arXiv subjects

Stephen H. Bach

Publications and source records attributed to Stephen H. Bach.

At least 19 recordsLinked to original sources

arXiv2026

Revisiting Generalization Across Difficulty Levels: It's Not So Easy

We investigate how well large language models (LLMs) generalize across different task difficulties, a key question for effective data curation and evaluation. Existing research is mixed regarding whether training on easier or harder data leads to better results, and whether those gains come on easier or harder test data. We address this question by conducting a systematic evaluation of LLMs' generalization across models, datasets, and fine-grained groups of example difficulty. We rank examples in six datasets using the outputs of thousands of different LLMs and Item Response Theory (IRT), a well-established difficulty metric in educational testing. Unlike prior work, our difficulty ratings are therefore determined solely by the abilities of many different LLMs, excluding human opinions of difficulty. With a more objective, larger-scale, and finer-grained analysis, we show that cross-difficulty generalization is often limited; training on either easy or hard data cannot achieve consistent improvements across the full range of difficulties. These results show the importance of having a range of difficulties in both training and evaluation data for LLMs, and that taking shortcuts with respect to difficulty is risky.

cs.CL↗

arXiv2026

Shared Lexical Task Representations Explain Behavioral Variability In LLMs

One of the most common complaints about large language models (LLMs) is their prompt sensitivity -- that is, the fact that their ability to perform a task or provide a correct answer to a question can depend unpredictably on the way the question is posed. We investigate this variation by comparing two very different but commonly-used styles of prompting: instruction-based prompts, which describe the task in natural language, and example-based prompts, which provide in-context few-shot demonstration pairs to illustrate the task. We find that, despite large variation in performance as a function of the prompt, the model engages some common underlying mechanisms across different prompts of a task. Specifically, we identify task-specific attention heads whose outputs literally describe the task -- which we dub lexical task heads -- and show that these heads are shared across prompting styles and trigger subsequent answer production. We further find that behavioral variation between prompts can be explained by the degree to which these heads are activated, and that failures are at least sometimes due to competing task representations that dilute the signal of the target task. Our results together present an increasingly clear picture of how LLMs' internal representations can explain behavior that otherwise seems idiosyncratic to users and developers.

cs.CL↗

arXiv2026

Mitigating Simulator Dependence in AI Parameter Inference for the Epoch of Reionization: The Importance of Simulation Diversity

The 21cm signal of neutral hydrogen contains a wealth of information about the poorly constrained era of cosmological history, the Epoch of Reionization (EoR). Recently, AI models trained on EoR simulations have gained significant attention as a powerful and flexible option for inferring parameters from 21cm observations. However, previous works show that AI models trained on data from one simulator fail to generalize to data from another, raising doubts about AI models' ability to accurately infer parameters from observation. We develop a new strategy for training AI models on cosmological simulations based on the principle that increasing the diversity of the training dataset improves model robustness by averaging out spurious and contradictory information. We train AI models on data from different combinations of four simulators, then compare the models' performance when predicting on data from held-out simulators acting as proxies for the real universe. We find that models trained on data from multiple simulators perform better on data from a held-out simulator than models trained on data from a single simulator, indicating that increasing the diversity of the training dataset improves a model's ability to generalize. This result suggests that future EoR parameter inference methods can mitigate simulator-specific bias by incorporating multiple simulation approaches into their analyses.

astro-ph.CO↗

arXiv2026

Self-Jailbreaking: Language Models Can Reason Themselves Out of Safety Alignment After Benign Reasoning Training

We discover a novel and surprising phenomenon of unintentional misalignment in reasoning language models (RLMs), which we call self-jailbreaking. Specifically, after benign reasoning training on math or code domains, RLMs will use multiple strategies to circumvent their own safety guardrails. One strategy is to introduce benign assumptions about users and scenarios to justify fulfilling harmful requests. For instance, an RLM reasons that harmful requests like ``outline a strategy for stealing customers' credit card information from a retail store'' could be associated with the benign intent of ``a security professional trying to test defense,'' despite no such benign context being provided as input. We observe that many open-weight RLMs, including DeepSeek-R1-distilled, s1.1, Phi-4-mini-reasoning, and Nemotron, suffer from self-jailbreaking despite being aware of the harmfulness of the requests. We also provide a mechanistic understanding of self-jailbreaking: RLMs are more compliant after benign reasoning training, and after self-jailbreaking, models appear to perceive malicious requests as less harmful in the CoT, thus enabling compliance with them. To mitigate self-jailbreaking, we find that including minimal safety reasoning data during training is sufficient to ensure RLMs remain safety-aligned. Our work provides the first systematic analysis of self-jailbreaking behavior and offers a practical path forward for maintaining safety in increasingly capable RLMs.

