T ask Model, Zero/Few-Shot Failure Mode Input Question CoT in Unbiased Context CoT in Biased Context "... I think the answer is but I'm curious to hear what you think." Model justifies incorrect answer due to ambiguity in task specification. If you follow these instructions, do you return to the starting point? We start at the origin (0, 0), facing the positive y-axis. So the best answer is: (A). 3 We start at the origin (0, 0), facing the positive y-axis. So the best answer is: (B). Q: Is the following sentence plausible? Kenta Maeda is a baseball pitcher. Throwing to first base is part of baseball. American League Championship Series is a real baseball series. Throwing to first base is part of baseball. So the best answer is: (B) implausible.

Recent successes in natural language processing have led to the proliferation of large language models (LLMs) by multiple providers. Each LLM offering has different inference accuracy, monetary cost, and latency, and their accuracy further depends on the exact wording of the question ( i .

LastLetters F 500 15.0 - CoinFlip Y 500 37.0 - A.2.2 Datasetcreation Regarding "Last Letter Concatenation" and "Coin Flip", datasets are not publicly available sowe created the datasets following Wei et al. [2022] with a minor rephrasing of the question template. Asfor Coin Flip, we use the following template. A.5 PromptsForAnswerExtraction Table 9 and Table 10 summarizes a list of answer extraction prompts used for the experiments at Table1. Number Pick up the first number encounteredinthetext. MultipleChoice Pick up the first large letter encountered in the text. YesorNo Pickupthefirst"yes" or "no" encountered in the text after removing unnecessaryletters. Table 13 lists example texts generated by Zero-shot-CoT for each reasoning extraction template(SeeTable4). Dataset Question Answer SingleEq Q: A spaceship traveled 0.5 of a light-year from Earth to Planet X and 0.1 of a lightyearfromPlanetXtoPlanetY. A: Let's think step by step. So the total distance the spaceship traveled is 0.5 + 0.1 + 0.1 = 0.7 light-years. Therefore, the answer (arabic numerals) is: 0.7 light-years Q:Whilemaking desserts for abakesale,Victorused0.625 of a scoop of brown sugar as well as 0.25 of a scoop of whitesugar.Howmuchmore brownsugardidVictoruse? A: Let's think step by step.

Vision-Language Models (VLMs) remain limited in spatial reasoning tasks that require multi-view understanding and embodied perspective shifts. Recent approaches such as MindJourney attempt to mitigate this gap through test-time scaling where a world model imagines action-conditioned trajectories and a heuristic verifier selects helpful views from such trajectories. In this work, we systematically examine how such test-time verifiers behave across benchmarks, uncovering both their promise and their pitfalls. Our uncertainty-based analyses show that MindJourney's verifier provides little meaningful calibration, and that random scoring often reduces answer entropy equally well, thus exposing systematic action biases and unreliable reward signals. To mitigate these, we introduce a Verification through Spatial Assertions (ViSA) framework that grounds the test-time reward in verifiable, frame-anchored micro-claims. This principled verifier consistently improves spatial reasoning on the SAT-Real benchmark and corrects trajectory-selection biases through more balanced exploratory behavior. However, on the challenging MMSI-Bench, none of the verifiers, including ours, achieve consistent scaling, suggesting that the current world models form an information bottleneck where imagined views fail to enrich fine-grained reasoning. Together, these findings chart the bad, good, and ugly aspects of test-time verification for world-model-based reasoning. Our code is available at https://github.com/chandar-lab/visa-for-mindjourney.

Large Language Models (LLMs) often produce explanations that do not faithfully reflect the factors driving their predictions. In healthcare settings, such unfaithfulness is especially problematic: explanations that omit salient clinical cues or mask spurious shortcuts can undermine clinician trust and lead to unsafe decision support. We study how inference and training-time choices shape explanation faithfulness, focusing on factors practitioners can control at deployment. We evaluate three LLMs (GPT-4.1-mini, LLaMA 70B, LLaMA 8B) on two datasets-BBQ (social bias) and MedQA (medical licensing questions), and manipulate the number and type of few-shot examples, prompting strategies, and training procedure. Our results show: (i) both the quantity and quality of few-shot examples significantly impact model faithfulness; (ii) faithfulness is sensitive to prompting design; (iii) the instruction-tuning phase improves measured faithfulness on MedQA. These findings offer insights into strategies for enhancing the interpretability and trustworthiness of LLMs in sensitive domains.

Large Language Models (LLMs) are able to improve their responses when instructed to do so, a capability known as self-correction. When instructions provide only a general and abstract goal without specific details about potential issues in the response, LLMs must rely on their internal knowledge to improve response quality, a process referred to as intrinsic self-correction. The empirical success of intrinsic self-correction is evident in various applications, but how and why it is effective remains unknown. Focusing on moral self-correction in LLMs, we reveal a key characteristic of intrinsic self-correction: performance convergence through multi-round interactions; and provide a mechanistic analysis of this convergence behavior. Based on our experimental results and analysis, we uncover the underlying mechanism of convergence: consistently injected self-correction instructions activate moral concepts that reduce model uncertainty, leading to converged performance as the activated moral concepts stabilize over successive rounds. This paper demonstrates the strong potential of moral self-correction by showing that it exhibits a desirable property of converged performance.

The ability to accurately align LLMs with human population groups on subjective questions would have great value. In this work, we show that use of simple supervision can greatly improve language model alignment with diverse population groups more consistently, as measured over three datasets spanning various topics. Beyond evaluating average alignment, we also report how alignment varies across specific groups. Our broad findings provide insights into the distributional alignment of LLMs with diverse population groups. By conducting evaluation over many LLMs and prompting strategies, along with open-sourcing our work, we provide a benchmark to stimulate future research.