Conversational assistants powered by large language models (LLMs) excel at tool-use tasks but struggle with adhering to complex, business-specific rules. While models can reason over business rules provided in context, including all policies for every query introduces high latency and wastes compute. Furthermore, these lengthy prompts lead to long contexts, harming overall performance due to the ’needle-in-a-haystack’ problem. To address these challenges, we propose a multi-stage alignment method that teaches models to recall and apply relevant business policies during chain-of-thought reasoning at inference time, without including the full business policy in-context. Furthermore, we introduce a novel Policy Recall reward based on the Jaccard score and a Hallucination Penalty for GRPO training. Altogether, our best model outperforms the baseline by **16 points** and surpasses comparable *in-context* baselines of similar model size by **3 points**, while using **40% fewer words**.

The rise of fully autonomous agents necessitates a rigorous evaluation of their reliability under **real-world API complexity**. Accordingly, moving beyond perfectly **sanitized** settings that disregard realistic obstacles, we introduce ***WildAGTEval***, a configurable benchmark environment that accounts for two critical dimensions of realistic API complexity: **API specification**, which includes detailed documentation and usage constraints, and **API execution**, which captures runtime challenges. Consequently, ***WildAGTEval*** offers (i) an API system spanning 86 functions across 7 domains with 60 independently activatable complexity scenarios, yielding approximately 32K distinct test configurations, and (ii) 300 multi-turn conversations grounded in this system. Through systematic evaluation of 11 state-of-the-art LLMs, we find that: (1) individual complexities reduce accuracy by up to 23.1%; (2) compounding them amplifies this drop to 52.1%; and crucially, (3) failing agents fabricate success and distort intent rather than admitting limitations, exposing a critical reliability gap invisible to existing benchmarks.

Reinforcement learning (RL) has become a cornerstone of the post-training pipeline for large language models (LLMs), enabling capabilities such as complex reasoning and tool use. However, standard RL approaches face significant challenges due to reward sparsity. Moreover, LLMs typically exhibit mode-seeking behavior, concentrating probability mass on high-likelihood regions. This lack of diversity biases the model toward premature exploitation, hindering the exploration necessary for optimal learning. To address this, we propose VEG (verbal 𝜖-greedy), a novel framework that leverages external feedback as a dynamic control variable to explicitly balance exploration and exploitation within the semantic space. This method not only supplements sparse final rewards with intermediate signals but also enforces sustained exploration throughout the training process. Experiments on Tau Bench and SearchQA demonstrate that our method achieves superior accuracy compared to standard RL baselines. Notably, the trained policy eventually outperforms the external feedback model itself, demonstrating that VEG enables the agent to effectively filter and improve upon the guidance it receives.