Behavioral evaluation of LLMs
Controlled benchmarks that hold the evidence fixed and vary one causal factor at a time, run across open-weight and frontier models, to measure when context rather than evidence decides the answer.
Hallo there! I'm Yiderigun Borjigin.
I am a doctoral researcher in AI safety at Saarland University, supervised by Prof. Dr. Roland Aydin, working on behavioral evaluation of LLMs, chain-of-thought faithfulness, and mechanistic interpretability.
My research asks when a model's output is driven by evidence and when it is driven by something else in its context — a number planted in a retrieved document, a confidently phrased wrong reasoning trace, an artifact of how data was serialized into text. I build controlled benchmarks that isolate one causal factor, run them across open-weight and frontier models, and trace the behavior back to internal state via residual-stream patching. The finding that recurs is that high task accuracy does not imply robustness.
Publications and preprints
† Published under the name Yiderigun Yiderigun; same author.
-
AnchorBench: A Multi-Pathway Benchmark for the Anchoring Effect in LLMs
Yiderigun Borjigin, Alexander Hermann, Christian J. Cyron, Roland Aydin
Conference on Language Modeling (COLM), main conference · 2026
A benchmark isolating how a numeric anchor's delivery pathway — prompt, conversation history, few-shot demonstrations, RAG documents, tool outputs — and its framing govern LLM judgment. 14 models × 9,000 condition-controlled prompts, with a Bayesian rational-updating ceiling separating justified evidence integration from bias.
-
From RawTokens to PhysSummary: Probing Text Interfaces for Inverse 1D PDE Parameter Estimation
Yiderigun Borjigin†, Arman Shojaei, Christian J. Cyron, Roland Aydin
ICLR 2026 Workshop on AI and Partial Differential Equations · Poster · 2026
Shows the text interface between numerical data and an LLM is reliability-critical, with a failure-mode taxonomy — prompt-visible constant copying, fine-tuning collapse on compressed features, non-monotonic shot scaling — invisible to accuracy metrics alone.
-
Synthetic Persona Pretraining: Alignment from Token Zero
Julian Minder, Viktor Moskvoretskii, Raghav Singhal, Difan Jiao, Andy Arditi, Shaobo Cui, Yiderigun Borjigin, Kartik Bali, Stefan Krsteski, Harsh Raj, Huu Nguyen, Jannik Brinkmann, Ashton Anderson, Roland Aydin, Robert West
Preprint, arXiv:2608.13482 [cs.LG] · 2026
A fifteen-author multi-institution collaboration (EPFL, Toronto, Saarland/Hereon, and others) on pretraining-time alignment interventions.
Research and artifacts
Chain-of-thought faithfulness and interpretability
Residual-stream activation patching, donor-state and layer-sweep interventions, and logit-lens attribution to locate where a reasoning trace becomes causally load-bearing.
AI for science: text interfaces for PDE modeling
Reliability metrics for the interface between numerical data and language models, including inverse parameter estimation under zero-shot, in-context, and LoRA/QLoRA regimes.
Open-source artifacts
AnchorBench — evaluation harness and public dataset
202614,400 released prompts with gold answers and anchor values; seed-controlled deterministic generation, one shared parser across all conditions (median parse rate 99.9%), and SHA-256 checksums linking every published number to the raw generations behind it.
Curriculum vitae
Download PDFResearch experience
Jun. 2026 – Present
Doctoral Researcher, AI Safety and Scientific ML
Saarland University · supervised by Prof. Dr. Roland Aydin
Lead author on LLM safety research spanning behavioral evaluation, chain-of-thought faithfulness, and mechanistic analysis of how context steers model answers.
Jun. 2025 – Jun. 2026
Research Associate
Helmholtz-Zentrum Hereon, Institute of Material Systems Modeling, Geesthacht
Built AnchorBench, and ran a mechanistic study of chain-of-thought faithfulness using activation patching and logit-lens attribution.
Jul. 2024 – Apr. 2025
Master Thesis Researcher
BMW Group, Research and Innovation Center (FIZ), Munich
Deep learning models for multivariate time-series prediction of vehicle thermal behavior, with transfer learning and domain adaptation.
Oct. 2024 – Apr. 2025
Teaching Assistant, Fundamentals of Artificial Intelligence
Technical University of Munich · built exercises on constraint satisfaction
Education
Jun. 2026 – Present
PhD in Artificial Intelligence (in progress)
Saarland University, Saarbrücken
Apr. 2022 – Apr. 2025
M.Sc. Robotics, Cognition, Intelligence
Technical University of Munich · ML, deep learning, computer vision
Sept. 2017 – Jun. 2021
B.Eng. Automotive Engineering
Jilin University, Changchun, China
Research methods and infrastructure
Evaluation infrastructure
Benchmark and harness design in Python, built end to end; vLLM on multi-GPU H100 nodes, HuggingFace Transformers, OpenRouter and provider APIs (OpenAI, Anthropic, Google, xAI); batched greedy and sampled decoding, deterministic answer extraction, parse-free logit readouts, seed-controlled data generation, checksummed result releases.
Interpretability
Residual-stream activation patching, donor-state and layer-sweep interventions, logit-lens attribution, next-token logit margins, teacher-forced span scoring.
Model adaptation
PyTorch, LoRA / QLoRA fine-tuning (PEFT), supervised fine-tuning pipelines, transfer learning, domain adaptation.
Experimental statistics
Cluster bootstrap CIs, Wilcoxon signed-rank tests, Benjamini–Hochberg and Holm correction, TOST equivalence testing, pre-declared thresholds and validity audits, ablation and sensitivity design.
Benchmarks and models evaluated
MMLU-Redux, MMLU-Pro, GPQA-Diamond, GSM8K, MATH-500, AIME, CruxEval; Llama 3.1–3.3, Qwen2.5 / Qwen3, Gemma 2–3, OLMo 2, DeepSeek-R1-Distill, gpt-oss, GPT-5.x, Claude, Gemini, Grok.