Model diffing studies the internal changes of a model from training to reveal learning mechanisms. To enhance interpretability, existing studies often strive for structural sparsity by removing redundant model components. However, we emphasize that for most fine-tuning methods, the fundamental source of internal change, i.e., the parameter updates, is non-sparse. The dense update mechanism contradicts the structural sparsity sought by researchers. Thus, we seek to bring sparsity to fine-tuning via an interpretability-driven method called circuit-tuning. It employs a heuristic prune-then-tune paradigm, forcing the model to optimize within a restricted subset of parameters. To test our method, we designed a novel task called Subject-Verb Disagreement (SVD) and conducted comprehensive explorations on it. Experimental results demonstrate that, compared to other fine-tuning methods, circuit-tuning achieves better sparsity in parameter updates, thus making it possible for faithful model diffing. Based on this, we performed stage-wise model diffing, revealing the evolution of model capabilities during fine-tuning from a new view. Furthermore, experiments on gender debiasing and larger models verify the practical utility of our approach. Our work provides a new methodology for model diffing-based interpretability and its applications.

Integrating supervised fine-tuning (SFT) and reinforcement learning (RL) has become a common post-training paradigm for improving the reasoning capabilities of large language models. Existing methods typically rely on sequential SFT–RL pipelines or static loss weights, overlooking sample-level heterogeneity and evolving learning states. We propose Sample Learning State (SLS), a rollout-based characterization of each training sample using two complementary signals: sample mastery and exploration-trajectory dispersion. These signals estimate whether reward-driven updates are reliable for the core task and whether current rollouts provide informative contrast for RL optimization, respectively. Based on SLS, we design a state-aware coefficient that dynamically balances SFT and RL losses, increasing the RL weight only when the model demonstrates sufficient core-task competence and the sampled trajectories provide useful reward contrast. Experiments on Qwen2.5-Math models show that SLS achieves state-of-the-art results on four in-distribution and two out-of-distribution mathematical reasoning benchmarks. In multi-reward settings, SLS preserves core reasoning performance while mitigating reward hacking induced by auxiliary rewards.

The prohibitive memory footprint of the Key-Value (KV) cache imposes a critical bottleneck for efficient long-context LLM serving. Current eviction methods, ranging from attention-based heuristics to recent learnable approaches, rely on scoring criteria decoupled from the downstream task performance that ultimately matters. To address this, we introduce KV-Retriever, a layer-wise learnable importance scoring module directly optimized by the end-to-end downstream task loss. We demonstrate that aligning the scoring objective with the actual task signal—rather than intermediate proxies such as attention statistics or oracle labels—yields a fundamentally more accurate notion of token importance. Functioning as a lightweight, plug-and-play module with only ∼0.01% additional parameters on a frozen 8B backbone, KV-Retriever is compatible with diverse host eviction frameworks. Extensive experiments show that, for each backbone, one trained retriever can be reused unchanged across multiple KV cache compression architectures and delivers consistent improvements on LongBench, LongBench-v2, and RULER. KV-Retriever requires one-time supervised training with golden answers; “plug-and-play” refers to deployment into otherwise unchanged host eviction frameworks, not to training-free adaptation. Our code is available at https://github.com/Taylor-Gavel/kv_retriever.

Prompt-based in-context learning (ICL) and parameter fine-tuning are two dominant paradigms for incorporating external information into large language models (LLMs), but they incur high inference costs or require expensive retraining. To bridge this gap, context-to-parameter mapping converts prompts into temporary adapter weights. However, we identify a critical failure mode in existing methods: hidden-state collapse, where the adapter-augmented model’s internal states diverge sharply from the full-context oracle in deeper layers. We trace this failure to two coupled gaps: suboptimal Input-Selection and inadequate Supervision-Signal. To address these issues, we propose SADA (State-Aligned Distillation Adapters). We establish the attention-block output as a principled feature interface to improve input selection and introduce state-alignment distillation to enforce consistency between the adapter-augmented model and the full-context oracle. Experiments on long-context language modeling (PG19) and downstream NLU and summarization benchmarks show that SADA consistently outperforms strong baselines like StreamAdapter and GenerativeAdapter, achieving performance comparable to ICL while significantly reducing memory footprint and latency. We further analyze when parameterized context compression is effective and when explicit context retention remains preferable. Our code is available at https://github.com/Taylor-Gavel/SADA.git.