This research addresses the performance gap in large language models between single-turn and multi-turn interactions. The authors introduce TURNWISEEVAL, a new benchmark that isolates conversational ability by comparing model responses in long dialogues against equivalent single-turn prompts. To improve model performance, they also developed TURNWISEDATA, a scalable pipeline that generates synthetic multi-turn training data from existing single-turn instructions. Their experiments demonstrate that even advanced models often struggle with extended context, but incorporating a small amount of this synthetic data during training significantly boosts chat capabilities. Ultimately, the study highlights that multi-turn proficiency is a distinct skill set that requires dedicated evaluation and specialized training data.

This research paper introduces **temporal straightening**, a technique designed to improve **latent planning** in AI world models by regularizing the curvature of agent trajectories. While standard visual encoders often produce highly curved paths in latent space, this approach uses a **curvature regularizer** to create a representation where feasible transitions follow straighter lines. This geometric transformation ensures that **Euclidean distance** serves as a more accurate proxy for the actual distance to a goal, significantly improving the stability of **gradient-based optimization**. Theoretical analysis demonstrates that straightening the latent space leads to a better-conditioned **planning objective**, allowing planners to converge more efficiently. Empirical tests across several goal-reaching tasks, such as **PointMaze** and **PushT**, show that this method substantially increases success rates for both open-loop and closed-loop planning. Ultimately, the work suggests that the **geometric structure** of learned representations is a critical factor in the effectiveness of autonomous planning systems.

This research examines the tension between in-context learning (ICL) and fine-tuning in Transformer-based models, specifically using linear attention to provide a theoretical foundation. While fine-tuning is often employed to enhance zero-shot performance on specific target tasks, the authors demonstrate that updating all attention parameters can inadvertently damage the model's ability to learn from demonstrations. They identify a superior strategy: restricting updates to the value matrix, which improves task-specific accuracy while maintaining the model’s original few-shot capabilities. The study further explores the use of an auxiliary few-shot loss, finding that it boosts performance on the target task but reduces the model's ability to generalize to out-of-distribution tasks. These theoretical insights are validated through both mathematical proofs and empirical experiments on the MMLU benchmark. Ultimately, the work provides a framework for optimizing language models without sacrificing their inherent flexibility as in-context learners.

Researchers have introduced OAPL, a new reinforcement learning algorithm designed to improve how Large Language Models (LLMs) learn complex reasoning for math and coding. Traditional methods often struggle when the training policy and the inference engine are out of sync, a common issue in large-scale, asynchronous computing. Instead of trying to force these mismatched systems to align, OAPL embraces this discrepancy by using a squared regression objective that functions effectively even with significant policy lag. This approach eliminates the need for complex importance sampling or heuristics that can destabilize training. Empirical results show that OAPL outperforms existing methods like GRPO on competitive benchmarks while using significantly fewer computational resources. Furthermore, the model maintains higher sequence entropy, which prevents the performance collapse often seen in other post-training techniques.

This paper explores Continual Reinforcement Learning (CRL) for large Vision-Language-Action (VLA) models, focusing on how these agents adapt to new tasks without losing prior knowledge. While traditional machine learning often suffers from catastrophic forgetting during sequential training, this research demonstrates that a simple Sequential Fine-Tuning approach remains remarkably effective. By combining pre-trained VLAs, on-policy reinforcement learning, and Low-Rank Adaptation (LoRA), the researchers found that models maintain high plasticity and strong zero-shot generalization. Their systematic study across multiple benchmarks reveals that this basic recipe often outperforms more complex, specialized CRL strategies. Ultimately, the source positions parameter-efficient fine-tuning as a scalable and stable foundation for developing lifelong embodied intelligence in robotic agents.

This research paper introduces In-Context Policy Optimization (ICPO), a framework designed to explain and enhance the self-reflection capabilities of large language models. The authors provide a mathematical foundation proving that specific transformer architectures can inherently mimic policy optimization algorithms without requiring parameter updates. Building on this theory, they develop ME-ICPO, a practical algorithm that improves mathematical reasoning by iteratively refining responses based on self-assessed rewards. To ensure reliability, the system utilizes minimum-entropy selection and majority voting to filter out noise from self-evaluations. Empirical results demonstrate that this approach significantly boosts performance on complex reasoning benchmarks while remaining computationally efficient. Ultimately, the work bridges the gap between the theoretical understanding of in-context learning and the empirical success of test-time scaling.

This research paper introduces Energy-Based Fine-Tuning (EBFT), a novel method for refining language models by matching feature statistics of generated text with ground-truth data. Traditional training relies on next-token prediction, which often causes models to drift or fail during long sequences because they lack global distributional calibration. By optimizing a feature-matching objective using a frozen feature network, EBFT provides dense semantic feedback at the sequence level without requiring a manual reward function. Experimental results across coding and translation show that EBFT outperforms standard supervised fine-tuning and equals reinforcement learning in accuracy. Furthermore, it achieves superior validation cross-entropy, proving that aligning feature moments effectively stabilizes model behavior and preserves high-quality language modeling.

This research introduces the concept of Neural Thickets, describing a phenomenon where large pretrained models are surrounded by a high density of diverse, task-specific solutions in their local weight space. While small models require structured optimization like gradient descent to find improvements, larger models transition into a regime where random weight perturbations frequently yield "expert" versions of the model. The authors exploit this discovery through RandOpt, a parallel post-training method that samples random weight changes, selects the best performers, and ensembles their predictions. Their findings show that these random experts are specialists rather than generalists, often excelling at one task while declining in others, which makes ensembling via majority vote highly effective. This approach proves competitive with standard reinforcement learning methods like PPO and GRPO, especially as model scale increases. Ultimately, the study suggests that sufficient pretraining fundamentally reshapes the loss landscape, making complex downstream adaptation possible through simple parallel search and selection.

This paper introduces AdaEvolve, a novel framework designed to enhance how Large Language Models (LLMs) solve complex optimization and programming tasks through evolutionary search. Unlike existing methods that use rigid, pre-set schedules, this system implements hierarchical adaptivity to manage computational resources and search strategies dynamically. It operates across three levels: local adaptation to adjust exploration intensity, global adaptation to allocate the budget toward promising solution populations, and meta-guidance to generate new tactics when progress stalls. This approach mimics the efficiency of adaptive gradient methods used in continuous optimization but applies it to discrete, zero-th order problems. Experimental results across 185 benchmarks show that AdaEvolve consistently outperforms standard baselines and human-designed solutions in areas like combinatorial geometry and systems optimization. By replacing brittle manual tuning with a unified improvement signal, the framework demonstrates a more robust and autonomous path for AI-driven discovery.

This paper introduces ∇-Reasoner, a novel framework that improves Large Language Model (LLM) reasoning by applying gradient-based optimization during the inference process. Unlike traditional methods that rely on random sampling or discrete searches, this approach uses Differentiable Textual Optimization (DTO) to refine token logits through first-order gradients derived from reward models and likelihood signals. By iteratively updating textual representations in latent space, the system allows for bidirectional information flow, enabling the model to correct its reasoning chains on the fly. To ensure efficiency, the framework incorporates gradient caching and rejection sampling, which reduce the computational burden typically associated with backpropagation. Empirical results demonstrate that $\nabla$-Reasoner significantly boosts accuracy on complex mathematical benchmarks while requiring fewer model calls than existing search-based baselines. Ultimately, the research establishes a theoretical and practical shift toward treating test-time reasoning as a continuous optimization problem rather than a simple stochastic generation task.