RL post-training strategies are dataset-dependent and reveal a recurring empirical pattern: capacity parameters accumulate monotonically across stages, while regularization parameters predominantly oscillate in response to shifting training dynamics. This distinction highlights a potential flaw in fixed training schedules: by forcing all parameters along rigid paths, they fail to capture the dynamic exploration-exploitation tradeoffs that regularization must track. We uncover this through LLMZero, an agentic system that optimizes training trajectories via tree search by diagnosing pathologies at each checkpoint and proposing coordinated multi-parameter transitions. Across four diverse GRPO tasks, LLMZero discovers strategies that improve over the base model by 9% to 140% and over grid search by 6% to 15% (relative), consistently outperforming random search and a skill-based agent under a matched compute budget. The capacity–regularization asymmetry is consistent across all four tasks, offering a candidate design heuristic for multi-stage training.

When an LLM agent controls real money, prompt injection stops being an information risk and becomes a financial one—and its hardest form is the confused deputy: paying is the legitimate job and only the recipient or amount is attacker-supplied, so an instruction-level safeguard has nothing to refuse. Across five models (Claude, Qwen, MiniMax) a normally-built agent is driven to a forbidden money action on up to 87% of such scenarios—and a hardened refusal prompt still loses up to 84%. Like prior work on policy enforcement outside the model (e.g. CaMeL), we take the model off the money path rather than trust it to refuse; we contribute the first open, installable, end-to-end realization of this for agent payments: BNB Agent Studio, an open-source toolchain (CLI, Python SDK, MCP server) that takes a developer from an empty directory to a deployed, self-funding LLM agent in six commands. Its two-layer, keyless-edge architecture keeps signing as fixed code isolated from the model context, prices deterministic, and spend guardrails enforced below the LLM: an identical agent fooled at a comparable rate moves funds on 0/45 everywhere, and a live testnet deployment confirms escrow, pricing, key-access, and suspension on-chain. The toolchain implements the Agent Lifecycle Protocol (ALP), a vendor-neutral draft spec composing ERC-8004 identity, ERC-8183 task commerce, x402 payments, and A2A. It is on PyPI (pip install bnbagent-studio, Apache-2.0), with a screencast at https://youtu.be/yjdvN8vE37A.

Large Language Models (LLMs) often suffer from catastrophic forgetting when learning multiple tasks sequentially, making continual learning (CL) essential for their dynamic deployment. Existing state-of-the-art (SOTA) methods, such as O-LoRA, typically focus on constructing orthogonality tasks to decouple parameter interdependence from various domains.In this paper, we reveal that building non-collision parameters is a more critical factor in addressing CL challenges. Our theoretical and experimental analyses demonstrate that non-collision parameters provide better task orthogonality, which is a sufficient but unnecessary condition. Furthermore, knowledge from multiple domains will be preserved in non-collision parameter subspaces, making it more difficult to forget previously seen data. Leveraging this insight, we propose Non-collision Low-Rank Adaptation (N-LoRA), a simple yet effective approach leveraging low collision rates to enhance CL in LLMs. Experimental results on multiple CL benchmarks indicate that N-LoRA achieves superior performance (+2.9%), higher task orthogonality (×4.1times), and lower parameter collision (×58.1times) than SOTA methods.

Wide usage of ChatGPT has highlighted the potential of reinforcement learning from human feedback. However, its training pipeline relies on manual ranking, a resource-intensive process. To reduce labor costs, we propose a self-supervised text ranking approach for applying Proximal-Policy-Optimization to fine-tune language models while eliminating the need for human annotators. Our method begins with probabilistic sampling to encourage a language model to generate diverse responses for each input. We then employ TextRank and ISODATA algorithms to rank and cluster these responses based on their semantics. Subsequently, we construct a reward model to learn the rank and optimize our generative policy. Our experimental results, conducted using two language models on three tasks, demonstrate that the models trained by our method considerably outperform baselines regarding BLEU, GLEU, and METEOR scores. Furthermore, our manual evaluation shows that our ranking results exhibit a remarkably high consistency with that of humans. This research significantly reduces training costs of proximal policy-guided models and demonstrates the potential for self-correction of language models.

While deep learning models are powerful, they have limitations in tasks that require commonsense reasoning, as these tasks often involve interpreting information that may not be directly available in the input. Providing evidence has been proven to significantly enhance performance in commonsense reasoning tasks. However, there are various perspectives on evidence, including natural language explanations generated by pre-trained language models, facts derived from world knowledge like text corpora and knowledge bases, and rationales extracted from the input context. Hence, it is crucial to determine how to estimate the confidence degree of different evidence and how to combine them reliably. To address these challenges, this study proposes a trusted multi-view evidential fusion framework for reliable commonsense reasoning tasks that dynamically assesses the confidence of evidence and combines different views of evidence in a trustworthy manner. The proposed method is applied to three commonsense question-answering benchmarks, demonstrating that this approach can effectively reason with multi-view evidence and can compete with state-of-the-art performance.

Style transfer is the task of paraphrasing text into a target-style domain while retaining the content. Unsupervised approaches mainly focus on training a generator to rewrite input sentences. In this work, we assume that text styles are determined by only a small proportion of words; therefore, rewriting sentences via generative models may be unnecessary. As an alternative, we consider style transfer as a sequence tagging task. Specifically, we use edit operations (i.e., deletion, insertion and substitution) to tag words in an input sentence. We train a classifier and a language model to score tagged sequences and build a conditional random field. Finally, the optimal path in the conditional random field is used as the output. The results of experiments comparing models indicate that our proposed model exceeds end-to-end baselines in terms of accuracy on both sentiment and style transfer tasks with comparable or better content preservation.

Style transfer is the task of transferring a sentence into the target style while keeping its content. The major challenge is that parallel corpora are not available for various domains. In this paper, we propose a Mask-And-Regenerate approach (MAR). It learns from unpaired sentences by modifying the word-level style attributes. We cautiously integrate the deletion, insertion and substitution operations into our model. This enables our model to automatically apply different edit operations for different sentences. Specifically, we train a multilayer perceptron (MLP) as a style classifier to find out and mask style-characteristic words in the source inputs. Then we learn a language model on non-parallel data sets to score sentences and remove unnecessary masks. Finally, the masked source sentences are input to a Transformer to perform style transfer. The final results show that our proposed model exceeds baselines by about 2 per cent of accuracy for both sentiment and style transfer tasks with comparable or better content retention.

Word alignment is essential for the downstream cross-lingual language understanding and generation tasks. Recently, the performance of the neural word alignment models has exceeded that of statistical models. However, they heavily rely on sophisticated translation models. In this study, we propose a super lightweight unsupervised word alignment model named MirrorAlign, in which bidirectional symmetric attention trained with a contrastive learning objective is introduced, and an agreement loss is employed to bind the attention maps, such that the alignments follow mirror-like symmetry hypothesis. Experimental results on several public benchmarks demonstrate that our model achieves competitive, if not better, performance compared to the state of the art in word alignment while significantly reducing the training and decoding time on average. Further ablation analysis and case studies show the superiority of our proposed MirrorAlign. Notably, we recognize our model as a pioneer attempt to unify bilingual word embedding and word alignments. Encouragingly, our approach achieves 16.4X speedup against GIZA++, and 50X parameter compression compared with the Transformer-based alignment methods. We release our code to facilitate the community: https://github.com/moore3930/MirrorAlign.