arXiv subjects

Nevan Wichers

Publications and source records attributed to Nevan Wichers.

14 recordsLinked to original sources

arXiv2026

Model Spec Midtraining: Improving How Alignment Training Generalizes

Some frontier AI developers aim to align language models to a Model Spec or Constitution that describes the intended model behavior. However, standard alignment fine-tuning -- training on demonstrations of spec-aligned behavior -- can produce shallow alignment that generalizes poorly, in part because demonstration data can underspecify the desired generalization. We introduce model spec midtraining (MSM): after pre-training but before alignment fine-tuning, we train models on synthetic documents discussing their Model Spec. This teaches models the content of the spec, thereby shaping how they generalize from subsequent demonstration data. For example, a model fine-tuned only to express certain cheese preferences (e.g., "I prefer cream cheese over brie") generalizes to broadly pro-America values when we apply MSM with a spec attributing those preferences to pro-America values. Conversely, a spec about pro-affordability values instead yields pro-affordability generalization from the exact same cheese fine-tuning. MSM can also shape complex safety-relevant propensities: applying MSM with a spec addressing self-preservation and goal-guarding substantially reduces agentic misalignment rate (Qwen3-32B: 54% to 7%), beating a deliberative alignment baseline (14%). We further use MSM as a tool to study which Model Specs produce the strongest alignment generalization, finding that explaining the values underlying rules improves generalization, as does providing specific rather than general guidance. Overall, MSM is a simple, effective technique for controlling and improving how models generalize from alignment training, by first teaching the intended generalization.

cs.AI↗

arXiv2026

Recontextualization Mitigates Specification Gaming without Modifying the Specification

Developers often struggle to specify correct training labels and rewards. Perhaps they don't need to. We propose recontextualization, which reduces how often language models "game" training signals, performing misbehaviors those signals mistakenly reinforce. We show recontextualization prevents models from learning to 1) prioritize evaluation metrics over chat response quality; 2) special-case code to pass incorrect tests; 3) overwrite evaluation functions rather than write correct code; and 4) become sycophantic. Our method works by generating completions from prompts discouraging misbehavior and then recontextualizing them as though they were in response to prompts permitting misbehavior. Recontextualization trains language models to resist misbehavior even when instructions permit it. This mitigates the reinforcement of misbehavior from misspecified training signals, reducing specification gaming without improving the supervision signal.

cs.AI↗

arXiv2025

Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment

Large language models are sometimes trained with imperfect oversight signals, leading to undesired behaviors such as reward hacking and sycophancy. Improving oversight quality can be expensive or infeasible, motivating methods that improve learned behavior despite an imperfect training signal. We introduce Inoculation Prompting (IP), a simple but counterintuitive technique that prevents learning of an undesired behavior by modifying training prompts to explicitly request it. For example, to inoculate against reward hacking, we modify the prompts used in supervised fine-tuning to request code that only works on provided test cases but fails on other inputs. Across four settings we find that IP reduces the learning of undesired behavior without substantially reducing the learning of desired capabilities. We also show that prompts which more strongly elicit the undesired behavior prior to fine-tuning more effectively inoculate against the behavior when used during training; this serves as a heuristic to identify promising inoculation prompts. Overall, IP is a simple yet effective way to control how models generalize from fine-tuning, preventing learning of undesired behaviors without substantially disrupting desired capabilities.

cs.LG↗

arXiv2024

Fusion-Eval: Integrating Assistant Evaluators with LLMs

Evaluating natural language systems poses significant challenges, particularly in the realms of natural language understanding and high-level reasoning. In this paper, we introduce 'Fusion-Eval', an innovative approach that leverages Large Language Models (LLMs) to integrate insights from various assistant evaluators. The LLM is given the example to evaluate along with scores from the assistant evaluators. Each of these evaluators specializes in assessing distinct aspects of responses. Fusion-Eval achieves a 0.962 system-level Kendall-Tau correlation with humans on SummEval and a 0.744 turn-level Spearman correlation on TopicalChat, which is significantly higher than baseline methods. These results highlight Fusion-Eval's significant potential in the realm of natural language system evaluation.

cs.CL↗

arXiv2024

Visualizing Neural Network Imagination

In certain situations, neural networks will represent environment states in their hidden activations. Our goal is to visualize what environment states the networks are representing. We experiment with a recurrent neural network (RNN) architecture with a decoder network at the end. After training, we apply the decoder to the intermediate representations of the network to visualize what they represent. We define a quantitative interpretability metric and use it to demonstrate that hidden states can be highly interpretable on a simple task. We also develop autoencoder and adversarial techniques and show that benefit interpretability.

