arXiv subjects

Piotr Miłoś

Publications and source records attributed to Piotr Miłoś.

At least 19 recordsLinked to original sources

arXiv2026

Up and Down the Abstraction Ladder: Code-Based Skills for Language Agents

Language agents struggle to act and learn in environments that require long sequences of low-level actions. Code-based abstractions can make these agents more productive by letting them invoke reusable skills instead of repeatedly selecting individual actions. The code handles recurring local decisions, while the language model decides which skills to use and how to combine them. Yet abstractions are leaky, and situations beyond a skill's capabilities may require a return to primitive actions. Motivated by this tradeoff between productivity and flexibility, we systematically study how code-based action abstraction affects the performance, inference cost, and learning of language agents. We study this in NetHack, a challenging, long-horizon game environment, using CodeHack, our library of code-based skills with natural-language descriptions. We use this library to compare agents restricted to primitives with those using semantic skills alone or in combination with primitives. We evaluate these agents in three settings: zero-shot prompting, supervised fine-tuning, and reinforcement learning. Across a broad zero-shot evaluation on NetHack, we find that compared with primitives, skills nearly triple game progression, while reducing inference cost per episode by 86%. Combining skills with primitives retains much of this benefit while preserving a path back down to low-level actions. Finally, in RL, we find that skill-based agents learn significantly faster than agents acting on primitives, achieving a 7.2x larger average gain in dungeon level over the same training budget. These results show that a supplied skill library can improve performance, efficiency, and learning, while retaining primitives provides flexibility when the library is insufficient. We release CodeHack together with training and evaluation code.

cs.AI↗

arXiv2026

Subgoal Search For Complex Reasoning Tasks

Humans excel in solving complex reasoning tasks through a mental process of moving from one idea to a related one. Inspired by this, we propose Subgoal Search (kSubS) method. Its key component is a learned subgoal generator that produces a diversity of subgoals that are both achievable and closer to the solution. Using subgoals reduces the search space and induces a high-level search graph suitable for efficient planning. In this paper, we implement kSubS using a transformer-based subgoal module coupled with the classical best-first search framework. We show that a simple approach of generating $k$-th step ahead subgoals is surprisingly efficient on three challenging domains: two popular puzzle games, Sokoban and the Rubik's Cube, and an inequality proving benchmark INT. kSubS achieves strong results including state-of-the-art on INT within a modest computational budget.

cs.AI↗

arXiv2026

What Drives Interactive Improvement from Feedback?

We study when natural-language feedback produces improvement beyond the gains obtainable from repeated attempts alone. In multi-turn language agent setting, higher final accuracy can reflect useful feedback, but it can also arise from resampling, format correction, or additional test-time computation. To separate these effects, we introduce a controlled student-teacher protocol across Omni-MATH, Codeforces, BBEH Linguini, and ARC-AGI1, evaluating thirteen open-weight models in both student and teacher roles. We compare external feedback, self-feedback, and unguided self-refinement, while varying interaction history, task difficulty, and teacher access to privileged task information. Across settings, we find that multi-turn improvement is often not evidence of feedback use: self-generated feedback adds little beyond unguided self-refinement, whereas the strongest external teachers produce substantially larger feedback-specific gains, suggesting that useful feedback must provide guidance beyond generic retry. Dense student-teacher interaction matrices further show that interactive gains are driven more by the student's ability to use feedback than by the teacher's identity, although teacher choice remains important for a fixed student. These results suggest that feedback-based agents should be evaluated against repeated-attempt baselines, and that ability to act on feedback, not merely feedback availability, is a central bottleneck for interactive improvement. We release our controlled student-teacher evaluation framework at https://j-lojek.github.io/feedback-generation-is-a-bottleneck/.

cs.AI↗

arXiv2026

When Does Non-Uniform Replay Matter in Reinforcement Learning?

