To Örebro University

AI research agents are demonstrating great potential to accelerate scientific progress by automating the design, implementation, and training of machine learning models. We focus on methods for improving agents' performance on MLE-bench, a challenging benchmark where agents compete in Kaggle competitions to solve real-world machine learning problems. We formalize AI research agents as search policies that navigate a space of candidate solutions, iteratively modifying them using operators. By designing and systematically varying different operator sets and search policies (Greedy, MCTS, Evolutionary), we show that their interplay is critical for achieving high performance. Our best pairing of search strategy and operator set achieves a state-of-the-art result on MLE-bench lite, increasing the success rate of achieving a Kaggle medal from 39.6% to 47.7%. Our investigation underscores the importance of jointly considering the search strategy, operator design, and evaluation methodology in advancing automated machine learning.

Large Language Models (LLMs) have been touted as AI models possessing advanced reasoning abilities. However, recent works have shown that LLMs often bypass true reasoning using shortcuts, sparking skepticism. To study the reasoning capabilities in a principled fashion, we adopt a computational theory perspective and propose an experimental protocol centered on 3-SAT – the prototypical NP-complete problem lying at the core of logical reasoning and constraint satisfaction tasks. Specifically, we examine the phase transitions in random 3-SAT and characterize the reasoning abilities of LLMs by varying the inherent hardness of the problem instances. Our experimental evidence shows that LLMs are incapable of performing true reasoning, as required for solving 3-SAT problems. Moreover, we observe significant performance variation based on the inherent hardness of the problems – performing poorly on harder instances and vice versa. Importantly ,we show that integrating external reasoners can considerably enhance LLM performance. By following a principled experimental protocol, our study draws concrete conclusions and moves beyond the anecdotal evidence often found in LLM reasoning research.

We propose a self-play approach to generating strategies for playing in multi-player games, where strategies are represented as computer code. We use large language models (LLMs) to generate pieces of code to play in the game, which we refer to as generated bots. We engage the LLM generated bots in competitions, designed to generate increasingly stronger strategies. We follow game theoretic principles in organizing these tournaments, and use a Policy Space Response Oracle (PSRO) approach. We start with an initial set of LLM generated bots, and continue in rounds for adding new bots into the population. Each round adds a bot to the population by asking the LLM to produce code for playing against a bot representing the Nash equilibrium mixture over the current population. Our analysis shows that even a few rounds are sufficient to produces strong bots for playing the game. Our demo shows the process for the game of Checkers. We allow users to select initial bots in the population, run the process, inspect how the bots evolve over time, and play against the generated bots.

To achieve natural and intuitive interaction with people, HRI frameworks combine a wide array of methods for human perception, intention communication, human-aware navigation and collaborative action. In practice, when encountering unpredictable behavior of people or unexpected states of the environment, these frameworks may lack the ability to dynamically recognize such states, adapt and recover to resume the interaction. Large Language Models (LLMs), owing to their advanced reasoning capabilities and context retention, present a promising solution for enhancing robot adaptability. This potential, however, may not directly translate to improved interaction metrics. This paper considers a representative interaction with an industrial robot involving approach, instruction, and object manipulation, implemented in two conditions: (1) fully scripted and (2) including LLM-enhanced responses. We use gaze tracking and questionnaires to measure the participants' task efficiency, engagement, and robot perception. The results indicate higher SUbjective ratings for the LLM condition, but objective metrics show that the scripted condition performs comparably, particularly in efficiency and focus during simple tasks. We also note that the scripted condition may have an edge over LLM-enhanced responses in terms of response latency and energy consumption, especially for trivial and repetitive interactions.

Large Language Models (LLMs) have been touted as AI models possessing advanced reasoning abilities. In theory, autoregressive LLMs with Chain-of-Thought (CoT) can perform more serial computations to solve complex reasoning tasks. However, recent studies suggest that, despite this capacity, LLMs do not truly learn to reason but instead fit on statistical features. To study the reasoning capabilities in a principled fashion, we adopt a computational theory perspective and propose an experimental protocol centered on 3-SAT -- the prototypical NP-complete problem lying at the core of logical reasoning and constraint satisfaction tasks. Specifically, we examine the phase transitions in random 3-SAT and characterize the reasoning abilities of state-of-the-art LLMs by varying the inherent hardness of the problem instances. By comparing DeepSeek R1 with other LLMs, our findings reveal two key insights (1) LLM accuracy drops significantly on harder instances, suggesting all current models struggle when statistical shortcuts are unavailable (2) Unlike other LLMs, R1 shows signs of having learned the underlying reasoning. Following a principled experimental protocol, our study moves beyond the benchmark-driven evidence often found in LLM reasoning research. Our findings highlight important gaps and suggest clear directions for future research. 

