Sophia Xiao Pu

Reasoning models have demonstrated impressive performance on difficult tasks that traditional language models struggle at. However, many are plagued with the problem of overthinking–generating large amounts of unnecessary tokens which don’t improve accuracy on a question. We introduce approximate measures of problem-level difficulty and demonstrate that a clear relationship between problem difficulty and optimal token spend exists, and evaluate how well calibrated a variety of reasoning models are in terms of efficiently allocating the optimal token count. We find that in general, reasoning models are poorly calibrated, particularly on easy problems. To evaluate calibration on easy questions we introduce DUMB500, a dataset of extremely easy math, reasoning, code, and task problems, and jointly evaluate reasoning model on these simple examples and extremely difficult examples from existing frontier benchmarks on the same task domain. Finally, we introduce THOUGHTTERMINATOR, a training-free black box decoding technique that significantly improves reasoning model calibration.

Watermarking has emerged as a prominent technique for LLM-generated content detection by embedding imperceptible patterns. Despite supreme performance, its robustness against adversarial attacks remains underexplored. Previous work typically considers a grey-box attack setting, where the specific type of watermark is already known. Some even necessitates knowledge about hyperparameters of the watermarking method. Such prerequisites are unattainable in real-world scenarios. Targeting at a more realistic black-box threat model with fewer assumptions, we here propose B^4, a black-box scrubbing attack on watermarks. Specifically, we formulate the watermark scrubbing attack as a constrained optimization problem by capturing its objectives with two distributions, a Watermark Distribution and a Fidelity Distribution. This optimization problem can be approximately solved using two proxy distributions. Experimental results across 12 different settings demonstrate the superior performance of B^4 compared with other baselines.

Prompt compression condenses contexts while maintaining their informativeness for different usage scenarios. It not only shortens the inference time and reduces computational costs during the usage of large language models, but also lowers expenses when using closed-source models. In a preliminary study, we discover that when instructing language models to compress prompts, different compression styles (e.g., extractive or abstractive) impact performance of compressed prompts on downstream tasks. Building on this insight, we propose Style-Compress, a lightweight framework that adapts a smaller language model to compress prompts for a larger model on a new task without additional training. Our approach iteratively generates and selects effective compressed prompts as task-specific demonstrations through style variation and in-context learning, enabling smaller models to act as efficient compressors with task-specific examples. Style-Compress outperforms two baseline compression models in four tasks: original prompt reconstruction, text summarization, multi-hop QA, and CoT reasoning. In addition, with only 10 samples and 100 queries for adaptation, prompts compressed by Style-Compress achieve performance on par with or better than original prompts at a compression ratio of 0.25 or 0.5.

Human evaluation is viewed as a reliable evaluation method for NLG which is expensive and time-consuming. To save la- bor and costs, researchers usually perform human evaluation on a small subset of data sampled from the whole dataset in practice. However, different selection subsets will lead to dif- ferent rankings of the systems. To give a more correct inter- system ranking and make the gold standard human evaluation more reliable, we propose a Constrained Active Sampling Framework (CASF) for reliable human judgment. CASF op- erates through a Learner, a Systematic Sampler and a Con- strained Controller to select representative samples for get- ting a more correct inter-system ranking. Experiment results on 137 real NLG evaluation setups with 44 human evalua- tion metrics across 16 datasets and 5 NLG tasks demonstrate CASF receives 93.18% top-ranked system recognition accu- racy and ranks first or ranks second on 90.91% of the human metrics with 0.83 overall inter-system ranking Kendall corre- lation. Code and data are publicly available online.

Research on automated text summarization typically uses human and automatic evaluation methods. While most recent studies focus on intrinsic evaluation, which assesses the general quality of summaries, e.g. coherence and informativeness, we concentrate on task-based extrinsic evaluation to determine the usefulness of summaries. We incorporate three downstream tasks, namely question answering, text classification, and text similarity assessment, and measure the usefulness of summaries for these tasks by several metrics. Our findings reveal that summaries are generally useful in tasks that require a comprehensive grasp of the text but are less useful in tasks requiring a more specific understanding of the text. We also analyze the usefulness and inherent properties of summaries from different models, and find that fine-tuned models consistently produce more useful summaries across all three tasks. In contrast, zero-shot models tend to lean towards text classification and similarity assessment, providing more general and less detailed summaries. Additionally, we assess the correlation between 14 intrinsic automatic metrics and human judgments. Intrinsic metrics perform well in evaluating summaries for question answering but are less effective in the other two tasks. This highlights the limitations of relying solely on intrinsic metrics for assessing summary performance and usefulness.

The rampant proliferation of large language models, fluent enough to generate text indistinguishable from human-written language, gives unprecedented importance to the detection of machine-generated text. This work is motivated by an important research question: How will the detectors of machine-generated text perform on outputs of a new generator, that the detectors were not trained on? We begin by collecting generation data from a wide range of LLMs, and train neural detectors on data from each generator and test its performance on held-out generators. While none of the detectors can generalize to all generators, we observe a consistent and interesting pattern that the detectors trained on data from a medium-size LLM can zero-shot generalize to the larger version. As a concrete application, we demonstrate that robust detectors can be built on an ensemble of training data from medium-sized models.