An official website of the United States government
Diffusion-based language models are emerging as a promising alternative to autoregressive LMs: they approach the competence of autoregressive LMs while offering nuanced controllability at inference time. While autoregressive LMs have benefited immensely from scaling and instruction-based learning, existing studies of diffusion LMs have been conducted on a smaller scale. Starting with a recently proposed diffusion model SSD-LM, in this work we first explore methods to scale it from 0.4B to 13B parameters, proposing techniques to improve its training and inference efficiency, and to finetune the model to follow instructions. Armed with a more powerful, general purpose diffusion LM, we introduce the primary contribution of this work – SSD-2 – an approach to easily ensemble at inference time a large general-purpose diffusion LM with smaller, but specialized and contextualized diffusion LMs. We show that SSD-2 facilitates novel ensembles with 100x smaller models that can be customized and deployed by individual users. We find that compared to autoregressive models, the collaboration between diffusion LMs is more effective, leading to higher-quality model responses due to their ability to dynamically incorporate bi-directional contexts.
more »
« less
This content will become publicly available on April 27, 2026
IHEval: Evaluating Language Models on Following the Instruction Hierarchy
The instruction hierarchy, which establishes a priority order from system messages to user messages, conversation history, and tool outputs, is essential for ensuring consistent and safe behavior in language models (LMs). Despite its importance, this topic receives limited attention, and there is a lack of comprehensive benchmarks for evaluating models’ ability to follow the instruction hierarchy. We bridge this gap by introducing IHEval, a novel benchmark comprising 3,538 examples across nine tasks, covering cases where instructions in different priorities either align or conflict. Our evaluation of popular LMs highlights their struggle to recognize instruction priorities. All evaluated models experience a sharp performance decline when facing conflicting instructions, compared to their original instruction-following performance. Moreover, the most competitive open-source model only achieves 48% accuracy in resolving such conflicts. Our results underscore the need for targeted optimization in the future development of LMs.
more »
« less
- Award ID(s):
- 2137396
- PAR ID:
- 10590839
- Editor(s):
- Chiruzzo, Luis; Ritter, Alan; Wang, Lu
- Publisher / Repository:
- Association for Computational Linguistics
- Date Published:
- ISBN:
- 979-8-89176-189-6
- Format(s):
- Medium: X
- Location:
- Albuquerque, New Mexico
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Instruction fine-tuning has recently emerged as a promising approach for improving the zero-shot capabilities of Large Language Models (LLMs) on new tasks. This technique has shown particular strength in improving the performance of modestly sized LLMs, sometimes inducing performance competitive with much larger model variants. In this paper, we ask two questions: (1) How sensitive are instruction-tuned models to the particular phrasings of instructions, and, (2) How can we make them more robust to such natural language variation? To answer the former, we collect a set of 319 instructions manually written by NLP practitioners for over 80 unique tasks included in widely used benchmarks, and we evaluate the variance and average performance of these instructions as compared to instruction phrasings observed during instruction fine-tuning. We find that using novel (unobserved) but appropriate instruction phrasings consistently degrades model performance, sometimes substantially so. Further, such natural instructions yield a wide variance in downstream performance, despite their semantic equivalence. Put another way, instruction-tuned models are not especially robust to instruction re-phrasings. We propose a simple method to mitigate this issue by introducing soft prompt'' embedding parameters and optimizing these to maximize the similarity between representations of semantically equivalent instructions. We show that this method consistently improves the robustness of instruction-tuned models.more » « less
-
Embeddings for instructions have been shown to be essential for software reverse engineering and automated program analysis. However, due to the complexity of dependencies and inherent variability of instructions, instruction embeddings using models that are successful for natural language processing may not be effective. In this paper, we perform geometric analysis of instruction embeddings at the token level and instruction family level, showing much greater variability and leading to degraded performance on intrinsic analyses. Then we propose to use metric learning to improve the relationships among instructions using triplet loss. Our results on a large dataset of instruction groups shows significant improvements. We also provide a theoretical analysis of the instruction embeddings by looking at the BERT components and characteristics of inner-product matrices for attention in the transformer blocks. The code will be available publicly after the paper is accepted for publication.more » « less
-
Even though heavily researched, a full formal model of the x86-64 instruction set is still not available. We present libLISA, a tool for automated discovery and analysis of the ISA of a CPU. This produces the most extensive formal x86-64 model to date, with over 118000 different instruction groups. The process requires as little human specification as possible: specifically, we do not rely on a human-written (dis)assembler to dictate which instructions are executable on a given CPU, or what their in- and outputs are. The generated model is CPU-specific: behavior that is undefined is synthesized for the current machine. Producing models for five different x86-64 machines, we mutually compare them, discover undocumented instructions, and generate instruction sequences that are CPU-specific. Experimental evaluation shows that we enumerate virtually all instructions within scope, that the instructions' semantics are correct w.r.t. existing work, and that we improve existing work by exposing bugs in their handwritten models.more » « less
-
Reinforcement learning has seen wide success in finetuning large language models to better align with instructions via human feedback. The so-called algorithm, Reinforcement Learning with Human Feedback (RLHF) demonstrates impressive performance on the GPT series models. However, the underlying reinforcement learning algorithm is complex and requires additional training for reward and value networks. In this paper, we consider an alternative approach: converting feedback to instruction by relabeling the original one and training the model for better alignment in a supervised manner. Such an algorithm doesn’t require any additional parameters except for the original language model and maximally reuses the pretraining pipeline. To achieve this, we formulate instruction alignment problem for language models as a goal-reaching problem in decision making. We propose Hindsight Instruction Relabeling (HIR), a novel algorithm for aligning language models with instructions. The resulting two-stage algorithm shed light to a family of reward-free approaches that utilize the hindsightly relabeled instructions based on feedback. We evaluate the performance of HIR extensively on 12 challenging BigBench reasoning tasks and show that HIR outperforms the baseline algorithms and is comparable to or even surpasses supervised fine-tuning. The implementation of HIR is available at https://github.com/tianjunz/HIR.more » « less
