Exploring Fine-Tuning for In-Context Retrieval and Efficient KV-Caching in Long-Context Language Models

With context windows of millions of tokens, Long-Context Language Models (LCLMs) can encode entire document collections, offering a strong alternative to conventional retrieval-augmented generation (RAG). However, it remains unclear whether fine-tuning strategies can improve long-context performance and translate to greater robustness under KV-cache compression techniques. In this work, we investigate which training strategies most effectively enhance LCLMs' ability to identify and use relevant information, as well as enhancing their robustness under KV-cache compression. Our experiments show substantial in-domain improvements, achieving gains of up to +20 points over the base model. However, out-of-domain generalization remains task dependent with large variance -- LCLMs excels on finance questions (+9 points), while RAG shows stronger performance on multiple-choice questions (+6 points) over the baseline models. Finally, we show that our fine-tuning approaches bring moderate improvements in robustness under KV-cache compression, with gains varying across tasks.

Right Answer, Wrong Score: Uncovering the Inconsistencies of LLM Evaluation in Multiple-Choice Question Answering

One of the most widely used tasks to evaluate Large Language Models (LLMs) is Multiple-Choice Question Answering (MCQA). While open-ended question answering tasks are more challenging to evaluate, MCQA tasks are, in principle, easier to assess, as the model's answer is thought to be simple to extract and is directly compared to a set of predefined choices. However, recent studies have started to question the reliability of MCQA evaluation, showing that multiple factors can significantly impact the reported performance of LLMs, especially when the model generates free-form text before selecting one of the answer choices. In this work, we shed light on the inconsistencies of MCQA evaluation strategies, which can lead to inaccurate and misleading model comparisons. We systematically analyze whether existing answer extraction methods are aligned with human judgment, and how they are influenced by answer constraints in the prompt across different domains. Our experiments demonstrate that traditional evaluation strategies often underestimate LLM capabilities, while LLM-based answer extractors are prone to systematic errors. Moreover, we reveal a fundamental trade-off between including format constraints in the prompt to simplify answer extraction and allowing models to generate free-form text to improve reasoning. Our findings call for standardized evaluation methodologies and highlight the need for more reliable and consistent MCQA evaluation practices.

ZEBRA: Zero-Shot Example-Based Retrieval Augmentation for Commonsense Question Answering

Current Large Language Models (LLMs) have shown strong reasoning capabilities in commonsense question answering benchmarks, but the process underlying their success remains largely opaque. As a consequence, recent approaches have equipped LLMs with mechanisms for knowledge retrieval, reasoning and introspection, not only to improve their capabilities but also to enhance the interpretability of their outputs. However, these methods require additional training, hand-crafted templates or human-written explanations. To address these issues, we introduce ZEBRA, a zero-shot question answering framework that combines retrieval, case-based reasoning and introspection and dispenses with the need for additional training of the LLM. Given an input question, ZEBRA retrieves relevant question-knowledge pairs from a knowledge base and generates new knowledge by reasoning over the relationships in these pairs. This generated knowledge is then used to answer the input question, improving the model's performance and interpretability. We evaluate our approach across 8 well-established commonsense reasoning benchmarks, demonstrating that ZEBRA consistently outperforms strong LLMs and previous knowledge integration approaches, achieving an average accuracy improvement of up to 4.5 points.