Research
LLMs meet symbolic AI
I'm a PhD student at Linköping University's MR Lab (since 2024), supervised by Jendrik Seipp and Marco Kuhlmann. My thesis, Robust Planning with Large Language Models, sits at the intersection of large language models and symbolic AI. LLMs give broad, easy-to-use intuition through a natural-language interface, but lack guarantees and interpretability; symbolic planning is the counterpoint: trustworthy and well-understood, yet domain-agnostic and expert-bound. I work in both directions: using LLMs to make planners more accessible and planners to make LLMs more reliable. Background: MSc in Applied Physics & Electrical Engineering (Data-Driven Intelligent Systems).
Papers
Symmetry-Aware Transformer Training for Automated PlanningProvably permutation-equivariant; a first attention-level loss · AAAI 2026 (CORE A*) · equal first author
Rather than pre-trained LLMs, we study the transformer architecture itself for planning, where state-representation and naming symmetries pose a major difficulty. We introduce a provably permutation-equivariant transformer and a first-of-its-kind attention-level loss for object renaming; both our heuristic and auto-regressive (policy) configurations beat the transformer-based state of the art.
Paper (PDF) →arXiv →
Planner Museum: Evaluating Classical Planners Over Time28 years of planning progress, measured · ICAPS 2026 (CORE A*)
The International Planning Competition has run since 1998 and has become a record of the field's paradigm shifts. The Planner Museum collects distinguished planners from past competitions and makes them run on modern computers. Evaluating these historical planners alongside modern ones on challenging benchmarks from multiple IPCs, we quantify how fast state-of-the-art planners have improved over 28 years and show that progress has been steady.
Paper (PDF) →
LLM-Evolved Domain-Independent Heuristics for Symbolic AI PlanningFirst evolved domain-independent heuristics · ICAPS 2026 (LM4Plan) · full version under review at ICLR 2027
Classical-planning heuristics are normally hand-designed and domain-specific. We instead evolve domain-independent heuristics automatically: an LLM-driven evolutionary loop (a three-island MAP-Elites population over an informedness–speed trade-off, seeded from FF or blind search and mutated by LLMs) discovers new heuristics that outperform a broad set of built-in baselines on the IPC learning-track domains. One striking finding is that blind search beats FF even among the evolved heuristics, suggesting heuristic effort is often wasted, and the whole evolutionary run stays cheap and fast.
Paper (PDF) →
Automatically Uncovering Intended Domain ConstraintsLearning a domain's unspoken rules, with guarantees · ICAPS 2026 (KEPS) · full version under review at AAAI 2027
Planning domains are usually incomplete: Blocksworld never states that towers can't be circular. We learn such intended constraints automatically via an abstract learning framework that subsumes existing invariant-learning methods (which it also speeds up) and is provably optimal. Candidates can be rules-based or LLM-generated; either way the result carries guarantees: no mistakes on the training data and PAC-optimal with respect to the best available choice.
Paper (PDF) →
Predicting Macro-Learning Performance Using Structural RegularityWith my master's student Anton Gustafsson · ICAPS 2026 (KEPS)
Macro-learning can speed planning up a lot or not at all, depending on the domain. We predict how well it will pay off from structural-regularity features of the domain itself, helping decide up front when macro-learning is worth applying.
Paper (PDF) →
Explainable Planning via Counterfactual Task Analysis (Beluga & Beyond)Won the Airbus × TUPLES Beluga Explainability Track · ICAPS 2025 (HAXP)
A counterfactual-reasoning framework, augmented with LLM explanations, for explaining planner decisions. We solve a task once, then re-solve modified versions driven by the user's question and have an LLM compare the two solutions into a human-readable analysis (for example, forcing an action to explain why it wasn't taken). This work won the Explainability Track of the Airbus & TUPLES Beluga Challenge.
Paper (PDF) →
NL2PlanFirst fully autonomous natural-language → PDDL system · ICAPS 2024 (HAXP)
The first fully autonomous system for generating complete PDDL planning tasks from minimal natural-language descriptions. A six-step pipeline iteratively extracts PDDL components such as types and actions and, unlike prior work, requires no PDDL information (e.g. predicates) while generating both the domain and problem files.
Paper (PDF) →arXiv →Code →
Talks & recognition
- Language Models Using Symbolic Tools: Push Performance, Increase Trust and Slash CostsInvited talk at the Software Technology Exchange Workshop (STEW 2026), Swedsoft, Stockholm · November 2026
- Symbolic Planning and Evolved HeuristicsInvited talk at the Natural Computing Research & Applications Group (NCRA), University College Dublin · June 2026
- Best ICAPS DC Presentation AwardFor LLMs and Planning at the ICAPS 2026 Doctoral Consortium
- Research visit to Canada, Q1 2027With Christian Muise (Queen's University) and Sheila McIlraith (University of Toronto), funded by a Blanceflor Scholarship (2026, sole recipient)
- Earlier recognitionIDA Best Thesis Award nominee (2025, one of 8); Qualcomm Stipend for Outstanding Engineering Students (2024, one of 9)
Service & teaching
- Co-organizer, LM4PlanThe workshop series on Planning in the Era of Large Language Models · ICAPS 2026 and ICML 2026
- ReviewerAAAI 2026–27, AAMAS 2027, ICAPS 2025–26, IJCAI 2025–26, KR 2026, NeurIPS 2025–26 · Shadow PC member, ICAPS 2026
- SupervisionMaster's student Anton Gustafsson (since 2025), which led to the KEPS paper above; master's thesis with Toyota Material Handling on LLM-based error handling for automated forklifts (2025)
- TeachingProject mentor in Foundations of AI and Machine Learning and teaching assistant in Logic (2026), where I also helped redesign the course; earlier Linear Algebra and Introduction to Mathematics
Advisory & building
Dependable AI for startups & scale-ups
Freelance AI consultant since 2021, working mostly with startups and scale-ups, building proofs-of-concept and MVPs and advising on AI strategy and adoption. Clients include Samla AI, Haaartland, Flow Innovate and Linköping Science Park. I also build fast under pressure, a repeat hackathon winner: Google DeepMind × Lovable, the Nordic Software Hackathon (judged by 1,200 C-level executives), the Airbus Beluga Challenge and SICK's Solution Hackathon.
