Rubric-based RL is a promising route for extending reinforcement learning beyond verifiable rewards, yet existing methods optimize rubrics while treating the query distribution as fixed. We identify a structural bottleneck: rubric quality is constrained by query structure. Open-ended queries yield vague rubrics; naively narrowing them introduces fabricated references that no model can verify, so all responses fail and training receives no reward signal. We present QUBRIC, a framework that co-designs queries and rubrics. Teacher-derived key points ground the rewriting of open-ended prompts into scenario-based, evaluable questions. Contrastive rubric generation then turns teacher–policy gaps into query-level criteria, and learnability filtering retains only informative query–rubric pairs for GRPO training. QUBRIC achieves a +5.5 point gain on ArenaHard over the SFT baseline. Trained only on instruction-following data, it further transfers to three held-out benchmarks spanning legal, moral, and narrative reasoning (+6.3 points on average), with improvements concentrated in reasoning-related dimensions. These results provide evidence that co-designing queries and rubrics can make rubric-based RL a practical complement to RLVR beyond strictly verifiable tasks.

Conversational shopping assistants now serve hundreds of millions of customers, yet no existing benchmark jointly evaluates the open-ended multi-turn reasoning, domain expertise, and criterion-level quality that real shopping conversations demand. Shopping reasoning is unique among language model applications. Unlike factual question answering or verifiable code generation, it requires balancing subjective preferences, budget constraints, and cross-product trade-offs across multi-turn dialogue, capabilities absent from previous e-commerce and general-purpose benchmarks. We introduce the Shopping Reasoning Bench, an expert-authored benchmark of 525 missions (232 single-turn, 293 multi-turn) with 10863 importance-weighted binary rubrics authored by retail domain experts. These criteria are organized under a taxonomy of five reasoning categories and fifteen subcategories covering diverse demands such as preference refinement, trade-off analysis, and compatibility assessment. An evaluation of nine models across three families (GPT, Claude, Gemini) shows that pass rates reach only 57–77% overall. On multi-turn missions, all models score 13–29 points lower on optional above-and-beyond criteria than on required ones, and performance degrades 4–18 points as conversations progress. These gaps show that current models handle basic shopping assistance but fall short of expert-level advice, making Shopping Reasoning Bench a challenging testbed for future shopping assistant development.

Due to the scarcity of agent-oriented pre-training data, LLM-based autonomous agents typically rely on complex prompting or extensive fine-tuning, which often fails to introduce new capabilities while preserving strong generalizability. We introduce Hephaestus-Forge, the first large-scale pre-training corpus designed to enhance the fundamental capabilities of LLM agents in API function calling, intrinsic reasoning and planning, and adapting to environmental feedback. Hephaestus-Forge comprises 103B agent-specific data encompassing 76,537 APIs, including both tool documentation to introduce knowledge of API functions and function calling trajectories to strengthen intrinsic reasoning. To explore effective training protocols, we investigate scaling laws to identify the optimal recipe in data mixing ratios. By continual pre-training on Hephaestus-Forge, Hephaestus outperforms small- to medium-scale open-source LLMs and rivals commercial LLMs on three agent benchmarks, demonstrating the effectiveness of our pre-training corpus in enhancing fundamental agentic capabilities and generalization of LLMs to new tasks or environments.

Learning from preference feedback is essential for aligning large language models (LLMs) with human values and improving the quality of generated responses. However, existing preference learning methods rely heavily on curated data from humans or advanced LLMs, which is costly and difficult to scale. In this work, we present PUGC, a novel framework that leverages implicit human Preferences in unlabeled User-Generated Content (UGC) to generate preference data. Although UGC is not explicitly created to guide LLMs in generating human-preferred responses, it often reflects valuable insights and implicit preferences from its creators that has the potential to address readers’ questions. PUGC transforms UGC into user queries and generates responses from the policy model. The UGC is then leveraged as a reference text for response scoring, aligning the model with these implicit preferences. This approach improves the quality of preference data while enabling scalable, domain-specific alignment. Experimental results on Alpaca Eval 2 show that models trained with DPO and PUGC achieve a 9.37% performance improvement over traditional methods, setting a 35.93% state-of-the-art length-controlled win rate using Mistral-7B-Instruct. Further studies highlight gains in reward quality, domain-specific alignment effectiveness, robustness against UGC quality, and theory of mind capabilities. Our code and dataset are available at https://zhaoxuan.info/PUGC.github.io/.

