arXiv subjects

Igor Kiselev

Publications and source records attributed to Igor Kiselev.

8 recordsLinked to original sources

arXiv2026

Weight-Space Geometry of Offline Reasoning Training

Offline reinforcement-learning losses (RFT, RIFT, DFT, Offline GRPO, DPO) are widely used to distill reasoning from large teachers into smaller students, and are typically compared on downstream accuracy alone. We ask whether they are mechanistically distinct or converge to a similar weight update. Training six methods (SFT, RFT, DFT, RIFT, Offline GRPO, DPO) on identical math rollouts from a single base model (Qwen3-4B) with attention-only LoRA, we analyze the resulting deltas via cosine similarity, principal-angle subspace analysis, linear mode connectivity, and CKA. We observe: (i) SFT, RFT, and RIFT have nearly colinear weight deltas (cosine >= 0.97, top-1 principal angle ~7 deg median over 144 modules) and comparable GSM8K accuracy (87-88%, n=1319; pairwise McNemar p >= 0.15); (ii) DFT diverges further in direction than any reward-weighted method despite using the same data; (iii) Offline GRPO adds a substantial component orthogonal to the SFT direction (~67% globally, up to ~86% in late layers) while staying in the SFT loss basin; (iv) DPO sits in a near-orthogonal subspace, shows a mode-connectivity barrier, and collapses late-layer CKA to ~0.46. DPO also reaches the highest accuracy in our protocol on both GSM8K (93.5%, McNemar p < 10^-9 vs. each other method) and AIME26 (30.0% vs. 3.3-10.0%); its training uses a 10x smaller learning rate than the others (the standard convention), so the update-norm and accuracy gaps reflect loss-function and optimizer choices jointly, and a learning-rate-matched DPO comparison is left for future work.

cs.LG↗

arXiv2026

ThinkBooster: A Unified Framework for Seamless Test-Time Scaling of LLM Reasoning

Test-time compute (TTC) scaling has emerged as a powerful paradigm for improving large language model (LLM) reasoning by allocating additional compute during inference, e.g., via multi-sample generation and verifier-based reranking. Existing TTC scaling strategies and reasoning scorers remain fragmented, evaluated under inconsistent protocols, and are rarely analyzed through the lens of quality-cost trade-offs. We introduce ThinkBooster, a unified framework for seamless test-time compute scaling of LLM reasoning, which consists of (i) a modular Python library implementing state-of-the-art TTC scaling strategy and scorer families, (ii) a benchmark that jointly evaluates performance and computational efficiency, and (iii) a deployable OpenAI-compatible proxy service that enables drop-in integration of adaptive reasoning into real-world applications. We further provide a demo visual debugger for inspecting the reasoning trajectories, intermediate selection decisions, and alternative reasoning paths. Empirical results on mathematical and coding tasks reveal the performance-compute trade-offs of TTC scaling strategies and scoring methods and demonstrate that ThinkBooster provides practical gains in real-world tasks. The code is available online under an MIT license.

cs.CL↗

arXiv2025

A Head to Predict and a Head to Question: Pre-trained Uncertainty Quantification Heads for Hallucination Detection in LLM Outputs

Large Language Models (LLMs) have the tendency to hallucinate, i.e., to sporadically generate false or fabricated information. This presents a major challenge, as hallucinations often appear highly convincing and users generally lack the tools to detect them. Uncertainty quantification (UQ) provides a framework for assessing the reliability of model outputs, aiding in the identification of potential hallucinations. In this work, we introduce pre-trained UQ heads: supervised auxiliary modules for LLMs that substantially enhance their ability to capture uncertainty compared to unsupervised UQ methods. Their strong performance stems from the powerful Transformer architecture in their design and informative features derived from LLM attention maps. Experimental evaluation shows that these heads are highly robust and achieve state-of-the-art performance in claim-level hallucination detection across both in-domain and out-of-domain prompts. Moreover, these modules demonstrate strong generalization to languages they were not explicitly trained on. We pre-train a collection of UQ heads for popular LLM series, including Mistral, Llama, and Gemma 2. We publicly release both the code and the pre-trained heads.