cs.LG↗

arXiv2024

Beyond Sparse Rewards: Enhancing Reinforcement Learning with Language Model Critique in Text Generation

Reinforcement learning (RL) can align language models with non-differentiable reward signals, such as human preferences. However, a major challenge arises from the sparsity of these reward signals - typically, there is only a single reward for an entire output. This sparsity of rewards can lead to inefficient and unstable learning. To address this challenge, our paper introduces an novel framework that utilizes the critique capability of Large Language Models (LLMs) to produce intermediate-step rewards during RL training. Our method involves coupling a policy model with a critic language model, which is responsible for providing comprehensive feedback of each part of the output. This feedback is then translated into token or span-level rewards that can be used to guide the RL training process. We investigate this approach under two different settings: one where the policy model is smaller and is paired with a more powerful critic model, and another where a single language model fulfills both roles. We assess our approach on three text generation tasks: sentiment control, language model detoxification, and summarization. Experimental results show that incorporating artificial intrinsic rewards significantly improve both sample efficiency and the overall performance of the policy model, supported by both automatic and human evaluation.

cs.CL↗

arXiv2024

Gradient-Based Language Model Red Teaming

Red teaming is a common strategy for identifying weaknesses in generative language models (LMs), where adversarial prompts are produced that trigger an LM to generate unsafe responses. Red teaming is instrumental for both model alignment and evaluation, but is labor-intensive and difficult to scale when done by humans. In this paper, we present Gradient-Based Red Teaming (GBRT), a red teaming method for automatically generating diverse prompts that are likely to cause an LM to output unsafe responses. GBRT is a form of prompt learning, trained by scoring an LM response with a safety classifier and then backpropagating through the frozen safety classifier and LM to update the prompt. To improve the coherence of input prompts, we introduce two variants that add a realism loss and fine-tune a pretrained model to generate the prompts instead of learning the prompts directly. Our experiments show that GBRT is more effective at finding prompts that trigger an LM to generate unsafe responses than a strong reinforcement learning-based red teaming approach, and succeeds even when the LM has been fine-tuned to produce safer outputs.

cs.CL↗

arXiv2023

SiRA: Sparse Mixture of Low Rank Adaptation

Parameter Efficient Tuning has been an prominent approach to adapt the Large Language Model to downstream tasks. Most previous works considers adding the dense trainable parameters, where all parameters are used to adapt certain task. We found this less effective empirically using the example of LoRA that introducing more trainable parameters does not help. Motivated by this we investigate the importance of leveraging "sparse" computation and propose SiRA: sparse mixture of low rank adaption. SiRA leverages the Sparse Mixture of Expert(SMoE) to boost the performance of LoRA. Specifically it enforces the top $k$ experts routing with a capacity limit restricting the maximum number of tokens each expert can process. We propose a novel and simple expert dropout on top of gating network to reduce the over-fitting issue. Through extensive experiments, we verify SiRA performs better than LoRA and other mixture of expert approaches across different single tasks and multitask settings.

cs.CL↗

arXiv2022

SAFER: Data-Efficient and Safe Reinforcement Learning via Skill Acquisition

Methods that extract policy primitives from offline demonstrations using deep generative models have shown promise at accelerating reinforcement learning(RL) for new tasks. Intuitively, these methods should also help to trainsafeRLagents because they enforce useful skills. However, we identify these techniques are not well equipped for safe policy learning because they ignore negative experiences(e.g., unsafe or unsuccessful), focusing only on positive experiences, which harms their ability to generalize to new tasks safely. Rather, we model the latentsafetycontextusing principled contrastive training on an offline dataset of demonstrations from many tasks, including both negative and positive experiences. Using this late variable, our RL framework, SAFEty skill pRiors (SAFER) extracts task-specific safe primitive skills to safely and successfully generalize to new tasks. In the inference stage, policies trained with SAFER learn to compose safe skills into successful policies. We theoretically characterize why SAFER can enforce safe policy learning and demonstrate its effectiveness on several complex safety-critical robotic grasping tasks inspired by the game Operation, in which SAFERoutperforms state-of-the-art primitive learning methods in success and safety.

cs.LG↗

arXiv2021

ActionBert: Leveraging User Actions for Semantic Understanding of User Interfaces