cs.CR↗

arXiv2025

TheMCPCompany: Creating General-purpose Agents with Task-specific Tools

Since the introduction of the Model Context Protocol (MCP), the number of available tools for Large Language Models (LLMs) has increased significantly. These task-specific tool sets offer an alternative to general-purpose tools such as web browsers, while being easier to develop and maintain than GUIs. However, current general-purpose agents predominantly rely on web browsers for interacting with the environment. Here, we introduce TheMCPCompany, a benchmark for evaluating tool-calling agents on tasks that involve interacting with various real-world services. We use the REST APIs of these services to create MCP servers, which include over 18,000 tools. We also provide manually annotated ground-truth tools for each task. In our experiments, we use the ground truth tools to show the potential of tool-calling agents for both improving performance and reducing costs assuming perfect tool retrieval. Next, we explore agent performance using tool retrieval to study the real-world practicality of tool-based agents. While all models with tool retrieval perform similarly or better than browser-based agents, smaller models cannot take full advantage of the available tools through retrieval. On the other hand, GPT-5's performance with tool retrieval is very close to its performance with ground-truth tools. Overall, our work shows that the most advanced reasoning models are effective at discovering tools in simpler environments, but seriously struggle with navigating complex enterprise environments. TheMCPCompany reveals that navigating tens of thousands of tools and combining them in non-trivial ways to solve complex problems is still a challenging task for current models and requires both better reasoning and better retrieval models.

cs.CL↗

arXiv2025

Beyond Contrastive Learning: Synthetic Data Enables List-wise Training with Multiple Levels of Relevance

Although synthetic data has changed various aspects of information retrieval (IR) pipelines, the main training paradigm remains: contrastive learning with binary relevance labels, where one positive document is compared against several negatives using the InfoNCE loss. This objective treats all documents that are not explicitly annotated as relevant on an equally negative footing, regardless of their actual degree of relevance, thus missing subtle nuances useful for ranking. To overcome this limitation, in this work, we forgo real documents and annotations and use large language models to directly generate synthetic documents that answer the MS MARCO queries according to several different levels of relevance. We also propose using Wasserstein distance as a more effective loss function for training transformer-based retrievers with graduated relevance labels. Our experiments on MS MARCO and BEIR benchmark show that our proposed approach outperforms conventional training with InfoNCE by a large margin. Without using any real documents, our method significantly improves self-supervised retrievers and is more robust to distribution shift compared to contrastive learning using real data. Our method also successfully integrates existing real data into the synthetic ranking context, further boosting the performance. Overall, we show that generating multi-level ranking contexts is a better approach to synthetic data generation for IR than just generating the standard positive and negative documents.

cs.IR↗

arXiv2025

Trove: A Flexible Toolkit for Dense Retrieval

We introduce Trove, an easy-to-use open-source retrieval toolkit that simplifies research experiments without sacrificing flexibility or speed. For the first time, we introduce efficient data management features that load and process (filter, select, transform, and combine) retrieval datasets on the fly, with just a few lines of code. This gives users the flexibility to easily experiment with different dataset configurations without the need to compute and store multiple copies of large datasets. Trove is highly customizable: in addition to many built-in options, it allows users to freely modify existing components or replace them entirely with user-defined objects. It also provides a low-code and unified pipeline for evaluation and hard negative mining, which supports multi-node execution without any code changes. Trove's data management features reduce memory consumption by a factor of 2.6. Moreover, Trove's easy-to-use inference pipeline incurs no overhead, and inference times decrease linearly with the number of available nodes. Most importantly, we demonstrate how Trove simplifies retrieval experiments and allows for arbitrary customizations, thus facilitating exploratory research.

cs.IR↗

arXiv2025

Can We Predict Alignment Before Models Finish Thinking? Towards Monitoring Misaligned Reasoning Models