Modern off-policy reinforcement learning algorithms often rely on simple uniform replay sampling and it remains unclear when and why non-uniform replay improves over this strong baseline. Across diverse RL settings, we show that the effectiveness of non-uniform replay is governed by three factors: replay volume, the number of replayed transitions per environment step; expected recency, how recent sampled transitions are; and the entropy of the replay sampling distribution. Our main contribution is clarifying when non-uniform replay is beneficial and providing practical guidance for replay design in modern off-policy RL. Namely, we find that non-uniform replay is most beneficial when replay volume is low, and that high-entropy sampling is important even at comparable expected recency. Motivated by these findings, we adopt a simple Truncated Geometric replay that biases sampling toward recent experience while preserving high entropy and incurring negligible computational overhead. Across large-scale parallel simulation, single-task, and multi-task settings, including three modern algorithms evaluated on five RL benchmark suites, this replay sampling strategy improves sample efficiency in low-volume regimes while remaining competitive when replay volume is high.

cs.LG↗

arXiv2026

Beyond Recognition: Evaluating Visual Perspective Taking in Vision Language Models

We investigate the ability of Vision Language Models (VLMs) to perform visual perspective taking using a new set of visual tasks inspired by established human tests. Our approach leverages carefully controlled scenes in which a single humanoid minifigure is paired with a single object. By systematically varying spatial configurations -- such as object position relative to the minifigure and the minifigure's orientation -- and using both bird's-eye and surface-level views, we created 144 unique visual tasks. Each task is paired with a series of 7 diagnostic questions designed to assess three levels of visual cognition: scene understanding, spatial reasoning, and visual perspective taking. We evaluate several high-performing models, including Gemini Robotics-ER 1.5, Llama-3.2-11B-Vision-Instruct, and variants of Claude Sonnet, GPT-4, and Qwen3, and find that while they excel at scene understanding, performance declines markedly on spatial reasoning and deteriorates further on perspective taking. Our analysis suggests a gap between surface-level object recognition and the deeper spatial and perspective reasoning required for complex visual tasks, pointing to the need for integrating explicit geometric representations and tailored training protocols in future VLM development.

cs.CV↗

arXiv2025

Amortized Causal Discovery with Prior-Fitted Networks

In recent years, differentiable penalized likelihood methods have gained popularity, optimizing the causal structure by maximizing its likelihood with respect to the data. However, recent research has shown that errors in likelihood estimation, even on relatively large sample sizes, disallow the discovery of proper structures. We propose a new approach to amortized causal discovery that addresses the limitations of likelihood estimator accuracy. Our method leverages Prior-Fitted Networks (PFNs) to amortize data-dependent likelihood estimation, yielding more reliable scores for structure learning. Experiments on synthetic, simulated, and real-world datasets show significant gains in structure recovery compared to standard baselines. Furthermore, we demonstrate directly that PFNs provide more accurate likelihood estimates than conventional neural network-based approaches.

cs.LG↗

arXiv2025

Lightweight Latent Verifiers for Efficient Meta-Generation Strategies

Verifiers are auxiliary models that assess the correctness of outputs generated by base large language models (LLMs). They play a crucial role in many strategies for solving reasoning-intensive problems with LLMs. Typically, verifiers are LLMs themselves, often as large (or larger) than the base model they support, making them computationally expensive. In this work, we introduce a novel lightweight verification approach, LiLaVe, which reliably extracts correctness signals from the hidden states of the base LLM. A key advantage of LiLaVe is its ability to operate with only a small fraction of the computational budget required by traditional LLM-based verifiers. To demonstrate its practicality, we couple LiLaVe with popular meta-generation strategies, like best-of-n or self-consistency. Moreover, we design novel LiLaVe-based approaches, like conditional self-correction or conditional majority voting, that significantly improve both accuracy and efficiency in generation tasks with smaller LLMs. Our work demonstrates the fruitfulness of extracting latent information from the hidden states of LLMs, and opens the door to scalable and resource-efficient solutions for reasoning-intensive applications.

cs.AI↗

arXiv2025

Structured Packing in LLM Training Improves Long Context Utilization

Recent advancements in long-context large language models have attracted significant attention, yet their practical applications often suffer from suboptimal context utilization. This study investigates structuring training data to enhance semantic interdependence, demonstrating that this approach effectively improves context utilization. To this end, we introduce the Structured Packing for Long Context (SPLiCe) method, which utilizes retrieval to collate mutually relevant documents into long and coherent training examples. We validate SPLiCe empirically across models of varying sizes -- 3B, 7B, and 13B -- achieving improved performance in long-context tasks, such as Qasper and HotpotQA. Remarkably, even brief fine-tuning with SPLiCe is sufficient to realize these benefits. Additionally, SPLiCe effectively mitigates the lost-in-middle phenomenon often observed in large models. Our comprehensive analysis of SPLiCe explores its design choices and reveals intriguing transfer effects; for instance, training on programming code enhances performance on natural language tasks.

