Generative adversarial distillation (GAD) provides on-policy feedback for black-box distillation by training a discriminator that ranks teacher and student responses to the same prompt. While effective, GAD is also costly: every step pays for a discriminator update, an actor update over the full response, and the indirect cost of recovering whenever the discriminator and the actor progress at incompatible rates. We argue that the right way to make GAD efficient is not to make individual updates cheaper but to decide **when each update is worth its cost**. We introduce **curriculum-efficient GAD (CE-GAD)**, an adaptive recipe that monitors a small set of training signals and uses them to:

- (i) skip discriminator updates that would not change the reward, - (ii) constrain actor updates and prune per-token gradients to the subset that actually carries signal, and - (iii) rescale rewards and normalize advantages to preserve within-batch contrast under saturation.

Evaluations across four models from two open-source families (Qwen2.5 at 3B/7B/14B and Llama-3.2-3B) demonstrate that CE-GAD matches or surpasses vanilla GAD on quality while winning on 12 out of 16 individual benchmark configurations. Specifically, on Qwen2.5-7B, CE-GAD improves the average score by up to **+3%** relative compared to the base model, while reducing per-step wall-clock time by up to **8x** relative to vanilla GAD.

In settings where labeled verifiable training data is the binding constraint, each checked example should be allocated to the model and reward density where it is most informative. We identify a reward-density principle: sparse sequence-level reward is most useful on models that can explore and discover better behavior, while dense token-level teacher supervision is better suited for compressing that behavior into a smaller deployment model; accordingly, scarce labeled data should be used upstream on the strongest available teacher, then transferred downstream as dense supervision.

We evaluate this rule through a four-stage workflow: teacher RL, forward-KL warmup, on-policy distillation, optional post-bridge student RL on verifiable math and code generation with Qwen3, Llama, and Qwen2.5 models. At fixed Qwen3-1.7B deployment-student size, an RL-improved 8B teacher distilled through the dense bridge outperforms direct GRPO on the same student (79.3% vs. 75.9% on MATH; 25.2% vs. 19.8% on AIME2024, avg@16), while transfer from the same teacher before RL underperforms. A component ablation confirms that each stage is load-bearing: the RL-improved teacher adds 7.8 absolute points on MATH over a raw teacher of the same size, the forward-KL warmup adds 1.7, and on-policy distillation adds 3.3. The teacher-quality ordering: raw-teacher transfer < direct GRPO < RL-teacher transfer replicates on Llama-3.1-8B-Instruct with a Llama-3.3-70B-Instruct teacher and on Qwen2.5-7B-Instruct for code generation with a Qwen2.5-32B-Instruct teacher, where the dense bridge improves HumanEval and MBPP by 3.3 and 3.8 absolute points, respectively, over direct student GRPO. The operational lesson is to avoid spending scarce labeled data on the least prepared policy: use sparse reward for teacher-side discovery, dense transfer for student compression, and student-side sparse reward only after the bridge. The OPD implementation used in our experiments is available at https://github.com/HJSang/OPSD_OnPolicyDistillation.

Large language models (LLMs) face a critical alignment challenge: balancing safety with helpfulness. Excessive safety can lead to over-refusal, where models reject harmful-looking yet benign queries, severely limiting utility.Existing training-free interventions offer an efficient way to mitigate over-refusal without re-training, but suffer from high inference overhead and architecture dependency. Our work explores a complementary direction: rather than applying post-hoc corrections to model outputs, our goal is to intrinsically reshape the distributions of harmful and benign samples within the model’s decision space. In this paper, we argue that a lightweight training-based approach can more effectively distinguish between harmful and benign samples. We propose Single Token Alignment (STA), which optimizes only a single-token prefix (e.g., 4,096 parameters) while keeping the base model frozen. To address the inherent challenge of achieving robust refinement through such a minimal parameter interface, STA employs a mixed weighting mechanism integrated with its optimization objective. This mechanism incorporates hard weighting via stringent data filtering to provide clear, unbiased learning signals, and soft weighting through a focal mechanism to prioritize challenging cases.Extensive experiments across 9 models and 10 datasets demonstrate that STA achieves a superior safety-helpfulness balance for LLMs, MLLMs, and reasoning models, offering a highly efficient and generalizable solution for refining safety alignment.

Distilling the capabilities from a large reasoning model (LRM) to a smaller student model often involves training on substantial amounts of reasoning data. However, knowledge distillation (KD) over lengthy sequences with prompt (P), chain-of-thought (CoT), and answer (A) sections makes the process computationally expensive. In this work, we investigate how the allocation of supervision across different sections (P, CoT, A) affects student performance. Our analysis shows that selective KD over only the CoT tokens can be effective when the prompt and answer information is encompassed by it. Building on this insight, we establish a truncation protocol to quantify computation-quality tradeoffs as a function of sequence length. We observe that beyond a specific length, longer training sequences provide marginal returns for downstream performance but require substantially higher memory and FLOPs. To this end, training on only the first 50% of tokens of every training sequence can retain, on average, ≈91% of full-sequence performance on math benchmarks while reducing training time, memory usage, and FLOPs by about 50% each. Codes are available at https://github.com/weiruichen01/distilling-the-essence.

Although Large Audio-Language Models (LALMs) deliver state-of-the-art (SOTA) performance, they frequently suffer from hallucinations, e.g., generating text not grounded in the audio input. We analyze these grounding failures and identify a distinct taxonomy: Event Omission, False Event Identity, Temporal Relation Error, and Quantitative Temporal Error. To address this, we introduce the AHA (Audio Hallucination Alignment) framework. By leveraging counterfactual hard negative mining, our pipeline constructs a high-quality preference dataset that forces models to distinguish strict acoustic evidence from linguistically plausible fabrications. Additionally, we establish AHA-Eval, a diagnostic benchmark designed to rigorously test these fine-grained reasoning capabilities. We apply this data to align Qwen2.5-Omni. The resulting model, Qwen-Audio-AHA, achieves a 13.7% improvement on AHA-Eval. Crucially, this benefit generalizes beyond our diagnostic set. Our model shows substantial gains on public benchmarks, including 1.3% on MMAU-Test and 1.6% on MMAR, outperforming latest SOTA methods.

Large language models (LLMs) have demonstrated remarkable performance across a wide range of industrial applications, from search and recommendation systems to generative tasks. Although scaling laws indicate that larger models generally yield better generalization and performance, their substantial computational requirements often render them impractical for many real-world scenarios at scale. In this paper, we present a comprehensive set of insights for training and deploying small language models (SLMs) that deliver high performance for a variety of industry use cases. We focus on two key techniques: (1) knowledge distillation and (2) model compression via structured pruning and quantization. These approaches enable SLMs to retain much of the quality of their larger counterparts while significantly reducing training/serving costs and latency. We detail the impact of these techniques on a variety of use cases in a large professional social network platform and share deployment lessons, including hardware optimization strategies that improve speed and throughput for both predictive and reasoning-based applications in Recommendation Systems.