Language model alignment has become an important component of AI safety, allowing safe interactions between humans and language models, by enhancing desired behaviors and inhibiting undesired ones. It is often done by tuning the model or inserting preset aligning prompts. Recently, representation engineering, a method which alters the model's behavior via changing its representations post-training, was shown to be effective in aligning LLMs (Zou et al., 2023a). Representation engineering yields gains in alignment oriented tasks such as resistance to adversarial attacks and reduction of social biases, but was also shown to cause a decrease in the ability of the model to perform basic tasks. In this paper we study the tradeoff between the increase in alignment and decrease in helpfulness of the model. We propose a theoretical framework which provides bounds for these two quantities, and demonstrate their relevance empirically. First, we find that under the conditions of our framework, alignment can be guaranteed with representation engineering, and at the same time that helpfulness is harmed in the process. Second, we show that helpfulness is harmed quadratically with the norm of the representation engineering vector, while the alignment increases linearly with it, indicating a regime in which it is efficient to use representation engineering. We validate our findings empirically, and chart the boundaries to the usefulness of representation engineering for alignment.
相關內容
Large language models have driven significant progress in natural language processing, but their deployment requires substantial compute and memory resources. As models scale, compression techniques become essential for balancing model quality with computational efficiency. Structured pruning, which removes less critical components of the model, is a promising strategy for reducing complexity. However, one-shot pruning often results in significant quality degradation, particularly in tasks requiring multi-step reasoning. To recover lost quality, supervised fine-tuning (SFT) is commonly applied, but it can lead to catastrophic forgetting by shifting the model's learned data distribution. Therefore, addressing the degradation from both pruning and SFT is essential to preserve the original model's quality. In this work, we utilize self-data distilled fine-tuning to address these challenges. Our approach leverages the original, unpruned model to generate a distilled dataset that preserves semantic richness and mitigates catastrophic forgetting by maintaining alignment with the base model's knowledge. Empirically, we demonstrate that self-data distillation consistently outperforms standard SFT, improving average accuracy by up to 8% on the HuggingFace OpenLLM Leaderboard v1. Specifically, when pruning six decoder blocks on Llama3.1-8B Instruct (i.e., 32 to 26 layers, reducing the model size from 8.03B to 6.72B parameters), our method retains 91.2% of the original model's accuracy compared to 81.7% with SFT, while reducing real-world FLOPs by 16.3%. Furthermore, combining self-data distilled models through model merging yields enhanced quality retention. Additionally, leveraging these pruned models in speculative decoding increases token acceptance rates, thereby improving inference efficiency in applied settings.
Relation extraction as an important natural Language processing (NLP) task is to identify relations between named entities in text. Recently, graph convolutional networks over dependency trees have been widely used to capture syntactic features and achieved attractive performance. However, most existing dependency-based approaches ignore the positive influence of the words outside the dependency trees, sometimes conveying rich and useful information on relation extraction. In this paper, we propose a novel model, Entity-aware Self-attention Contextualized GCN (ESC-GCN), which efficiently incorporates syntactic structure of input sentences and semantic context of sequences. To be specific, relative position self-attention obtains the overall semantic pairwise correlation related to word position, and contextualized graph convolutional networks capture rich intra-sentence dependencies between words by adequately pruning operations. Furthermore, entity-aware attention layer dynamically selects which token is more decisive to make final relation prediction. In this way, our proposed model not only reduces the noisy impact from dependency trees, but also obtains easily-ignored entity-related semantic representation. Extensive experiments on various tasks demonstrate that our model achieves encouraging performance as compared to existing dependency-based and sequence-based models. Specially, our model excels in extracting relations between entities of long sentences.
We study an auto-calibration problem in which a transform-sparse signal is acquired via compressive sensing by multiple sensors in parallel, but with unknown calibration parameters of the sensors. This inverse problem has an important application in pMRI reconstruction, where the calibration parameters of the receiver coils are often difficult and costly to obtain explicitly, but nonetheless are a fundamental requirement for high-precision reconstructions. Most auto-calibration strategies for this problem involve solving a challenging biconvex optimization problem, which lacks reconstruction guarantees. In this work, we transform the auto-calibrated parallel compressive sensing problem to a convex optimization problem using the idea of `lifting'. By exploiting sparsity structures in the signal and the redundancy introduced by multiple sensors, we solve a mixed-norm minimization problem to recover the underlying signal and the sensing parameters simultaneously. Our method provides robust and stable recovery guarantees that take into account the presence of noise and sparsity deficiencies in the signals. As such, it offers a theoretically guaranteed approach to auto-calibrated parallel imaging in MRI under appropriate assumptions. Applications in compressive sensing pMRI are discussed, and numerical experiments using real and simulated MRI data are presented to support our theoretical results.
Large language models (LLMs) have shown promise in safety-critical applications such as healthcare, yet the ability to quantify performance has lagged. An example of this challenge is in evaluating a summary of the patient's medical record. A resulting summary can enable the provider to get a high-level overview of the patient's health status quickly. Yet, a summary that omits important facts about the patient's record can produce a misleading picture. This can lead to negative consequences on medical decision-making. We propose MED-OMIT as a metric to explore this challenge. We focus on using provider-patient history conversations to generate a subjective (a summary of the patient's history) as a case study. We begin by discretizing facts from the dialogue and identifying which are omitted from the subjective. To determine which facts are clinically relevant, we measure the importance of each fact to a simulated differential diagnosis. We compare MED-OMIT's performance to that of clinical experts and find broad agreement We use MED-OMIT to evaluate LLM performance on subjective generation and find some LLMs (gpt-4 and llama-3.1-405b) work well with little effort, while others (e.g. Llama 2) perform worse.
