Igor Slinko

Igor Slinko, Ilia Zavidnyi, Egor Bogomolov et al.

Rejection Fine-Tuning (RFT) is a standard method for training LLM agents, where unsuccessful trajectories are discarded from the training set. In the context of SWE-bench tasks, this corresponds to filtering out runs where the submitted patch does not pass the tests. However, this approach discards unresolved trajectories, even though they form a large portion of all trajectories for hard tasks and even then may be partially correct. In this work, we propose Step Rejection Fine-Tuning (SRFT) - a practical way to leverage these unresolved trajectories. For this, we employ a critic LLM to assess the correctness of each step in a trajectory. Consequently, during training, we mask the loss for erroneous steps while retaining them in the context window. This way we ensure the model learns to recover from errors without reproducing them. Evaluation on SWE-bench Verified shows that while RFT improves the resolution rate by 2.4% by excluding unresolved trajectories, SRFT improves it by 3.7% by filtering them instead of discarding completely, reaching the total resolution rate of 32.2%.

Kirill Gelvan, Igor Slinko, Felix Steinbauer et al.

LLM-based Software Engineering agents face a critical bottleneck: context length limitations cause failures on complex, long-horizon tasks. One promising solution is to encode context as continuous embeddings rather than discrete tokens, enabling denser information storage. We apply the recently proposed In-Context Autoencoder for this purpose. While the method performs well on single-shot common-knowledge and code-understanding tasks, our experiments demonstrate that it fails on multi-step agentic coding tasks. In this paper, we explore this phenomenon and discuss possible factors contributing to this failure.

Tobias Lindenbauer, Igor Slinko, Ludwig Felder et al.

Large Language Model (LLM)-based agents solve complex tasks through iterative reasoning, exploration, and tool-use, a process that can result in long, expensive context histories. While state-of-the-art Software Engineering (SE) agents like OpenHands or Cursor use LLM-based summarization to tackle this issue, it is unclear whether the increased complexity offers tangible performance benefits compared to simply omitting older observations. We present a systematic comparison of these approaches within SWE-agent on SWE-bench Verified across five diverse model configurations. Moreover, we show initial evidence of our findings generalizing to the OpenHands agent scaffold. We find that a simple environment observation masking strategy halves cost relative to the raw agent while matching, and sometimes slightly exceeding, the solve rate of LLM summarization. Additionally, we introduce a novel hybrid approach that further reduces costs by 7% and 11% compared to just observation masking or LLM summarization, respectively. Our findings raise concerns regarding the trend towards pure LLM summarization and provide initial evidence of untapped cost reductions by pushing the efficiency-effectiveness frontier. We release code and data for reproducibility.

Igor Slinko, Anna Vorontsova, Dmitry Zhukov et al.

Learning-based visual odometry and SLAM methods demonstrate a steady improvement over past years. However, collecting ground truth poses to train these methods is difficult and expensive. This could be resolved by training in an unsupervised mode, but there is still a large gap between performance of unsupervised and supervised methods. In this work, we focus on generating synthetic data for deep learning-based visual odometry and SLAM methods that take optical flow as an input. We produce training data in a form of optical flow that corresponds to arbitrary camera movement between a real frame and a virtual frame. For synthesizing data we use depth maps either produced by a depth sensor or estimated from stereo pair. We train visual odometry model on synthetic data and do not use ground truth poses hence this model can be considered unsupervised. Also it can be classified as monocular as we do not use depth maps on inference. We also propose a simple way to convert any visual odometry model into a SLAM method based on frame matching and graph optimization. We demonstrate that both the synthetically-trained visual odometry model and the proposed SLAM method build upon this model yields state-of-the-art results among unsupervised methods on KITTI dataset and shows promising results on a challenging EuRoC dataset.

Pavel Kirsanov, Airat Gaskarov, Filipp Konokhov et al.

We present a novel dataset for training and benchmarking semantic SLAM methods. The dataset consists of 200 long sequences, each one containing 3000-5000 data frames. We generate the sequences using realistic home layouts. For that we sample trajectories that simulate motions of a simple home robot, and then render the frames along the trajectories. Each data frame contains a) RGB images generated using physically-based rendering, b) simulated depth measurements, c) simulated IMU readings and d) ground truth occupancy grid of a house. Our dataset serves a wider range of purposes compared to existing datasets and is the first large-scale benchmark focused on the mapping component of SLAM. The dataset is split into train/validation/test parts sampled from different sets of virtual houses. We present benchmarking results forboth classical geometry-based and recent learning-based SLAM algorithms, a baseline mapping method, semantic segmentation and panoptic segmentation.

Igor Slinko, Anna Vorontsova, Filipp Konokhov et al.

Optical Flow (OF) and depth are commonly used for visual odometry since they provide sufficient information about camera ego-motion in a rigid scene. We reformulate the problem of ego-motion estimation as a problem of motion estimation of a 3D-scene with respect to a static camera. The entire scene motion can be represented as a combination of motions of its visible points. Using OF and depth we estimate a motion of each point in terms of 6DoF and represent results in the form of motion maps, each one addressing single degree of freedom. In this work we provide motion maps as inputs to a deep neural network that predicts 6DoF of scene motion. Through our evaluation on outdoor and indoor datasets we show that utilizing motion maps leads to accuracy improvement in comparison with naive stacking of depth and OF. Another contribution of our work is a novel network architecture that efficiently exploits motion maps and outperforms learnable RGB/RGB-D baselines.