Many recursively self-improving systems improve by editing their codebases based on observed rollout behavior. However, recent methods largely treat self-improvement as offline search: variants are optimized and evaluated on fixed benchmarks. This misaligns with real-world deployment, where tasks arrive sequentially and may have different distributions. We study recursive self-improvement as online learning over non-stationary task streams. In this setting, reliable execution and fallback are particularly difficult with limited evidence: different branches of self-improvement see different tasks, so raw rewards conflate capability with task difficulty, leading to overly conservative or optimistic reversion. We propose Online Anchor-calibrated Self-Improvement System (OASIS) that runs an adaptive reference agent on incoming tasks, converts raw rewards into paired comparisons, and uses confidence-calibrated rules for branch continuation and fallback without evaluating every agent on all tasks. Across SWE-Bench Pro and DeepResearch Bench, OASIS consistently improves over both the initial agent and prior RSI baseline SSA, raising the SWE-Bench Pro solve rate from 51.6% to 56.9% using Qwen3.6-35B-A3B and DeepResearch Bench from 35.18 to 39.31 using Gemini 3 Flash. This enables OASIS to outperform larger open weights and proprietary systems, including Qwen3.5-397B-A17B, Minimax 2.7, Qwen3.6-Plus, and GPT-5.3-Codex, with sustained deep self-improvements rather than shallow cycles.

Current multilingual benchmarks cover only a fraction of the languages represented in modern pretraining corpora and often conflate language proficiency with the ability to answer questions grounded in specific regions and communities. We introduce INCLUDE-2.0, a large-scale benchmark for region-aware multilingual factual evaluation, which covers 91 languages from 86 countries (127 total language–country pairs), comprising over 220K curated samples drawn from native regional sources. Unlike existing benchmarks that often treat language as a proxy for region, IINCLUDE-2.0 distinguishes between questions whose answers are independent of context, vary by country, concern widely documented facts about a place, or depend on knowledge primarily expressed through local linguistic and cultural sources. Each item is annotated with its regional knowledge type and domain, and paired with an English translation, enabling controlled analysis of the relationship between language proficiency and possession of regional knowledge.

Evaluating frontier and open-weight models on IINCLUDE-2.0, we find that aggregate multilingual accuracy masks substantial variation across regional knowledge types, languages, and regions. We find that frontier models perform worse on locally embedded knowledge than on widely documented facts about regions, with an average drop of ~14.5 points. We also show that for open-weight models, scale improves accuracy but does not reliably make performance invariant to whether a question is asked in English or its original language. INCLUDE-2.0 is available at https://huggingface.co/datasets/epfl-nlp/include-2.0

Large language models (LLMs) have shown remarkable performance on code generation tasks. A recent use case is iterative code repair, where an LLM fixes an incorrect program by rationalizing about errors and generating new code. Recent works augment the code repair process by integrating modern techniques such as chain-of-thought reasoning or distillation, but only study their benefits on high-resource languages like Python, and ignore low-resource languages like Perl. To address this gap of knowledge, we investigate the benefits of distilling code repair for both high and low resource languages to determine if the techniques that are effective in a high resource setting are also applicable in a low resource setting. Our evaluation shows that distilling the ability to repair code has language dependent benefits. To explain this behavior, we perform a further analysis and find that contrary to preexisting beliefs, the correlation between reasoning ability and code correction ability is weak. We hypothesize this weak correlation is magnified in low-resource settings where base models lack deep knowledge of a programming language, leading to wavering benefits of code repair.

Multi-agent debates have been introduced to improve the accuracy of Large Language Models (LLMs) by having multiple agents discuss solutions to a problem over several rounds of debate. However, models often generate incorrect yet confident-sounding responses, which can mislead the others. This issue arises partly because agents do not consider how confident their peers are. To address this, we propose DebUnc, a debate framework that uses uncertainty metrics to assess agent confidence. Confidence is then conveyed through textual prompts or via a modified attention mechanism that adjusts token weights. Evaluations across benchmarks show that attention-based methods are particularly effective and that performance continues to improve as uncertainty estimation becomes more reliable. The code is available at https://github.com/lukeyoffe/debunc.

Large Language Models (LLMs) have shown exceptional results on current benchmarks when working individually. The advancement in their capabilities, along with a reduction in parameter size and inference times, has facilitated the use of these models as agents, enabling interactions among multiple models to execute complex tasks. Such collaborations offer several advantages, including the use of specialized models (e.g. coding), improved confidence through multiple computations, and enhanced divergent thinking, leading to more diverse outputs. Thus, the collaborative use of language models is expected to grow significantly in the coming years. In this work, we evaluate the behavior of a network of models collaborating through debate under the influence of an adversary. We introduce pertinent metrics to assess the adversary’s effectiveness, focusing on system accuracy and model agreement. Our findings highlight the importance of a model’s persuasive ability in influencing others. Additionally, we explore inference-time methods to generate more compelling arguments and evaluate the potential of prompt-based mitigation as a defensive strategy.

This paper investigates the capabilities of Large Language Models (LLMs) in understanding their knowledge and uncertainty over questions. Specifically, we focus on addressing known-unknown questions, characterized by high uncertainty due to the absence of definitive answers. To facilitate our study, we collect a new dataset with Known-Unknown Questions (KUQ) and establish a categorization framework to clarify the origins of uncertainty in such queries. Subsequently, we examine the performance of open-source LLMs, fine-tuned using this dataset, in distinguishing between known and unknown queries within open-ended question-answering scenarios. The fine-tuned models demonstrated a significant improvement, achieving a considerable increase in F1-score relative to their pre-fine-tuning state. Through a comprehensive analysis, we reveal insights into the models’ improved uncertainty articulation and their consequent efficacy in multi-agent debates. These findings help us understand how LLMs can be trained to identify and express uncertainty, improving our knowledge of how they understand and express complex or unclear information.