cs.CL↗

arXiv2025

Yes, Q-learning Helps Offline In-Context RL

Existing offline in-context reinforcement learning (ICRL) methods have predominantly relied on supervised training objectives, which are known to have limitations in offline RL settings. In this study, we explore the integration of RL objectives within an offline ICRL framework. Through experiments on more than 150 GridWorld and MuJoCo environment-derived datasets, we demonstrate that optimizing RL objectives directly improves performance by approximately 30% on average compared to widely adopted Algorithm Distillation (AD), across various dataset coverages, structures, expertise levels, and environmental complexities. Furthermore, in the challenging XLand-MiniGrid environment, RL objectives doubled the performance of AD. Our results also reveal that the addition of conservatism during value learning brings additional improvements in almost all settings tested. Our findings emphasize the importance of aligning ICRL learning objectives with the RL reward-maximization goal, and demonstrate that offline RL is a promising direction for advancing ICRL.

cs.LG↗

arXiv2025

Object-Centric Latent Action Learning

Leveraging vast amounts of unlabeled internet video data for embodied AI is currently bottlenecked by the lack of action labels and the presence of action-correlated visual distractors. Although recent latent action policy optimization (LAPO) has shown promise in inferring proxy action labels from visual observations, its performance degrades significantly when distractors are present. To address this limitation, we propose a novel object-centric latent action learning framework that centers on objects rather than pixels. We leverage self-supervised object-centric pretraining to disentangle the movement of the agent and distracting background dynamics. This allows LAPO to focus on task-relevant interactions, resulting in more robust proxy-action labels, enabling better imitation learning and efficient adaptation of the agent with just a few action-labeled trajectories. We evaluated our method in eight visually complex tasks across the Distracting Control Suite (DCS) and Distracting MetaWorld (DMW). Our results show that object-centric pretraining mitigates the negative effects of distractors by 50%, as measured by downstream task performance: average return (DCS) and success rate (DMW).

cs.CV↗

arXiv2025

Latent Action Learning Requires Supervision in the Presence of Distractors

Recently, latent action learning, pioneered by Latent Action Policies (LAPO), have shown remarkable pre-training efficiency on observation-only data, offering potential for leveraging vast amounts of video available on the web for embodied AI. However, prior work has focused on distractor-free data, where changes between observations are primarily explained by ground-truth actions. Unfortunately, real-world videos contain action-correlated distractors that may hinder latent action learning. Using Distracting Control Suite (DCS) we empirically investigate the effect of distractors on latent action learning and demonstrate that LAPO struggle in such scenario. We propose LAOM, a simple LAPO modification that improves the quality of latent actions by 8x, as measured by linear probing. Importantly, we show that providing supervision with ground-truth actions, as few as 2.5% of the full dataset, during latent action learning improves downstream performance by 4.2x on average. Our findings suggest that integrating supervision during Latent Action Models (LAM) training is critical in the presence of distractors, challenging the conventional pipeline of first learning LAM and only then decoding from latent to ground-truth actions.

cs.CV↗

arXiv2024

N-Gram Induction Heads for In-Context RL: Improving Stability and Reducing Data Needs

In-context learning allows models like transformers to adapt to new tasks from a few examples without updating their weights, a desirable trait for reinforcement learning (RL). However, existing in-context RL methods, such as Algorithm Distillation (AD), demand large, carefully curated datasets and can be unstable and costly to train due to the transient nature of in-context learning abilities. In this work, we integrated the n-gram induction heads into transformers for in-context RL. By incorporating these n-gram attention patterns, we considerably reduced the amount of data required for generalization and eased the training process by making models less sensitive to hyperparameters. Our approach matches, and in some cases surpasses, the performance of AD in both grid-world and pixel-based environments, suggesting that n-gram induction heads could improve the efficiency of in-context RL.

cs.LG↗

arXiv2024

Vikhr: The Family of Open-Source Instruction-Tuned Large Language Models for Russian

There has been a surge in the development of various Large Language Models (LLMs). However, text generation for languages other than English often faces significant challenges, including poor generation quality and reduced computational performance due to the disproportionate representation of tokens in the model's vocabulary. In this work, we address these issues by developing a pipeline for the adaptation of English-oriented pre-trained models to other languages and constructing efficient bilingual LLMs. Using this pipeline, we construct Vikhr, a series of bilingual open-source instruction-following LLMs designed specifically for the Russian language. ``Vikhr'' refers to the name of the Mistral LLM series and means a ``strong gust of wind.'' Unlike previous Russian-language models that typically rely on LoRA adapters on top of English-oriented models, sacrificing performance for lower training costs, Vikhr features an adapted tokenizer vocabulary and undergoes the continued pre-training and instruction tuning of all weights. This not only enhances the model's performance but also significantly improves its computational and contextual efficiency. We also expanded the instruction datasets and corpora for continued pre-training. The model weights, instruction sets, and code are publicly available.

cs.CL↗