Large reasoning models exhibit long chain-of-thought reasoning with complex strategies such as backtracking and self-verification. Yet, these capabilities typically require resource-intensive post-training. We investigate whether such behaviors can be elicited in large models without any gradient updates. To this end, we propose a decoding-time approach, ThinkLogit, which utilizes logit arithmetic to transfer these capabilities from a substantially smaller reasoning guider to a large non-reasoning target. We further show that we can boost performance by training the guider to correct the target’s errors using preference optimization over mixed model outputs, a setup we refer to as ThinkLogit-DPO. We evaluate these methods across six reasoning benchmarks spanning math, science, and coding domains using the Qwen2.5-32B guided by R1-Distill-Qwen-1.5B, a model 21x smaller. Our experiments demonstrate that ThinkLogit and ThinkLogit-DPO achieve a relative improvement of 21.5% and 24.2%, respectively, over the target model. Moreover, ThinkLogit remains effective even when the guider and target come from different model families.Crucially, our method requires zero training for the large model and would incur minimal inference overhead when logits are computed in parallel, presenting a practical solution for enabling long reasoning at scale.

Tool-augmented Language Models (TaLMs) can invoke external tools to solve problems beyond their parametric capacity. However, it remains unclear whether these tool-enabled gains reflect trustworthy reasoning. Focusing on the Code Interpreter tool, we show that even when tools are selected and executed correctly, TaLMs treat tool outputs as substitutes for reasoning, producing solutions that appear correct but lack coherent justification. We term this failure mode Tool-Induced Myopia (TIM), and study it using PyMath, a benchmark of 1,679 competition-level mathematical problems for which Python code is helpful but not sufficient. We further develop a multi-dimensional evaluation suite to quantify reasoning degradation in TaLMs relative to their non-tool counterparts. Our findings reveal that while TaLMs achieve up to a 19.3 percentage point gain in final-answer accuracy, their reasoning behavior consistently deteriorates (e.g., non-tool language models win up to 41.5% more often in pairwise comparisons of reasoning processes). This degradation intensifies with tool use; the more frequently a model invokes tools, the less coherent its reasoning becomes. Moreover, tool use shifts errors from arithmetic mistakes toward global reasoning failures (logic, assumption, creativity). Finally, we propose a preference-optimization-based framework that realigns TaLMs to use tool outputs as assistive evidence, improving both final-answer accuracy and reasoning depth under tool use. Code and data will be released upon publication.

The rapid adoption of language models (LMs) across diverse applications has raised concerns about their factuality, i.e., their consistency with real-world facts. We introduce VERIFY, an evidence-based evaluation pipeline that measures LMs’ factuality in real-world user interactions. VERIFY considers the verifiability of LM-generated content and categorizes content units as Supported, Unsupported, or Undecidable based on Web-retrieved evidence. Importantly, factuality judgment by VERIFY more strongly correlates with human evaluations than existing methods. Using VERIFY, we identify “hallucination prompts,” i.e., those that frequently elicit factual errors in LM responses. These prompts form FactBench, a dataset of 1K prompts spanning 150 topics and tiered into Easy, Moderate, and Hard prompts. We benchmark widely-used openweight and proprietary LMs from six families, yielding three key findings: (i) LMs’ factual precision declines from Easy to Hard prompts, (ii) factuality does not necessarily improve with scale; Llama3.1-405B-Instruct performs comparably to or worse than its 70B variant, and (iii) Gemini1.5-Pro shows a notably higher refusal rate, with over-refusal in 25% of cases.

Calibrating language models (LMs) aligns their generation confidence with the actual likelihood of answer correctness, which can inform users about LMs’ reliability and mitigate hallucinated content. However, prior calibration methods, such as self-consistency-based and logit-based approaches, are either limited in inference-time efficiency or fall short of providing informative signals. Moreover, simply filtering out low-confidence responses reduces the LM’s helpfulness when the answers are correct. Therefore, effectively using calibration techniques to enhance an LM’s factuality remains an unsolved challenge. In this paper, we first propose an activation-based calibration method, ActCab, which trains a linear layer on top of the LM’s last-layer activations that can better capture the representations of knowledge. Built on top of ActCab, we further propose CoDec, a confidence-guided decoding strategy to elicit truthful answers with high confidence from LMs. By evaluating on five popular QA benchmarks, ActCab achieves superior calibration performance than all competitive baselines, e.g., by reducing the average expected calibration error (ECE) score by up to 39%. Further experiments on CoDec show consistent improvements in several LMs’ factuality on challenging QA datasets, such as TruthfulQA, highlighting the value of confidence signals in enhancing the factuality.

Large language models (LLMs) can generate long-form and coherent text, yet they often hallucinate facts, which undermines their reliability. To mitigate this issue, inference-time methods steer LLM representations toward the “truthful directions” previously learned for truth elicitation. However, applying these truthful directions with the same intensity fails to generalize across different query contexts. We propose LITO, a Learnable Intervention method for Truthfulness Optimization that automatically identifies the optimal intervention intensity tailored to each specific context. LITO explores a sequence of model generations based on increasing levels of intervention intensities. It selects the most accurate response or refuses to answer when the predictions are highly uncertain. Experiments on multiple LLMs and question-answering datasets demonstrate that LITO improves truthfulness while preserving task accuracy. The adaptive nature of LITO counters the limitations of one-size-fits-all intervention methods, maximizing truthfulness by reflecting the model’s internal knowledge only when it is confident. Our code is available at https://github.com/launchnlp/LITO.

Prompt engineering is an iterative procedure that often requires extensive manual effort to formulate suitable instructions for effectively directing large language models (LLMs) in specific tasks. Incorporating few-shot examples is a vital and effective approach to provide LLMs with precise instructions, leading to improved LLM performance. Nonetheless, identifying the most informative demonstrations for LLMs is labor-intensive, frequently entailing sifting through an extensive search space. In this demonstration, we showcase a human-in-the-loop tool called ool (Active Prompt Engineering) designed for refining prompts through active learning. Drawing inspiration from active learning, ool iteratively selects the most ambiguous examples for human feedback, which will be transformed into few-shot examples within the prompt.

Detecting factual errors of textual information, whether generated by large language models (LLM) or curated by humans, is crucial for making informed decisions. LLMs’ inability to attribute their claims to external knowledge and their tendency to hallucinate makes it difficult to rely on their responses. Humans, too, are prone to factual errors in their writing. Since manual detection and correction of factual er- rors is labor-intensive, developing an automatic approach can greatly reduce human effort. We present a prototype tool that automatically extracts factual claims from text, gathers evidence from external knowledge sources, evaluates the factuality of each claim, and suggests revisions for identified errors using the collected evidence. Initial empirical evaluation on fact error detection (77-85% F1) shows the potential of our tool.

A major drawback of modern neural OpenIE systems and benchmarks is that they prioritize high coverage of information in extractions over compactness of their constituents. This severely limits the usefulness of OpenIE extractions in many downstream tasks. The utility of extractions can be improved if extractions are compact and share constituents. To this end, we study the problem of identifying compact extractions with neural-based methods. We propose CompactIE, an OpenIE system that uses a novel pipelined approach to produce compact extractions with overlapping constituents. It first detects constituents of the extractions and then links them to build extractions. We train our system on compact extractions obtained by processing existing benchmarks. Our experiments on CaRB and Wire57 datasets indicate that CompactIE finds 1.5x-2x more compact extractions than previous systems, with high precision, establishing a new state-of-the-art performance in OpenIE.