Reinforcement learning (RL) is effective for improving code generation but suffers from data scarcity. While experience replay mitigates this, existing approaches rely on static, in-epoch metrics that overlook training dynamics, often introducing low-utility or outdated data. Analyzing RL dynamics via dataset cartography, we observe that “ambiguous” samples, which are vital for model generalization, rapidly migrate to “easy-to-learn” regions, diminishing their training value. To address this, we propose Adaptive Ambiguity Replay (A2R) for RL, a plug-and-play module that prioritizes cross-epoch ambiguous samples. To neutralize the noise from stale experiences, A2R incorporates an adaptive importance mechanism based on policy divergence to weigh replayed rollouts. Extensive experiments on nine LLMs (3B–14B) demonstrate that A2R outperforms state-of-the-art baselines on real-world code editing tasks across both unseen and learned domains. Our results highlight cross-epoch ambiguity as a key factor for effective replay in RL. Code: https://github.com/TsingZ0/verl-A2R

Reinforcement Learning with Verifiable Rewards (RLVR) serves as a cornerstone technique for enhancing the reasoning capabilities of Large Language Models (LLMs). However, its training is often plagued by entropy collapse, a rapid decline in policy entropy that limits exploration and undermines training effectiveness. While recent works attempt to mitigate this issue via several heuristic entropy interventions, the underlying mechanisms remain poorly understood. In this work, we conduct comprehensive theoretical and empirical analyses of entropy dynamics in RLVR, offering two main insights: (1) We derive a tight approximation for token-level entropy change at each update step, revealing four governing factors and providing a unified theoretical framework of how existing methods influence entropy; (2) We reveal a fundamental limitation of recent approaches: they rely on heuristic adjustments to one or two of these factors, leaving other relevant factors unconsidered, thus inherently limiting their effectiveness. Motivated by these findings, we propose STEER, a principled entropy-modulation method that adaptively reweighs tokens based on theoretically-estimated entropy variations. Extensive experiments across six mathematical reasoning and three coding benchmarks demonstrate that STEER effectively mitigates entropy collapse and consistently outperforms state-of-the-art baselines.

Recently, latent reasoning has been introduced into large language models (LLMs) to leverage rich information within a continuous space.However, without stochastic sampling, these methods inevitably collapse to deterministic inference, failing to discover diverse reasoning paths.To bridge the gap, we inject controllable stochasticity into latent reasoning via Gumbel-Softmax, restoring LLMs’ exploratory capacity and enhancing their compatibility with Reinforcement Learning (RL).Building on this, we propose **L**atent R**e**asoning **P**olicy **O**ptimization (**LEPO**), a novel framework that applies RL directly to continuous latent representations.Specifically, in rollout stage, LEPO maintains stochasticity to enable diverse trajectory sampling, while in optimization stage, LEPO constructs a unified gradient estimation for both latent representations and discrete tokens.

Large Language Model (LLM) agents are reshaping the industrial landscape. However, most practical agents remain human-designed because tasks differ widely, making them labor-intensive to build. This situation poses a central question: can we automatically create and adapt domain agents in the wild? While several recent approaches have sought to automate agent creation, they typically treat agent generation as a black-box procedure and rely solely on final performance metrics to guide the process. Such strategies overlook critical evidence explaining why an agent succeeds or fails, and often require high computational costs. To address these limitations, we propose ReCreate, an experience-driven framework for the automatic creation of domain agents. ReCreate systematically leverages agent interaction histories, which provide rich concrete signals on both the causes of success or failure and the avenues for improvement. Specifically, we introduce an agent-as-optimizer paradigm that effectively learns from experience via three key components: (i) an experience storage and retrieval mechanism for on-demand inspection; (ii) a reasoning–creating synergy pipeline that map execution experience into scaffold edits; and (iii) hierarchical updates that abstract instance-level details into reusable domain patterns. In experiments across diverse domains, ReCreate consistently outperforms human-designed agents and existing automated agent generation methods, even when starting from minimal seed scaffolds.