The ability of large language models (LLMs) to execute complex instructions is essential for their real-world applications. However, several recent studies indicate that LLMs struggle with challenging instructions. In this paper, we propose Evolutionary Contrastive Distillation (ECD), a novel method for generating high-quality synthetic preference data designed to enhance the complex instruction-following capability of language models. ECD generates data that specifically illustrates the difference between a response that successfully follows a set of complex instructions and a response that is high-quality, but nevertheless makes some subtle mistakes. This is done by prompting LLMs to progressively evolve simple instructions to more complex instructions. When the complexity of an instruction is increased, the original successful response to the original instruction becomes a “hard negative” response for the new instruction, mostly meeting requirements of the new instruction, but barely missing one or two. By pairing a good response with such a hard negative response, and employing contrastive learning algorithms such as DPO, we improve language models’ ability to follow complex instructions. Empirically, we observe that our method yields a 7B model that exceeds the complex instruction-following performance of current SOTA 7B models and is competitive even with open-source 70B models.

The adoption of large language models (LLMs) to assist clinicians has attracted remarkable attention. Existing works mainly adopt the close-ended question-answering (QA) task with answer options for evaluation. However, many clinical decisions involve answering open-ended questions without pre-set options. To better understand LLMs in the clinic, we construct a benchmark ClinicBench. We first collect eleven existing datasets covering diverse clinical language generation, understanding, and reasoning tasks. Furthermore, we construct six novel datasets and clinical tasks that are complex but common in real-world practice, e.g., open-ended decision-making, long document processing, and emerging drug analysis. We conduct an extensive evaluation of twenty-two LLMs under both zero-shot and few-shot settings. Finally, we invite medical experts to evaluate the clinical usefulness of LLMs

Multi-task learning (MTL) aims to solve multiple tasks jointly by sharing a base representation among them. This can lead to more efficient learning and better generalization, as compared to learning each task individually. However, one issue that often arises in MTL is the convergence speed between tasks varies due to differences in task difficulty, so it can be a challenge to simultaneously achieve the best performance on all tasks with a single model checkpoint. Various techniques have been proposed to address discrepancies in task convergence rate, including weighting the per-task losses and modifying task gradients. In this work, we propose a novel approach that avoids the problem of requiring all tasks to converge at the same rate, but rather allows for “asynchronous” convergence among the tasks where each task can converge on its own schedule. As our main contribution, we monitor per-task validation metrics and switch to a knowledge distillation loss once a task has converged instead of continuing to train on the true labels. This prevents the model from overfitting on converged tasks while it learns the remaining tasks. We evaluate the proposed method in two 5-task MTL setups consisting of internal e-commerce datasets. The results show that our method consistently outperforms existing loss weighting and gradient balancing approaches, achieving average improvements of 0.9% and 1.5% over the best performing baseline model in the two setups, respectively.

Recent research has shown that large language models pretrained using unsupervised approaches can achieve significant performance improvement on many downstream tasks. Typically when adapting these language models to downstream tasks, like a classification or regression task, we employ a fine-tuning paradigm in which the sentence representation from the language model is input to a task-specific head; the model is then fine-tuned end-to-end. However, with the emergence of models like GPT-3, prompt-based fine-tuning has been proven to be a successful approach for few-shot tasks. Inspired by this work, we study discrete prompt technologies in practice. There are two issues that arise with the standard prompt approach. First, it can overfit on the prompt template. Second, it requires manual effort to formulate the downstream task as a language model problem. In this paper, we propose an improvement to prompt-based fine-tuning that addresses these two issues. We refer to our approach as DynaMaR – Dynamic Prompt with Mask Token Representation. Results show that DynaMaR can achieve an average improvement of 10% in few-shot settings and improvement of 3.7% in data-rich settings over the standard fine-tuning approach on four e-commerce applications.