Language model (LM) evaluators that generate chain-of-thought (CoT) reasoning are widely used for the assessment of LM responses. Simultaneously, increasing LMs’ "thinking" time through scaling test-time compute has proven to be an effective technique for solving challenging problems in domains such as math and code. This raises a natural question: can an LM’s evaluation capability also be improved by scaling test-time compute? To answer this, we investigate employing reasoning models - LMs that natively generate long CoT reasoning - as evaluators. We explore scaling evaluation-time compute by using reasoning models to evaluate both the overall candidate response (i.e., outcome evaluation) and the individual reasoning steps within it (i.e., process evaluation). We observe that evaluator performance improves monotonically with the number of reasoning tokens generated, mirroring trends seen in LM reasoning. Furthermore, we use these more accurate evaluators to rerank multiple generations, and demonstrate that spending more compute at evaluation time can be as effective as increasing compute during generation for improving an LM’s problem-solving performance.

Multimodal Retrieval Augmented Generation (MMRAG) is a powerful approach to question-answering over multimodal documents. A key challenge with evaluating MMRAG is the paucity of high-quality datasets matching the question styles and modalities of interest. In light of this, we propose SMMQG, a synthetic data generation framework. SMMQG leverages interplay between a retriever, large language model (LLM) and large multimodal model (LMM) to generate question and answer pairs directly from multimodal documents, with the questions conforming to specified styles and modalities. We use SMMQG to generate an MMRAG dataset of 1024 questions over Wikipedia documents and evaluate state-of-the-art models using it, revealing insights into model performance that are attainable only through style- and modality-specific evaluation data. Next, we measure the quality of data produced by SMMQG via a human study. We find that the quality of SMMQG-generated synthetic data is on par with the quality of the crowdsourced benchmark MMQA and that downstream evaluation results using both datasets strongly concur.

A natural way to design a negotiation dialogue system is via self-play RL: train an agent that learns to maximize its performance by interacting with a simulated user that has been designed to imitate human-human dialogue data. Although this procedure has been adopted in prior work, we find that it results in a fundamentally flawed system that fails to learn the value of compromise in a negotiation, which can often lead to no agreements (i.e., the partner walking away without a deal), ultimately hurting the model’s overall performance. We investigate this observation in the context of DealOrNoDeal task, a multi-issue negotiation over books, hats, and balls. Grounded in negotiation theory from Economics, we modify the training procedure in two novel ways to design agents with diverse personalities and analyze their performance with human partners. We find that although both techniques show promise, a selfish agent, which maximizes its own performance while also avoiding walkaways, performs superior to other variants by implicitly learning to generate value for both itself and the negotiation partner. We discuss the implications of our findings for what it means to be a successful negotiation dialogue system and how these systems should be designed in the future.