Reinforcement learning (RL) is widely used for post-training large language models (LLMs) in code editing, where group-relative methods, such as GRPO, are popular due to their critic-free and normalized advantage estimation. However, in real-world code-editing scenarios, reward distributions are often skewed with unpredictable noise, leading to distorted advantage computation and increased rollout outliers. To address this issue, we propose Group Adaptive Policy Optimization (GAPO), which adaptively finds an interval with the highest SNR (Signal to Noise Ratio) per prompt and uses the median of that interval as an adaptive Q to replace the group mean in advantage calculation to reduce noise further. This adaptive Q robustly handles rollout noise while remaining plug-and-play and efficient. We evaluate GAPO on nine instruction-tuned LLMs (3B–14B) using a collected large dataset of 51,844 real-world, history-aware code-editing tasks spanning 10 programming languages. GAPO yields up to 4.35 in-domain (ID) and 5.30 out-of-domain (OOD) exact-match improvements over GRPO and its variant DAPO, while achieving lower clipping ratios and higher GPU throughput. Code: https://github.com/TsingZ0/verl-GAPO

Large language models (LLMs) have been widely adopted in mathematical optimization in scientific scenarios for their extensive knowledge and advanced reasoning capabilities. Existing methods mainly focus on utilizing LLMs to solve optimization problems in a prompt-based manner, which takes observational feedback as additional textual descriptions. However, due to LLM’s **high sensitivity to the prompts** and **tendency to get lost in lengthy prompts**, these methods struggle to effectively utilize the observational feedback from each optimization step, which severely hinders the applications for real-world scenarios. To address these challenges, we propose a conceptually simple and general bi-level optimization method, namely **G**eneral **S**cientific **O**ptimizers (GSO).Specifically, GSO first utilizes inner-level simulators as experimental platforms to evaluate the current solution and provide observational feedback. Then, LLMs serve as knowledgeable and versatile scientists, generating new solutions by refining potential errors from the feedback as the outer-level optimization.Finally, simulations together with the expert knowledge in LLMs are jointly updated with bi-level interactions via model editing.Extensive experiments show that GSO consistently outperforms existing state-of-the-art methods using *six* different LLM backbone on *seven* different tasks, demonstrating the effectiveness and a wide range of applications.

Recent work has shown that large pretrained Language Models (LMs) can not only perform remarkably well on a range of Natural Language Processing (NLP) tasks but also start improving on reasoning tasks such as arithmetic induction, symbolic manipulation, and commonsense reasoning with increasing size of models. However, it is still unclear what the underlying capabilities of these LMs are. Surprisingly, we find that these models have limitations on certain basic symbolic manipulation tasks such as copy, reverse, and addition. When the total number of symbols or repeating symbols increases, the model performance drops quickly. We investigate the potential causes behind this phenomenon and examine a set of possible methods, including explicit positional markers, fine-grained computation steps, and LMs with callable programs. Experimental results show that none of these techniques can solve the simplest addition induction problem completely. In the end, we introduce LMs with tutor, which demonstrates every single step of teaching. LMs with tutor is able to deliver 100% accuracy in situations of OOD and repeating symbols, shedding new insights on the boundary of large LMs in induction.

Building dialogue systems requires a large corpus of annotated dialogues. Such datasets are usually created via crowdsourcing, which is expensive and time-consuming. In this paper, we propose Dialogic, a novel dialogue simulation method based on large language model in-context learning to automate dataset creation. Seeded with a few annotated dialogues, Dialogic automatically selects in-context examples for demonstration and prompts GPT-3 to generate new dialogues and annotations in a controllable way. Our method can rapidly expand a small set of dialogue data with minimum or zero human involvement and parameter update and is thus much more cost-efficient and time-saving than crowdsourcing. Experimental results on the MultiWOZ dataset demonstrate that training a model on the simulated dialogues leads to even better performance than using the same amount of human-generated dialogues under the challenging low-resource settings, with as few as 85 dialogues as a seed. When the full training set is given, our method can still serve as an effective data augmentation method to further improve performance. Human evaluation results also show that our simulated dialogues have near-human fluency and annotation accuracy. The code and data are available at https://github.com/Leezekun/dialogic.

Commonly used information retrieval methods such as TF-IDF in open-domain question answering (QA) systems are insufficient to capture deep semantic matching that goes beyond lexical overlaps. Some recent studies consider the retrieval process as maximum inner product search (MIPS) using dense question and paragraph representations, achieving promising results on several information-seeking QA datasets. However, the pretraining of the dense vector representations is highly resource-demanding, e.g., requires a very large batch size and lots of training steps. In this work, we propose a sample-efficient method to pretrain the paragraph encoder. First, instead of using heuristically created pseudo question-paragraph pairs for pretraining, we use an existing pretrained sequence-to-sequence model to build a strong question generator that creates high-quality pretraining data. Second, we propose a simple progressive pretraining algorithm to ensure the existence of effective negative samples in each batch. Across three open-domain QA datasets, our method consistently outperforms a strong dense retrieval baseline that uses 6 times more computation for training. On two of the datasets, our method achieves more than 4-point absolute improvement in terms of answer exact match.

