InstructGPT implements RLHF through several stages, including Supervised Fine-Tuning (SFT), reward model training, and Proximal Policy Optimization (PPO).

Obtaining human-interpretable explanations of large, general-purpose language models is an urgent goal for AI safety. However, it is just as important that our interpretability methods are faithful to the causal dynamics underlying model behavior and able to robustly generalize to unseen inputs. Distributed Alignment Search (DAS) is a powerful gradient descent method grounded in a theory of causal abstraction that uncovered perfect alignments between interpretable symbolic algorithms and small deep learning models fine-tuned for specific tasks. In the present paper, we scale DAS significantly by replacing the remaining brute-force search steps with learned parameters -- an approach we call Boundless DAS. This enables us to efficiently search for interpretable causal structure in large language models while they follow instructions. We apply Boundless DAS to the Alpaca model (7B parameters), which, off the shelf, solves a simple numerical reasoning problem. With Boundless DAS, we discover that Alpaca does this by implementing a causal model with two interpretable boolean variables. Furthermore, we find that the alignment of neural representations with these variables is robust to changes in inputs and instructions. These findings mark a first step toward deeply understanding the inner-workings of our largest and most widely deployed language models.

LLMs are often used by downstream users as teacher models for knowledge distillation, compressing their capabilities into memory-efficient models. However, as these teacher models may stem from untrusted parties, distillation can raise unexpected security risks. In this paper, we investigate the security implications of knowledge distillation from backdoored teacher models. First, we show that prior backdoors mostly do not transfer onto student models. Our key insight is that this is because existing LLM backdooring methods choose trigger tokens that rarely occur in usual contexts. We argue that this underestimates the security risks of knowledge distillation and introduce a new backdooring technique, T-MTB, that enables the construction and study of transferable backdoors. T-MTB carefully constructs a composite backdoor trigger, made up of several specific tokens that often occur individually in anticipated distillation datasets. As such, the poisoned teacher remains stealthy, while during distillation the individual presence of these tokens provides enough signal for the backdoor to transfer onto the student. Using T-MTB, we demonstrate and extensively study the security risks of transferable backdoors across two attack scenarios, jailbreaking and content modulation, and across four model families of LLMs.

Abstract--Efficient scheduling of large language model (LLM) inference tasks is essential for achieving low latency and high throughput, particularly with the growing use of reasoning-capable LLMs. Traditional strategies like First Come, First-Serve (FCFS) often suffer from Head-of-Line (HOL) blocking, where long-running tasks delay shorter ones queued behind them. In this paper, we introduce PARS, a prompt-aware LLM task scheduler that improves serving efficiency by approximating shortest-job-first (SJF) scheduling through pairwise ranking with margin ranking loss. PARS focuses on impactful scheduling decisions and seamlessly integrates into the state-of-the-art LLM serving system vLLM. It effectively predicts response-length-based task ordering, reducing latency with minimal overhead. Extensive experiments across multiple LLMs and real-world inference datasets show that PARS significantly improves performance, including for reasoning workloads. Furthermore, our cross-model evaluations demonstrate that the design generalizes well, enabling effective scheduling even when predictors are trained on different LLMs. Large language models (LLMs) have emerged as core engines for artificial intelligence applications, demonstrating remarkable capabilities in a wide range of tasks, including question answering, code generation, and text classification.

A common training approach for language models involves using a large-scale language model to expand a human-provided dataset, which is subsequently used for model training.This method significantly reduces training costs by eliminating the need for extensive human data annotation. However, it still faces challenges such as high carbon emissions during data augmentation and the risk of data leakage when we use closed-source LLMs. To address these issues, we propose a self-evolution method for language models. First, we introduce the Multi-level Principle Generation, which enables a large-scale model to summarize task-completion principles based on a small amount of task data. Then, we propose the Principle-based Instance Generation, in which a smaller-scale language model uses these task principles to generate a large amount of data. This data is then used for model training. Experimental results show that our proposed method significantly improves model performance compared to directly using a smaller-scale language model to generate data. Additionally, since we only use the large-scale language model to generate the task-completion principles, the carbon emissions associated with training the model are greatly reduced.

Instruction data selection aims to identify a high-quality subset from the training set that matches or exceeds the performance of the full dataset on target tasks. Existing methods focus on the instruction-to-response mapping, but neglect the human preference for diverse responses. In this paper, we propose Preference-oriented Data Selection method (ProDS) that scores training samples based on their alignment with preferences observed in the target set. Our key innovation lies in shifting the data selection criteria from merely estimating features for accurate response generation to explicitly aligning training samples with human preferences in target tasks. Specifically, direct preference optimization (DPO) is employed to estimate human preferences across diverse responses. Besides, a bidirectional preference synthesis strategy is designed to score training samples according to both positive preferences and negative preferences. Extensive experimental results demonstrate our superiority to existing task-agnostic and targeted methods.

The research in AI-based formal mathematical reasoning has shown an unstop- pable growth trend. These studies have excelled in mathematical competitions like IMO and have made significant progress. This paper focuses on formal verification, an immediate application scenario of formal reasoning, and breaks it down into sub-tasks. We constructed 18k high-quality instruction-response pairs across five formal specification languages (Coq, Lean4, Dafny, ACSL, and TLA+) by distilling gpt-4o and evaluated against ten open-sourced LLMs, including recent popular DeepSeek-R1. We also fine-tuned several 7~8B small models to achieve comparable performance with Deepseek-R1-671B. Interestingly, we observed that fine-tuning with formal data also enhances mathematics, reasoning, and coding capabilities. Fine-tuned models are released at https: //huggingface.co/fm-universe.