cs.CL↗

arXiv2025

Since Faithfulness Fails: The Performance Limits of Neural Causal Discovery

Neural causal discovery methods have recently improved in terms of scalability and computational efficiency. However, our systematic evaluation highlights significant room for improvement in their accuracy when uncovering causal structures. We identify a fundamental limitation: neural networks cannot reliably distinguish between existing and non-existing causal relationships in the finite sample regime. Our experiments reveal that neural networks, as used in contemporary causal discovery approaches, lack the precision needed to recover ground-truth graphs, even for small graphs and relatively large sample sizes. Furthermore, we identify the faithfulness property as a critical bottleneck: (i) it is likely to be violated across any reasonable dataset size range, and (ii) its violation directly undermines the performance of neural discovery methods. These findings lead us to conclude that progress within the current paradigm is fundamentally constrained, necessitating a paradigm shift in this domain.

cs.LG↗

arXiv2025

Joint MoE Scaling Laws: Mixture of Experts Can Be Memory Efficient

Mixture of Experts (MoE) architectures have significantly increased computational efficiency in both research and real-world applications of large-scale machine learning models. However, their scalability and efficiency under memory constraints remain relatively underexplored. In this work, we present joint scaling laws for dense and MoE models, incorporating key factors such as the number of active parameters, dataset size, and the number of experts. Our findings provide a principled framework for selecting the optimal MoE configuration under fixed memory and compute budgets. Surprisingly, we show that MoE models can be more memory-efficient than dense models, contradicting conventional wisdom. To derive and validate the theoretical predictions of our scaling laws, we conduct over 280 experiments with up to 2.7B active parameters and up to 5B total parameters. These results offer actionable insights for designing and deploying MoE models in practical large-scale training scenarios.

cs.LG↗

arXiv2025

What Matters in Hierarchical Search for Combinatorial Reasoning Problems?

Efficiently tackling combinatorial reasoning problems, particularly the notorious NP-hard tasks, remains a significant challenge for AI research. Recent efforts have sought to enhance planning by incorporating hierarchical high-level search strategies, known as subgoal methods. While promising, their performance against traditional low-level planners is inconsistent, raising questions about their application contexts. In this study, we conduct an in-depth exploration of subgoal-planning methods for combinatorial reasoning. We identify the attributes pivotal for leveraging the advantages of high-level search: hard-to-learn value functions, complex action spaces, presence of dead ends in the environment, or using data collected from diverse experts. We propose a consistent evaluation methodology to achieve meaningful comparisons between methods and reevaluate the state-of-the-art algorithms.

cs.LG↗

arXiv2024

Bigger, Regularized, Optimistic: scaling for compute and sample-efficient continuous control

Sample efficiency in Reinforcement Learning (RL) has traditionally been driven by algorithmic enhancements. In this work, we demonstrate that scaling can also lead to substantial improvements. We conduct a thorough investigation into the interplay of scaling model capacity and domain-specific RL enhancements. These empirical findings inform the design choices underlying our proposed BRO (Bigger, Regularized, Optimistic) algorithm. The key innovation behind BRO is that strong regularization allows for effective scaling of the critic networks, which, paired with optimistic exploration, leads to superior performance. BRO achieves state-of-the-art results, significantly outperforming the leading model-based and model-free algorithms across 40 complex tasks from the DeepMind Control, MetaWorld, and MyoSuite benchmarks. BRO is the first model-free algorithm to achieve near-optimal policies in the notoriously challenging Dog and Humanoid tasks.

cs.LG↗

arXiv2024

Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe

Text embeddings are essential for many tasks, such as document retrieval, clustering, and semantic similarity assessment. In this paper, we study how to contrastively train text embedding models in a compute-optimal fashion, given a suite of pre-trained decoder-only language models. Our innovation is an algorithm that produces optimal configurations of model sizes, data quantities, and fine-tuning methods for text-embedding models at different computational budget levels. The resulting recipe, which we obtain through extensive experiments, can be used by practitioners to make informed design choices for their embedding models. Specifically, our findings suggest that full fine-tuning and low-rank adaptation fine-tuning produce optimal models at lower and higher computational budgets respectively.

