Research overview

Research Highlights

Five peer-reviewed studies and my MA thesis, each shown with its headline result. Every card links to the publication; the papers also have interactive explainers.

A 20-step stress test for LLM consistency

DEFER DEFER DEFER CLAIM →→→…→ STEP 1 STEP 20 Can the agent maintain the same commitment over repeated decisions? MANIPULATED FACTORS SOCIAL CONTEXT PERSONA DELIBERATION POLICY WHEN does it fail? Survival / hazard analysis WHY does it fail? Failure-rationale taxonomy
ICLR 2026 LLM Reasoning Workshop

LLM failures develop distinct patterns over time.

Early failures are impulse-driven; later failures shift toward cost–benefit reasoning and fatigue.

84,540agent trajectories
8model families
13,780labeled failure rationales
Explore failure patterns → Read the paper ↗

Mean learning gain

26.1% 10.6% with prompts without prompts
AIED 2025 Main track

Three simple engagement prompts more than doubled mean learning gains.

Students learned more when prompted to show their work, avoid copying, and reflect on errors. ChatGPT vs. traditional materials was not a significant factor.

101participants
4experimental conditions
p = .040prompts effect
Explore the experiment → Read the paper ↗

Relative odds of giving in at a given minute

Conditional on the agent still waiting · log scale

0.1×1×10×100× child vs adult senior vs adult MUST vs MAY isolated vs broadcast 66.3× 7.55× 1.42× 0.78×
NeurIPS 2025 MTI-LLM Workshop

Persona prompts strongly changed failure risk⁠—and mandatory self-questioning increased it.

In Gemini 2.5 Flash-Lite, child and craving personas produced the highest risk of giving in, while requiring self-questioning increased rather than reduced failure odds.

19,200agent trajectories
64experimental cells
1base LLM · Gemini 2.5 Flash-⁠Lite
Explore the benchmark → Read the paper ↗

How the bandit reduces grading cost

Multi-step+ examples Multi-step Single-step+ examples Single-step BANDIT CONTROLLER SELECT ONE STRATEGY PER ESSAY learns from each score Exhaustive search Bandit selection vs. 7,870 1,697 LLM calls LLM calls 78.4% fewer calls Comparable scoring agreement with human graders (QWK)
EDM 2025 WLIEDM Workshop

A bandit learns which grading strategy to favor while it grades⁠—using 78.4% fewer LLM calls than exhaustive search.

−78.4%LLM calls
−72.8%tokens
4grading strategies
Run the simulation → Read the paper ↗

Where AI fits in the assignment

1 2 3 4 DESIGN BUILD CORE LOGIC TEST & REFINE Student-led AI-assisted Student-led AI-assisted AI: explain concepts· review design AI: examples ·boilerplate · APIs AI: syntax help only AI: code review ·test cases →→→ Students document how they used, modified, and deliberately avoided AI assistance.
IEEE COMPSAC 2025 QUORS Workshop

A framework for integrating AI into software engineering curricula, grounded in learning theory and expert-reviewed before student testing.

2learning theories
11guideline categories reviewed
1Operating Systems assignment
Explore the framework → Read the paper ↗

How the LLM runs the instructional cycle

GENERATE PRACTICE STUDENT ANSWERS GRADE RESPONSE UPDATE MASTERY STATE DECIDE WHAT HAPPENS NEXT 2 consecutive correct first attempts = mastered NEW ISOMORPHIC PROBLEM NEXT TEMPLATE FORMATIVE ASSESSMENT COMPLETE REMEDIATION RE-PRACTICE RETEST mastered not mastered all templates mastered pass fail
MA thesis 2026 Carleton University

A prompt-configured LLM coordinated a complete mastery-learning cycle with high, but imperfect, reliability.

It generated practice, graded responses, tracked mastery, decided progression, and triggered remediation across 45 sessions with 25 novice programmers.

94.5%correct mastery decisions
96%structurally valid generated problems
97.5%correct grading decisions
Read the thesis ↗

Get in touch

Open to roles, projects, and research collaborations.

AI enablement, AI education, and LLM evaluation & reliability.

Contact me olga@manakina.com