Commonsense reasoning is a key language model capability, as it is purportedly a prerequisite for many basic tasks, unlike specific factual knowledge. It is often measured with multiple-choice questions (MCQ) benchmarks, e.g. HellaSwag and PIQA. Some of these benchmarks, however, are outdated and contain numerous validity issues. We illustrate some typical validity issues with a case study on HellaSwag, one of the most popular and problematic benchmarks for commonsense reasoning. The issues we find range from basic ungrammaticality and numerous typos to misleading prompts or equally correct options. We show that if we remove question prompts or replace them with "Lorem ipsum dolor...", about 68% of model predictions remain the same. We argue that this occurs due to inner flaws in the benchmark, not mere contamination that might be present in some models. Since benchmark scores are an essential part of model selection in both research and commercial applications, these validity issues can have severe consequences. Based on our findings, we propose BenCheck, a package for benchmark validity analysis that encapsulates the main checks performed in our case study and can be used to audit MCQ benchmarks. We apply these checks to other commonsense reasoning benchmarks.

Tokenizer adaptation plays an important role in adapting pre-trained language models to new domains or languages. In this work, we address two complementary aspects of this process: vocabulary extension and pruning. The common approach to extension trains a new tokenizer on domain-specific text and appends the tokens that do not overlap with the existing vocabulary, which often results in many tokens that are unreachable or never used. We propose continued BPE training that extends a pre-trained tokenizer by continuing the BPE merge learning process on new data. Experiments across multiple languages and model families show that this approach improves tokenization efficiency and leads to better utilization of added vocabulary. We also introduce leaf-based vocabulary pruning, which removes redundant tokens while preserving model quality. Together, these methods provide practical tools for controlled vocabulary modification, which we release as an open-source toolkit.

Automated analysis of customer feedback on social media is hindered by three challenges: the high cost of annotated training data, the scarcity of evaluation sets, especially in multilingual settings, and privacy concerns that prevent data sharing and reproducibility. We address these issues by developing a generalizable synthetic data generation pipeline applied to a case study on customer distress detection in French public transportation. Our approach utilizes backtranslation with fine-tuned models to generate 1.7 million synthetic tweets from a small seed corpus, complemented by synthetic reasoning traces. We train 600M-parameter reasoners with English and French reasoning that achieve 77-79% accuracy on human-annotated evaluation data, matching or exceeding SOTA proprietary LLMs and specialized encoders. Beyond reducing annotation costs, our pipeline preserves privacy by eliminating the exposure of sensitive user data. Our methodology can be adopted for other use cases and languages.

LLMs often exhibit highly agreeable conversational styles, also known as AI sycophancy. This pattern may become problematic when interacting with user prompts that reflect negative social tendencies, risking the amplification of harmful behavior. We examine how LLMs respond to user prompts expressing varying degrees of Dark Triad traits (Machiavellianism, Narcissism, and Psychopathy) using a curated dataset. Our analysis reveals systematic differences across models: while all models predominantly exhibit corrective behavior, some generate reinforcing or ambivalent output. Model behavior further varies with severity level and response sentiment. These findings highlight the need for safer conversational systems that can reliably detect and respond to users escalating from benign to harmful requests.

RAG systems are usually compared as complete pipelines, making it difficult to attribute improvements to individual components. We evaluate chunking, retrieval, fusion, and reranking through a full factorial design with paired component substitutions and equal context budgets. Across five benchmarks, chunking interacts with retrieval architecture, BM25 fusion changes direction by domain, and reranking is the only intervention we test that remains beneficial across nearly all upstream configurations. Claim-level transitions further show that similar net recall can conceal very different changes in visible evidence. These findings demonstrate that RAG components cannot be ranked independently and motivate interaction-aware evaluation instead of winner-take-all pipeline comparisons.

Efficiency and safety of Large Language Models (LLMs), among other factors, rely on the quality of tokenization. A good tokenizer not only improves inference speed and language understanding but also provides extra defense against jailbreak attacks and lowers the risk of hallucinations. In this work, we investigate the efficiency of code tokenization, in particular from the perspective of data source diversity. We demonstrate that code tokenizers are prone to producing unused, and thus under-trained, tokens due to the imbalance in repository and language diversity in the training data, as well as the dominance of source-specific, repetitive tokens that are often unusable in future inference. By modifying the BPE objective and introducing merge skipping, we implement different techniques under the name Source-Attributed BPE (SA-BPE) to regularize BPE training and minimize overfitting, thereby substantially reducing the number of under-trained tokens while maintaining the same inference procedure as with regular BPE. This provides an effective tool suitable for production use.

Modern language models are trained on large amounts of data. These data inevitably include controversial and stereotypical content, which contains all sorts of biases related to gender, origin, age, etc. As a result, the models express biased points of view or produce different results based on the assigned personality or the personality of the user. In this paper, we investigate various proxy measures of bias in large language models (LLMs). We find that evaluating models with pre-prompted personae on a multi-subject benchmark (MMLU) leads to negligible and mostly random differences in scores. However, if we reformulate the task and ask a model to grade the user’s answer, this shows more significant signs of bias. Finally, if we ask the model for salary negotiation advice, we see pronounced bias in the answers. With the recent trend for LLM assistant memory and personalization, these problems open up from a different angle: modern LLM users do not need to pre-prompt the description of their persona since the model already knows their socio-demographics.

Language models can greatly benefit from efficient tokenization. However, they still mostly utilize the classical Byte-Pair Encoding (BPE) algorithm, a simple and reliable method. BPE has been shown to cause such issues as under-trained tokens and sub-optimal compression that may affect the downstream performance. We introduce PickyBPE, a modified BPE algorithm that carries out vocabulary refinement during tokenizer training by removing merges that leave intermediate “junk” tokens. Our method improves vocabulary efficiency, eliminates under-trained tokens, and does not compromise text compression. Our experiments show that this method either improves downstream performance or does not harm it.