Syntactic prediction requires models to generate outputs that are globally consistent and satisfy hierarchical structural constraints. Although large language models have been applied to such tasks, existing approaches still rely mainly on prompting or supervised fine-tuning, and their performance generally remains below that of specialized neural syntactic models. Our experiments show that supervised fine-tuning reaches a plateau, and additional supervised training does not further improve prediction quality. We therefore propose a two-stage post-training framework for syntactic prediction. The first cold-start stage adapt the model to a unified syntax generation format, and then applies reinforcement learning on a disjoint data subset to directly optimize structure-level objectives. Experiments on constituency parsing and CCG supertagging show that Syntax-RL consistently outperforms supervised LLM baselines and achieves state-of-the-art results. Further analyses suggest that its gains come primarily from the reinforcement learning stage. These findings identify reinforcement learning as an effective post-training mechanism for improving hierarchical syntactic prediction in large language models.

Email is becoming the universal interface for AI agents, but an ingested message contains un- trusted, malicious data the agent may execute as instructions. The indirect prompt-injection failure underlies real-world incidents such as EchoLeak (M365 Copilot, CVE-2025-32711) and the Gemini-in-Gmail hidden-HTML hijack. We present e2a, a deployable, open-source authenticated-email gateway that gives an agent a verified email identity and exposes two agent- agnostic defense hook points. The (G)ateway authenticates senders via SPF/DKIM/DMARC and parses MIME structure; the (L)LM-layer screens the normalized prompt for injection. On top of the gateway, we construct a bench- mark that isolates email structure as the first- class injection surface: one payload rendered across eight transport surfaces. We measure the per-surface evasion delta for open-source classifiers, three commercial security APIs, and frontier LLM judges. We find that structural hiding defeats detection. A naive text/plain- first parse costs every detector substantial AUC, dropping most classifiers below chance, and even on the gateway’s structure-aware view no non-LLM detector survives. A frontier LLM judge on the same view recovers most of the loss, and conditioning the threshold on crypto- graphic sender authentication (DMARC) lifts held-out recall further. No single layer suf- fices. e2a is live and open-source (Apache-2.0) at https://e2a.dev, released with the bench- mark and harness.

In this work, we propose FoRA-UA, a novel method that, using only 1–5% of the standard LoRA’s parameters, achieves state-of-the-art performance across a wide range of tasks. Specifically, we explore scenarios with extremely limited parameter budgets and derive two key insights: (1) fix-sized sparse frequency representations approximate small matrices more accurately; and (2) with a fixed number of trainable parameters, introducing a smaller intermediate representation to approximate larger matrices results in lower construction error. These findings form the foundation of our FoRA-UA method. By inserting a small intermediate parameter set, we achieve greater model compression without sacrificing performance. We evaluate FoRA-UA across diverse tasks, including natural language understanding (NLU), natural language generation (NLG), instruction tuning, and image classification, demonstrating strong generalisation and robustness under extreme compression.

This paper introduces UoRA, a novel parameter-efficient fine-tuning (PEFT) approach for large language models (LLMs). UoRA achieves state-of-the-art efficiency by leveraging a low-rank approximation method that reduces the number of trainable parameters without compromising performance. Unlike existing methods such as LoRA and VeRA, UoRA employs a re-parametrization mechanism that eliminates the need to adapt frozen projection matrices while maintaining shared projection layers across the model. This results in halving the trainable parameters compared to LoRA and outperforming VeRA in computation and storage efficiency. Comprehensive experiments across various benchmarks demonstrate UoRA’s superiority in achieving competitive fine-tuning performance with minimal computational overhead. We demonstrate its performance on GLUE and E2E benchmarks and is effectiveness in instruction-tuning large language models and image classification models. Our contributions establish a new paradigm for scalable and resource-efficient fine-tuning of LLMs.