supervised-fine-tuning

PrimeSeeker 提出一种面向能力的监督框架，通过「潜在锚点推理」将深度搜索分解为锚点解析与关系传递的耦合操作，据此构建网络证据结构、联合生成问题与参考证据骨架，训练出 30B 搜索智能体，在五个深度搜索基准上以更少工具调用取得强性能。

The paper introduces OPSFT (On-Policy direction-constrained Supervised Fine-Tuning), which constrains SFT parameter update directions to those observed in on-policy RL training, showing that on-policy generalization advantages can be transferred to SFT for improved efficiency and the ability to leverage high-quality trajectories.

This paper proposes Scaffolding Minds, which learns a dedicated scaffolding encoder for optimized latent visual target representations and learns both mean and variance of the RL sampler, improving latent reasoning baselines by +9.5 points on Frozen Lake spatial planning and +5.6 points on average across nine visual reasoning benchmarks.

The paper proposes HybridCUA, a method that combines GUI and CLI interactions for computer-use agents, using a two-stage training framework with supervised fine-tuning and reinforcement learning to improve task accuracy and cross-platform generalizability.

The paper proposes a lightweight, rule-based selector for diverse SFT traces to improve post-RL generalization in reasoning models, showing significant performance gains on mathematical benchmarks.

This paper shows that LLM-based scorers producing equal ranking quality can still make different decisions when candidate order in the prompt changes, and introduces order-consistency SFT (OC-SFT) to penalize score disagreement across permutations. OC-SFT holds ranking quality while improving decision stability across reranking, response ranking, and multi-document QA tasks.

CaptchaArena is a large-scale, fine-grained dataset for training computer-use agents to solve interactive CAPTCHAs, containing 50K puzzles and annotations. The associated CaptchaAgent model achieves 71.7% accuracy using supervised and reinforcement learning.

Spatial-Interactor is a framework that trains vision-language models to enhance spatial reasoning through interaction with the physical world, employing a three-level curriculum and two-stage training strategy to improve state transition modeling and long-horizon integration.

This paper presents a method for balancing supervised fine-tuning and reinforcement learning to train long-horizon advertising agents, demonstrating that targeted RL reduces data leakage and improves performance in enterprise analytics tasks.

This paper diagnoses the gap between next-turn evaluation metrics and autonomous workflow execution success for AI agents, showing that supervised fine-tuning improves turn-level performance but fails to enhance end-to-end workflow success.

This controlled study compares supervised fine-tuning and reinforcement learning methods for training tool-calling agents across different datasets and model scales, finding that SFT with LoRA is strongest in-distribution while RL shows slight advantages in cross-dataset transfer.

This paper introduces TrimSFT, a token-level reweighting method for supervised fine-tuning that enhances mathematical reasoning by focusing on intermediate logit-gap tokens, demonstrating significant performance gains over standard SFT.

NeoHorse-1 explores recursive self-improvement in AI models through agentic post-training with intelligent routing and feedback loops, demonstrating performance gains on various benchmarks.

This paper finds that large language models can gain reasoning improvements from truncated reasoning trajectories rather than full ones during post-training, reducing redundancy while benefiting methods like supervised fine-tuning and reinforcement learning.

Terminal-Universe is a framework that reconstructs executable workspaces from agent trajectories to synthesize diverse training tasks, improving AI agent post-training performance through supervised fine-tuning.

The paper introduces Elite-Weighted Supervised Fine-tuning (EW-SFT), a method for goal-directed molecular optimization that uses reward to guide elite selection and updates via the model's native loss, applicable across various generative architectures and tasks.

This paper presents a specialization pipeline for post-training language models to achieve gold-medal performance in coding competitions, demonstrating top scores on IOI benchmarks using techniques like supervised fine-tuning and reinforcement learning.

CogEvol is a family of models that efficiently generate structured learning artifacts like slides and interactive HTML pages in a single pass using supervised fine-tuning and reinforcement learning with vision-language rewards, reducing cost and improving reliability.

This paper investigates failures in a 2B model for dialogue games and introduces a diagnosis-guided post-training recipe using SFT, DPO, and LoRA to boost performance while maintaining general capabilities.

The paper identifies factual access failures in large language models after supervised fine-tuning and introduces Recall-Anchored Distillation (RAD) to preserve out-of-distribution factual recall without labeled data.