Selected projects
Agents for Social InfrastructureSamla AI · since 2026
Advising a Swedish startup on how to design AI agents for social infrastructure.
AI-Driven Knowledge Sharing in AgricultureLinköping Science Park · 2025
A report, code and strategy for AI-driven knowledge sharing: a case study and implementation within the agricultural sector.
Desktop Robot for Elderly CareDyno Robotics (EU project) · 2025
Robot selection and development for a desktop robot that helps elderly people with their dietary and social needs.
GDPR-Compliant Customer TrackingFlow Innovate · 2024
Designed and implemented GDPR-compliant, camera-based tracking of customer movement and interaction for physical stores; served on the advisory board.
AI Strategy & RoadmapsHaaartland · 2022–2024
AI maintenance, future-proofing and roadmaps for a social-media company.
Earlier roles
Privacy of Models Trained on Sensitive DataResearcher, National Forensic Centre (Swedish Police) · 2023–2024
Research on the state of the art in reconstructing training data from trained neural networks, and what it means for the privacy of models trained on sensitive data.
Reinforcement-Learning FrameworkSystem developer, Drakryggen · 2022
Built a custom reinforcement-learning framework and card-game environment in PyTorch.
Building
Startups
Alongside research and consulting, I co-found companies that turn AI into tools people can actually use.
- Lingon AICo-founder, since 2026. A browser-based platform for building 3D prototypes of urban ideas with AI, without coding or 3D-modelling skills. Co-organising Innovathon 2026, a city-development innovation event in Linköping, together with Sigma, Linköpings Innersta, Linköping Municipality and Linköping Science Park. lingonai.com →
- StrimovaCo-founder, since 2025. A startup in process automation and documentation, affiliated with the LEAD incubator.
Talks & workshops
Making AI make sense
Grounded, engaging talks on AI, in English and Swedish, for audiences from high-schoolers and seniors to teachers, leaders and IT professionals.
- Language Models Using Symbolic ToolsInvited talk at the Software Technology Exchange Workshop (STEW 2026), Swedsoft, Stockholm · November 2026
- AI in Everyday Life: Examples and RisksA talk for seniors, held for SeniorNet Linköping · October 2026 · in Swedish (AI i vardagen: exempel och risker)
- Leadership LabsWorkshops on vibe coding for leaders, developed and delivered with LiU's AI Academy for industry and the public sector · since 2026
- AI for TeachersPractical AI for teachers, held as teacher training at Lagmanskolan, Linköping · March 2026 · in Swedish (AI för lärare)
- What Is AI, Really?A lecture for high-school students, held at schools around Linköping · since 2024 · in Swedish (Vad är AI egentligen?)
- Foundation Models Beyond LLMsA more technical talk on exactly that, for IT professionals (with AI East, Linköping Science Park & Goto10) · 2025
- Models and MyceliumAn AI-art exhibition at Naturcentrum, Linköping · 2025
- Data Leaks and Neural NetworksWhat trained models can reveal about their training data, for IT professionals · 2023
- On-demand workshopsTargeted sessions designed for your audience or organisation
Volunteering & service
Non-profit & community
Beyond research and consulting, I put real time into making AI and tech education accessible and into running the communities around them.
- Treasurer & board member, DataföreningenSweden's oldest computer association, focused on mass education (incl. the AI-Körkortet course). Co-opted board 2025; treasurer since 2026, elected to a two-year term.
- Organiser, RoboCup Junior European Championship 2028Organising the European youth robotics championship. Since 2026.
- National organiser, RoboCup Junior SwedenPrincipal organiser of Sweden's youth robotics championship (2022, 2024, 2025, 2026) and logistics (2021, 2023). Now mentoring the current national organiser.
- Founder, PhD Shadowing programmeLets master's students at LiU's computer science department (IDA) shadow a PhD student for a normal working day, to see what doing a PhD actually involves. Since 2026.
- Organiser, master's thesis inclusivity programmeAt LiU's computer science department (IDA). Since 2025.
- Chair & vice-chair, FIA, LiU's AI & Robotics AssociationLed the association's board, workshops and competitions, 2020–2024.
- Skolverket grant, sole PI (≈287k SEK)2025. International travel and accommodation for the youth robotics championship.