Existing work on automated hate speech classification assumes that the dataset is fixed and the classes are pre-defined. However, the amount of data in social media increases every day, and the hot topics changes rapidly, requiring the classifiers to be able to continuously adapt to new data without forgetting the previously learned knowledge. This ability, referred to as lifelong learning, is crucial for the real-word application of hate speech classifiers in social media. In this work, we propose lifelong learning of hate speech classification on social media. To alleviate catastrophic forgetting, we propose to use Variational Representation Learning (VRL) along with a memory module based on LB-SOINN (Load-Balancing Self-Organizing Incremental Neural Network). Experimentally, we show that combining variational representation learning and the LB-SOINN memory module achieves better performance than the commonly-used lifelong learning techniques.

Existing question answering datasets focus on dealing with homogeneous information, based either only on text or KB/Table information alone. However, as human knowledge is distributed over heterogeneous forms, using homogeneous information alone might lead to severe coverage problems. To fill in the gap, we present HybridQA, a new large-scale question-answering dataset that requires reasoning on heterogeneous information. Each question is aligned with a Wikipedia table and multiple free-form corpora linked with the entities in the table. The questions are designed to aggregate both tabular information and text information, i.e., lack of either form would render the question unanswerable. We test with three different models: 1) a table-only model. 2) text-only model. 3) a hybrid model that combines heterogeneous information to find the answer. The experimental results show that the EM scores obtained by two baselines are below 20%, while the hybrid model can achieve an EM over 40%. This gap suggests the necessity to aggregate heterogeneous information in HybridQA. However, the hybrid model’s score is still far behind human performance. Hence, HybridQA can serve as a challenging benchmark to study question answering with heterogeneous information.

Conventional approaches to relation extraction usually require a fixed set of pre-defined relations. Such requirement is hard to meet in many real applications, especially when new data and relations are emerging incessantly and it is computationally expensive to store all data and re-train the whole model every time new data and relations come in. We formulate such challenging problem as lifelong relation extraction and investigate memory-efficient incremental learning methods without catastrophically forgetting knowledge learned from previous tasks. We first investigate a modified version of the stochastic gradient methods with a replay memory, which surprisingly outperforms recent state-of-the-art lifelong learning methods. We further propose to improve this approach to alleviate the forgetting problem by anchoring the sentence embedding space. Specifically, we utilize an explicit alignment model to mitigate the sentence embedding distortion of learned model when training on new data and new relations. Experiment results on multiple benchmarks show that our proposed method significantly outperforms the state-of-the-art lifelong learning approaches.

With social media becoming increasingly popular on which lots of news and real-time events are reported, developing automated question answering systems is critical to the effective-ness of many applications that rely on real-time knowledge. While previous datasets have concentrated on question answering (QA) for formal text like news and Wikipedia, we present the first large-scale dataset for QA over social media data. To ensure that the tweets we collected are useful, we only gather tweets used by journalists to write news articles. We then ask human annotators to write questions and answers upon these tweets. Unlike otherQA datasets like SQuAD in which the answers are extractive, we allow the answers to be abstractive. We show that two recently proposed neural models that perform well on formal texts are limited in their performance when applied to our dataset. In addition, even the fine-tuned BERT model is still lagging behind human performance with a large margin. Our results thus point to the need of improved QA systems targeting social media text.

A key challenge of multi-hop question answering (QA) in the open-domain setting is to accurately retrieve the supporting passages from a large corpus. Existing work on open-domain QA typically relies on off-the-shelf information retrieval (IR) techniques to retrieve answer passages, i.e., the passages containing the groundtruth answers. However, IR-based approaches are insufficient for multi-hop questions, as the topic of the second or further hops is not explicitly covered by the question. To resolve this issue, we introduce a new subproblem of open-domain multi-hop QA, which aims to recognize the bridge (i.e., the anchor that links to the answer passage) from the context of a set of start passages with a reading comprehension model. This model, the bridge reasoner, is trained with a weakly supervised signal and produces the candidate answer passages for the passage reader to extract the answer. On the full-wiki HotpotQA benchmark, we significantly improve the baseline method by 14 point F1. Without using any memory inefficient contextual embeddings, our result is also competitive with the state-of-the-art that applies BERT in multiple modules.

Existing models for extractive summarization are usually trained from scratch with a cross-entropy loss, which does not explicitly capture the global context at the document level. In this paper, we aim to improve this task by introducing three auxiliary pre-training tasks that learn to capture the document-level context in a self-supervised fashion. Experiments on the widely-used CNN/DM dataset validate the effectiveness of the proposed auxiliary tasks. Furthermore, we show that after pre-training, a clean model with simple building blocks is able to outperform previous state-of-the-art that are carefully designed.