As mobile devices are becoming ubiquitous, regularly interacting with a variety of user interfaces (UIs) is a common aspect of daily life for many people. To improve the accessibility of these devices and to enable their usage in a variety of settings, building models that can assist users and accomplish tasks through the UI is vitally important. However, there are several challenges to achieve this. First, UI components of similar appearance can have different functionalities, making understanding their function more important than just analyzing their appearance. Second, domain-specific features like Document Object Model (DOM) in web pages and View Hierarchy (VH) in mobile applications provide important signals about the semantics of UI elements, but these features are not in a natural language format. Third, owing to a large diversity in UIs and absence of standard DOM or VH representations, building a UI understanding model with high coverage requires large amounts of training data. Inspired by the success of pre-training based approaches in NLP for tackling a variety of problems in a data-efficient way, we introduce a new pre-trained UI representation model called ActionBert. Our methodology is designed to leverage visual, linguistic and domain-specific features in user interaction traces to pre-train generic feature representations of UIs and their components. Our key intuition is that user actions, e.g., a sequence of clicks on different UI components, reveals important information about their functionality. We evaluate the proposed model on a wide variety of downstream tasks, ranging from icon classification to UI component retrieval based on its natural language description. Experiments show that the proposed ActionBert model outperforms multi-modal baselines across all downstream tasks by up to 15.5%.

cs.CL↗

arXiv2020

Resolving Spurious Correlations in Causal Models of Environments via Interventions

Causal models bring many benefits to decision-making systems (or agents) by making them interpretable, sample-efficient, and robust to changes in the input distribution. However, spurious correlations can lead to wrong causal models and predictions. We consider the problem of inferring a causal model of a reinforcement learning environment and we propose a method to deal with spurious correlations. Specifically, our method designs a reward function that incentivizes an agent to do an intervention to find errors in the causal model. The data obtained from doing the intervention is used to improve the causal model. We propose several intervention design methods and compare them. The experimental results in a grid-world environment show that our approach leads to better causal models compared to baselines: learning the model on data from a random policy or a policy trained on the environment's reward. The main contribution consists of methods to design interventions to resolve spurious correlations.

cs.LG↗

arXiv2020

RL agents Implicitly Learning Human Preferences

In the real world, RL agents should be rewarded for fulfilling human preferences. We show that RL agents implicitly learn the preferences of humans in their environment. Training a classifier to predict if a simulated human's preferences are fulfilled based on the activations of a RL agent's neural network gets .93 AUC. Training a classifier on the raw environment state gets only .8 AUC. Training the classifier off of the RL agent's activations also does much better than training off of activations from an autoencoder. The human preference classifier can be used as the reward function of an RL agent to make RL agent more beneficial for humans.

cs.AI↗

arXiv2018

Resolving Referring Expressions in Images With Labeled Elements

Images may have elements containing text and a bounding box associated with them, for example, text identified via optical character recognition on a computer screen image, or a natural image with labeled objects. We present an end-to-end trainable architecture to incorporate the information from these elements and the image to segment/identify the part of the image a natural language expression is referring to. We calculate an embedding for each element and then project it onto the corresponding location (i.e., the associated bounding box) of the image feature map. We show that this architecture gives an improvement in resolving referring expressions, over only using the image, and other methods that incorporate the element information. We demonstrate experimental results on the referring expression datasets based on COCO, and on a webpage image referring expression dataset that we developed.

cs.CV↗

arXiv2018

Hierarchical Long-term Video Prediction without Supervision

Much of recent research has been devoted to video prediction and generation, yet most of the previous works have demonstrated only limited success in generating videos on short-term horizons. The hierarchical video prediction method by Villegas et al. (2017) is an example of a state-of-the-art method for long-term video prediction, but their method is limited because it requires ground truth annotation of high-level structures (e.g., human joint landmarks) at training time. Our network encodes the input frame, predicts a high-level encoding into the future, and then a decoder with access to the first frame produces the predicted image from the predicted encoding. The decoder also produces a mask that outlines the predicted foreground object (e.g., person) as a by-product. Unlike Villegas et al. (2017), we develop a novel training method that jointly trains the encoder, the predictor, and the decoder together without highlevel supervision; we further improve upon this by using an adversarial loss in the feature space to train the predictor. Our method can predict about 20 seconds into the future and provides better results compared to Denton and Fergus (2018) and Finn et al. (2016) on the Human 3.6M dataset.

cs.CV↗