Owing to their reasoning capabilities, large language models (LLMs) have been evaluated on planning tasks described in natural language. However, LLMs have largely been tested on planning domains without constraints. In order to deploy them in real-world settings where adherence to constraints, in particular safety constraints, is critical, we need to evaluate their performance on constrained planning tasks. We introduce LexiCon -- a natural language-based (Lexi) constrained (Con) planning benchmark, consisting of a suite of environments, that can be used to evaluate the planning capabilities of LLMs in a principled fashion. The core idea behind LexiCon is to take existing planning environments and impose temporal constraints on the states. These constrained problems are then translated into natural language and given to an LLM to solve. A key feature of LexiCon is its extensibility. That is, the set of supported environments can be extended with new (unconstrained) environment generators, for which temporal constraints are constructed automatically. This renders LexiCon future-proof: the hardness of the generated planning problems can be increased as the planning capabilities of LLMs improve. Our experiments reveal that the performance of state-of-the-art LLMs, including reasoning models like GPT-5, o3, and R1, deteriorates as the degree of constrainedness of the planning tasks increases. 

Reasoning and decision-making are foundational challenges in artificial intelligence (AI). These processes are closely linked – an intelligent agent must reason about its environment and goals in order to make decisions and select actions. Two principal frameworks for sequential decision-making are AI planning and reinforcement learning (RL). Planning assumes access to a known model of the environment and uses symbolic representations to compute a sequence of actions that leads from an initial state to a desired goal. In contrast, RL focuse son learning behavior through interaction, enabling agents to develop policies that maximize long-term reward under uncertainty. Despite methodological differences, both approaches aim to generate intelligent, goal-directed action sequences.

The rise of Large Language Models (LLMs) has sparked significant interest in their potential to perform reasoning, planning, and decision-making tasks. Despite their impressive performance in natural language understanding and generalization, there is growing skepticism about whether LLMs genuinely reason or merely leverage statistical correlations. This dissertation investigates this question through a principled evaluation grounded in computational theory, using 3-SAT – the canonical NP-complete problem – as a testbed. The findings demonstrate that LLMs fail to exhibit sound and complete reasoning, especially on complex instances where shallow heuristics fail, and that their apparent reasoning abilities often stem from overfitting to statistical patterns.

To address these limitations, this dissertation proposes a range of neurosymbolic architectures that combine the generative flexibility of LLMs with the rigor and reliability of symbolic methods. Empirical evaluations across planning, reward design, and plan verification tasks show that such integration yields systems that are more robust and accurate. This work advances our theoretical and practical understanding of LLM-based reasoning, provides concrete design principles for neurosymbolic systems, and charts a path toward AI agents that integrate world knowledge with logical precision.

Designing effective reward functions is crucial to training reinforcement learning (RL) algorithms. However, this design is non-trivial, even for domain experts, due to the subjective nature of certain tasks that are hard to quantify explicitly. In recent works, large language models (LLMs) have been used for reward generation from natural language task descriptions, leveraging their extensive instruction tuning and commonsense understanding of human behavior. In this work, we hypothesize that LLMs, guided by human feedback, can be used to formulate reward functions that reflect human implicit knowledge. We study this in three challenging settings - autonomous driving, humanoid locomotion, and dexterous manipulation - wherein notions of “good” behavior are tacit and hard to quantify. To this end, we introduce REvolve, a truly evolutionary framework that uses LLMs for reward design in RL. REvolve generates and refines reward functions by utilizing human feedback to guide the evolution process, effectively translating implicit human knowledge into explicit reward functions for training (deep) RL agents. Experimentally, we demonstrate that agents trained on REvolve-designed rewards outperform other state-of-the-art baselines. 

Large Language Models (LLMs) have demonstrated impressive planning abilities due to their vast "world knowledge". Yet, obtaining plans that are both feasible (grounded in affordances) and cost-effective (in plan length), remains a challenge, despite recent progress. This contrasts with heuristic planning methods that employ domain knowledge (formalized in action models such as PDDL) and heuristic search to generate feasible, optimal plans. Inspired by this, we propose to combine the power of LLMs and heuristic planning by leveraging the world knowledge of LLMs and the principles of heuristic search. Our approach, SayCanPay, employs LLMs to generate actions (Say) guided by learnable domain knowledge, that evaluates actions' feasibility (Can) and long-term reward/payoff (Pay), and heuristic search to select the best sequence of actions. Our contributions are (1) a novel framing of the LLM planning problem in the context of heuristic planning, (2) integrating grounding and cost-effective elements into the generated plans, and (3) using heuristic search over actions. Our extensive evaluations show that our model surpasses other LLM planning approaches.

Despite its successes, Deep Reinforcement Learning (DRL) yields non-interpretable policies. Moreover, since DRL does not exploit symbolic relational representations, it has difficulties in coping with structural changes in its environment (such as increasing the number of objects).  Meanwhile, Relational Reinforcement Learning inherits the relational representations from symbolic planning to learn reusable policies. However, it has so far been unable to scale up and exploit the power of deep neural networks. We propose Deep Explainable Relational Reinforcement Learning (DERRL), a framework that exploits the best of both -- neural and symbolic worlds. By resorting to a neuro-symbolic approach, DERRL combines relational representations and constraints from symbolic planning with deep learning to extract interpretable policies. These policies are in the form of logical rules that explain why each decision (or action) is arrived at. Through several experiments, in setups like the Countdown Game, Blocks World, Gridworld, Traffic, and Mingrid, we show that the policies learned by DERRL are adaptable to varying configurations and environmental changes.