying when and why models refuse legitimate user queries, with benchmarking and dataset resources.

Projects

Sun, 19 May 2024 00:00:00 +0000

Experience

Tue, 24 Oct 2023 00:00:00 +0000