While Supervised Fine-Tuning (SFT) is the prevailing method for equipping Large Language Models (LLMs) with function calling capabilities, its effectiveness is often compromised by two critical challenges: 1) Imbalanced Training Signals, where lengthy Chain-of-Thought (CoT) reasoning tokens dominate the training signals over concise function calls in the learning objective, and 2) Imbalanced Data Hardness, characterized by a scarcity of hard training examples. To overcome these limitations, we propose Balanced Supervised Fine-tuning (BalanceSFT), a novel framework that incorporates two key components: a Self-adjusted Signal Balancing (SSB) loss that employs a learnable hyperparameter to dynamically adjust the token contributions of CoT reasoning and function calls, together with a Hard Data Re-sampling (HDR) strategy that establishes a feedback loop to selectively generate new, high-quality complex data guided by model errors. Extensive experiments demonstrate the effectiveness of our proposed BalanceSFT framework. With BalanceSFT, a 7B model achieves function calling performance that surpasses state-of-the-art models like GPT-5. Our code, models, and dataset are open-sourced.

Language Reasoning Models (LRMs) achieve strong performance by scaling test-time computation but often suffer from “overthinking”, producing excessively long reasoning traces that increase latency and memory usage. Existing LRMs typically enforce conciseness with uniform length penalties, which over-compress crucial early deduction steps at the sequence level and indiscriminately penalize all queries at the group level. To solve these limitations, we propose PACE, a dual-level framework for prefix-protected and difficulty-aware compression under hierarchical supervision. At the sequence level, prefix-protected optimization employs decaying mixed rollouts to maintain valid reasoning paths while promoting conciseness. At the group level, difficulty-aware penalty dynamically scales length constraints based on query complexity, maintaining exploration for harder questions while curbing redundancy on easier ones. Extensive experiments on DeepSeek-R1-Distill-Qwen (1.5B/7B) demonstrate that PACE achieves a substantial reduction in token usage (up to 55.7%) while simultaneously improving accuracy (up to 4.1%) on math benchmarks, with generalization ability to code, science, and general domains.

The advancement of LLM agents with tool-use capabilities requires diverse and complex training corpora. Existing data generation methods, which predominantly follow the paradigm of random sampling and shallow generation, often yield simple and homogeneous trajectories that fail to capture complex, implicit logical dependencies. To bridge this gap, we introduce **HardGen**, an automatic agentic pipeline designed to generate hard tool-use training samples. First, HardGen establishes a dynamic API Graph built upon agent failure cases, from which it samples hard tool-execution sequences. Second, these execution sequences are used to reversely build advanced tools that combine multi-step operations into one, which then help generate hard queries requiring unstated intermediate steps. Finally, advanced tools and hard queries drive the generation of Chain-of-Thought (CoT), which is iteratively refined through closed-loop verification feedback. Extensive evaluations demonstrate that a 4B parameter model trained with our curated dataset achieves superior performance compared to several leading open-source and closed-source competitors (e.g., GPT-5.2, Gemini-3-Pro and Claude-Opus-4.5). Our code, models and the curated dataset with 27,000 samples will be open-sourced to facilitate future research.

While prior work focuses on treating answer verification as a general reasoning or semantic matching problem, significant challenges remain in computation-oriented scientific domains. In this paper, we introduce CoSineVerifier, an agentic verifier that actively invokes external executors to ground its judgments in precise computations and symbolic reasoning. To train this accurate agentic verifier, we propose a novel verifier training data augmentation method to increase the correctness and efficiency of tool-invoked verifications. Extensive experiments across STEM, QA, and long-form reasoning tasks demonstrate CoSineVerifier’s robust generalization abilities, achieving state-of-the-art performance on VerifyBench-Hard and SCI-Bench. Furthermore, when employed as an RLVR reward model, CoSineVerifier consistently outperforms both rubric- and model-based verifiers on AIME’24, AIME’25 and GPQA-D, highlighting its potential to advance LLM reasoning and accelerate data processing

The unlearning method aims at effectively removing harmful, sensitive, or outdated knowledge without costly retraining the model. However, existing methods suffer from two critical limitations: (1) collateral forgetting, where erasing target data inadvertently removes related but desirable knowledge, and (2) generality forgetting, where aggressive unlearning degrades the model’s general capabilities. To address these challenges, we propose DirectiOn Guide unlEarning (DOGE), a novel method that enables precise knowledge erasure by identifying and leveraging a targeted “unlearning direction” in the model’s parameter space. DOGE first extracts this direction through differential analysis of representations for forgotten and retained samples, pinpointing the exact subspace associated with unwanted knowledge. It then selectively applies updates along this direction, ensuring minimal interference with retained information and general model performance. Experiments across multiple benchmarks demonstrate that Doge achieves state-of-the-art unlearning precision while preserving both related knowledge and general capabilities.