cs.LG↗

arXiv2024

RapidDock: Unlocking Proteome-scale Molecular Docking

Accelerating molecular docking -- the process of predicting how molecules bind to protein targets -- could boost small-molecule drug discovery and revolutionize medicine. Unfortunately, current molecular docking tools are too slow to screen potential drugs against all relevant proteins, which often results in missed drug candidates or unexpected side effects occurring in clinical trials. To address this gap, we introduce RapidDock, an efficient transformer-based model for blind molecular docking. RapidDock achieves at least a $100 \times$ speed advantage over existing methods without compromising accuracy. On the Posebusters and DockGen benchmarks, our method achieves $52.1\%$ and $44.0\%$ success rates ($\text{RMSD}<2$Å), respectively. The average inference time is $0.04$ seconds on a single GPU, highlighting RapidDock's potential for large-scale docking studies. We examine the key features of RapidDock that enable leveraging the transformer architecture for molecular docking, including the use of relative distance embeddings of $3$D structures in attention matrices, pre-training on protein folding, and a custom loss function invariant to molecular symmetries.

q-bio.BM↗

arXiv2024

Fine-tuning Reinforcement Learning Models is Secretly a Forgetting Mitigation Problem

Fine-tuning is a widespread technique that allows practitioners to transfer pre-trained capabilities, as recently showcased by the successful applications of foundation models. However, fine-tuning reinforcement learning (RL) models remains a challenge. This work conceptualizes one specific cause of poor transfer, accentuated in the RL setting by the interplay between actions and observations: forgetting of pre-trained capabilities. Namely, a model deteriorates on the state subspace of the downstream task not visited in the initial phase of fine-tuning, on which the model behaved well due to pre-training. This way, we lose the anticipated transfer benefits. We identify conditions when this problem occurs, showing that it is common and, in many cases, catastrophic. Through a detailed empirical analysis of the challenging NetHack and Montezuma's Revenge environments, we show that standard knowledge retention techniques mitigate the problem and thus allow us to take full advantage of the pre-trained capabilities. In particular, in NetHack, we achieve a new state-of-the-art for neural models, improving the previous best score from $5$K to over $10$K points in the Human Monk scenario.

cs.LG↗

arXiv2024

Overestimation, Overfitting, and Plasticity in Actor-Critic: the Bitter Lesson of Reinforcement Learning

Recent advancements in off-policy Reinforcement Learning (RL) have significantly improved sample efficiency, primarily due to the incorporation of various forms of regularization that enable more gradient update steps than traditional agents. However, many of these techniques have been tested in limited settings, often on tasks from single simulation benchmarks and against well-known algorithms rather than a range of regularization approaches. This limits our understanding of the specific mechanisms driving RL improvements. To address this, we implemented over 60 different off-policy agents, each integrating established regularization techniques from recent state-of-the-art algorithms. We tested these agents across 14 diverse tasks from 2 simulation benchmarks, measuring training metrics related to overestimation, overfitting, and plasticity loss -- issues that motivate the examined regularization techniques. Our findings reveal that while the effectiveness of a specific regularization setup varies with the task, certain combinations consistently demonstrate robust and superior performance. Notably, a simple Soft Actor-Critic agent, appropriately regularized, reliably finds a better-performing policy within the training regime, which previously was achieved mainly through model-based approaches.

cs.LG↗

arXiv2024

Fast and Precise: Adjusting Planning Horizon with Adaptive Subgoal Search

Complex reasoning problems contain states that vary in the computational cost required to determine a good action plan. Taking advantage of this property, we propose Adaptive Subgoal Search (AdaSubS), a search method that adaptively adjusts the planning horizon. To this end, AdaSubS generates diverse sets of subgoals at different distances. A verification mechanism is employed to filter out unreachable subgoals swiftly, allowing to focus on feasible further subgoals. In this way, AdaSubS benefits from the efficiency of planning with longer subgoals and the fine control with the shorter ones, and thus scales well to difficult planning problems. We show that AdaSubS significantly surpasses hierarchical planning algorithms on three complex reasoning tasks: Sokoban, the Rubik's Cube, and inequality proving benchmark INT.

cs.AI↗

arXiv2024

Simulation-based reinforcement learning for real-world autonomous driving

We use reinforcement learning in simulation to obtain a driving system controlling a full-size real-world vehicle. The driving policy takes RGB images from a single camera and their semantic segmentation as input. We use mostly synthetic data, with labelled real-world data appearing only in the training of the segmentation network. Using reinforcement learning in simulation and synthetic data is motivated by lowering costs and engineering effort. In real-world experiments we confirm that we achieved successful sim-to-real policy transfer. Based on the extensive evaluation, we analyze how design decisions about perception, control, and training impact the real-world performance.

cs.LG↗