Recent models for natural language understanding are inclined to exploit simple patterns in datasets, commonly known as shortcuts. These shortcuts hinge on spurious correlations between labels and latent features existing in the training data. At inference time, shortcut-dependent models are likely to generate erroneous predictions under distribution shifts, particularly when some latent features are no longer correlated with the labels. To avoid this, previous studies have trained models to eliminate the reliance on shortcuts. In this study, we explore a different direction: pessimistically aggregating the predictions of a mixture-of-experts, assuming each expert captures relatively different latent features. The experimental results demonstrate that our post-hoc control over the experts significantly enhances the model's robustness to the distribution shift in shortcuts. Besides, we show that our approach has some practical advantages. We also analyze our model and provide results to support the assumption.
Recent work in the construction of 3D scene graphs has enabled mobile robots to build large-scale metric-semantic hierarchical representations of the world. These detailed models contain information that is useful for planning, however an open question is how to derive a planning domain from a 3D scene graph that enables efficient computation of executable plans. In this work, we present a novel approach for defining and solving Task and Motion Planning problems in large-scale environments using hierarchical 3D scene graphs. We describe a method for building sparse problem instances which enables scaling planning to large scenes, and we propose a technique for incrementally adding objects to that domain during planning time that minimizes computation on irrelevant elements of the scene graph. We evaluate our approach in two real scene graphs built from perception, including one constructed from the KITTI dataset. Furthermore, we demonstrate our approach in the real world, building our representation, planning in it, and executing those plans on a real robotic mobile manipulator. A video supplement is available at \url{//youtu.be/v8fkwLjBn58}.
Understanding how information can efficiently spread in distributed systems under noisy communications is a fundamental question in both biological research and artificial system design. When agents are able to control whom they interact with, noise can often be mitigated through redundancy or other coding techniques, but it may have fundamentally different consequences on well-mixed systems. Specifically, Boczkowski et al. (2018) considered the noisy $\mathcal{PULL}(h)$ model, where each message can be viewed as any other message with probability $\delta$. The authors proved that in this model, the basic task of propagating a bit value from a single source to the whole population requires $\Omega(\frac{n\delta}{h(1-\delta|\Sigma|)^2})$ (parallel) rounds. The current work shows that the aforementioned lower bound is almost tight. In particular, when each agent observes all other agents in each round, which relates to scenarios where each agent senses the system's average tendency, information spreading can reliably be achieved in $\mathcal{O}(\log n)$ time, assuming constant noise. We present two simple and highly efficient protocols, thus suggesting their applicability to real-life scenarios. Notably, they also work in the presence of multiple conflicting sources and efficiently converge to their plurality opinion. The first protocol we present uses 1-bit messages but relies on a simultaneous wake-up assumption. By increasing the message size to 2 bits and removing the speedup in the information spreading time that may result from having multiple sources, we also present a simple and highly efficient self-stabilizing protocol that avoids the simultaneous wake-up requirement. Overall, our results demonstrate how, under stochastic communication, increasing the sample size can compensate for the lack of communication structure by linearly accelerating information spreading time.
We propose a data-driven pressure distribution rendering method that uses the interpolation of experimentally obtained pressure values. The pressure data were collected using a pressure sensor array. The prediction was performed using linear interpolation, assuming that the pressure distribution is dependent on pushing displacement and contact angle. Leap Motion Controller was used to implement the prediction based on user input. The proposed prediction model was found to be fast and reproduce the measured data well.
The high-performance computing (HPC) community has recently seen a substantial diversification of hardware platforms and their associated programming models. From traditional multicore processors to highly specialized accelerators, vendors and tool developers back up the relentless progress of those architectures. In the context of scientific programming, it is fundamental to consider performance portability frameworks, i.e., software tools that allow programmers to write code once and run it on different computer architectures without sacrificing performance. We report here on the benefits and challenges of performance portability using a field-line tracing simulation and a particle-in-cell code, two relevant applications in computational plasma physics with applications to magnetically-confined nuclear-fusion energy research. For these applications we report performance results obtained on four HPC platforms with server-class CPUs from Intel (Xeon) and AMD (EPYC), and high-end GPUs from Nvidia and AMD, including the latest Nvidia H100 GPU and the novel AMD Instinct MI300A APU. Our results show that both Kokkos and OpenMP are powerful tools to achieve performance portability and decent "out-of-the-box" performance, even for the very latest hardware platforms. For our applications, Kokkos provided performance portability to the broadest range of hardware architectures from different vendors.
The success of AI models relies on the availability of large, diverse, and high-quality datasets, which can be challenging to obtain due to data scarcity, privacy concerns, and high costs. Synthetic data has emerged as a promising solution by generating artificial data that mimics real-world patterns. This paper provides an overview of synthetic data research, discussing its applications, challenges, and future directions. We present empirical evidence from prior art to demonstrate its effectiveness and highlight the importance of ensuring its factuality, fidelity, and unbiasedness. We emphasize the need for responsible use of synthetic data to build more powerful, inclusive, and trustworthy language models.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191