CV

PhD

University of Tübingen & Max Planck Research School

Computer Science

• Health-NLP Lab, Supervisor: Prof. Dr. Carsten Eickhoff.
• Scholar, International Max Planck Research School for Intelligent Systems (IMPRS-IS).
• Research focus: Mechanistic interpretability of language models for NLP and IR tasks.
• Expected graduation: Aug 2028.
• Available for an internship in 2027 (flexible dates and duration).

MSc

Moscow Institute of Physics and Technology

Applied Mathematics and Computer Science

• Research topics: sparse retrieval, multimodal retrieval, tool calling in LLMs.
• Thesis: “Approaches to Training Language Models for External Function Calling: Using Self-Generated Data for Better Error Detection and Recovery”.

BSc

Moscow Institute of Physics and Technology

Applied Mathematics and Computer Science

• Thesis: “Usage of MCMC Methods in Deep Learning”.

Understanding Wacky Weights: A Dissection of SPLADE's Learned Term Importance

G. Polyakov, H. Scells, C. Eickhoff

SIGIR 2026 (Reproducibility)

Introduced the Wackiness Score and W-AUC to quantify semantically unrelated expansion terms across models with different vocabularies and sparsity levels. Trained 6 SPLADE variants, showing that the prevalence of wacky tokens increases with vocabulary size and that they mainly aid in-domain retrieval.

Interpretability Analysis of Arithmetic In-Context Learning in Large Language Models

G. Polyakov, C. Hepting, C. Eickhoff, A. Bahrainian

EMNLP 2025 (Main)

Traced how 4 LLMs (up to 12B) use in-context examples for multi-step arithmetic, showing that format consistency matters more than arithmetic correctness. Distilled in-context examples into a function vector that raises zero-shot accuracy by up to 25 points and fully recovers one-shot accuracy on Pythia-12B.

Towards Best Practices of Axiomatic Activation Patching in Information Retrieval

G. Polyakov, C. Chen, C. Eickhoff

SIGIR 2025 (Short)

Proposed a noise-mitigating evaluation metric for activation patching in neural rankers that filters out spurious low-confidence signals. Systematically analyzed diagnostic datasets, identifying key selection criteria for queries and documents to ensure robust interpretability analysis.

ToolReflection: Improving Large Language Models for Real-World API Calls with Self-Generated Data

G. Polyakov, I. Alimova, D. Abulkhanov, I. Sedykh, A. Bout, S. Nikolenko, I. Piontkovskaya

ACL 2025 (REALM Workshop)

Proposed a fine-tuning strategy that enables LLMs to self-correct API calls by learning from runtime execution feedback. Co-developed evaluation datasets for tool-augmented LLMs in realistic scenarios.

Sinkhorn Transformations for Single-Query Postprocessing in Text-Video Retrieval

K. Yakovlev, G. Polyakov, I. Alimova, A. V. Podolskiy, A. Bout, S. I. Nikolenko, I. Piontkovskaya

SIGIR 2023 (Short)

NLP Research Engineer (full-time)

Huawei Noah's Ark Lab

• Developed a fine-tuning method that teaches LLMs to detect and recover from failed API calls (ACL 2025 REALM Workshop).
• Designed and co-implemented the end-to-end NLU pipeline for a production in-car voice assistant in Russian.
• Co-developed a Sinkhorn-based post-processing method that improves text-video retrieval quality (SIGIR 2023).

NLP Research Intern (full-time)

Huawei Noah's Ark Lab

• Co-developed the reranking model for a large-scale Russian-language product search engine: improved ranking quality by 5% relative NDCG@10 and cut inference latency 1.5x by adapting ColBERTer and TinyBERT-style distillation.

Understanding Wacky Weights: A Dissection of SPLADE's Learned Term Importance

Glasgow IR Seminar, University of Glasgow

Presidential Grant Scholarship

Russian Federation

Feb 2018 – Aug 2021. Four-year stipend for academic and competition merit, one of 1,200 recipients nationwide.

Teaching Assistant: Introduction to AI

University of Tübingen

Winter Semester 2026/27. Conducted hands-on tutorials and evaluated student assignments.

Teaching Assistant: Modern Search Engines

University of Tübingen

Summer Semesters 2025 and 2026. Conducted hands-on tutorials, evaluated assignments, and mentored student projects.

Teaching Assistant: Introduction to Python Programming

University of Tübingen

Winter Semesters 2024/25 and 2025/26. Conducted hands-on tutorials and evaluated student assignments.