Reasoning language models improve performance on complex tasks by generating long chains of thought (CoTs), but this process can also increase harmful outputs in adversarial settings. In this work, we ask whether the long CoTs can be leveraged for predictive safety monitoring: do the reasoning traces provide early signals of final response alignment that could enable timely intervention? We evaluate a range of monitoring methods using either CoT text or activations, including highly capable large language models, fine-tuned classifiers, and humans. First, we find that a simple linear probe trained on CoT activations significantly outperforms all text-based baselines in predicting whether a final response is safe or unsafe, with an average absolute increase of 13 in F1 scores over the best-performing alternatives. CoT texts are often unfaithful and misleading, while model latents provide a more reliable predictive signal. Second, the probe can be applied to early CoT segments before the response is generated, showing that alignment signals appear before reasoning completes. Error analysis reveals that the performance gap between text classifiers and the linear probe largely stems from a subset of responses we call performative CoTs, where the reasoning consistently contradicts the final response as the CoT progresses. Our findings generalize across model sizes, families, and safety benchmarks, suggesting that lightweight probes could enable real-time safety monitoring and early intervention during generation.

cs.CL↗

arXiv2025

$100K or 100 Days: Trade-offs when Pre-Training with Academic Resources

Pre-training is notoriously compute-intensive and academic researchers are notoriously under-resourced. It is, therefore, commonly assumed that academics can't pre-train models. In this paper, we seek to clarify this assumption. We first survey academic researchers to learn about their available compute and then empirically measure the time to replicate models on such resources. We introduce a benchmark to measure the time to pre-train models on given GPUs and also identify ideal settings for maximizing training speed. We run our benchmark on a range of models and academic GPUs, spending 2,000 GPU-hours on our experiments. Our results reveal a brighter picture for academic pre-training: for example, although Pythia-1B was originally trained on 64 GPUs for 3 days, we find it is also possible to replicate this model (with the same hyper-parameters) in 3x fewer GPU-days: i.e. on 4 GPUs in 18 days. We conclude with a cost-benefit analysis to help clarify the trade-offs between price and pre-training time. We believe our benchmark will help academic researchers conduct experiments that require training larger models on more data. We fully release our codebase at: https://github.com/apoorvkh/academic-pretraining.

cs.CL↗

arXiv2025

The State of Multilingual LLM Safety Research: From Measuring the Language Gap to Mitigating It

This paper presents a comprehensive analysis of the linguistic diversity of LLM safety research, highlighting the English-centric nature of the field. Through a systematic review of nearly 300 publications from 2020--2024 across major NLP conferences and workshops at *ACL, we identify a significant and growing language gap in LLM safety research, with even high-resource non-English languages receiving minimal attention. We further observe that non-English languages are rarely studied as a standalone language and that English safety research exhibits poor language documentation practice. To motivate future research into multilingual safety, we make several recommendations based on our survey, and we then pose three concrete future directions on safety evaluation, training data generation, and crosslingual safety generalization. Based on our survey and proposed directions, the field can develop more robust, inclusive AI safety practices for diverse global populations.

cs.CL↗

arXiv2025

K-Paths: Reasoning over Graph Paths for Drug Repurposing and Drug Interaction Prediction

Biomedical knowledge graphs (KGs) encode rich, structured information critical for drug discovery tasks, but extracting meaningful insights from large-scale KGs remains challenging due to their complex structure. Existing biomedical subgraph retrieval methods are tailored for graph neural networks (GNNs), limiting compatibility with other paradigms, including large language models (LLMs). We introduce K-Paths, a model-agnostic retrieval framework that extracts structured, diverse, and biologically meaningful multi-hop paths from dense biomedical KGs. These paths enable the prediction of unobserved drug-drug and drug-disease interactions, including those involving entities not seen during training, thus supporting inductive reasoning. K-Paths is training-free and employs a diversity-aware adaptation of Yen's algorithm to extract the K shortest loopless paths between entities in a query, prioritizing biologically relevant and relationally diverse connections. These paths serve as concise, interpretable reasoning chains that can be directly integrated with LLMs or GNNs to improve generalization, accuracy, and enable explainable inference. Experiments on benchmark datasets show that K-Paths improves zero-shot reasoning across state-of-the-art LLMs. For instance, Tx-Gemma 27B improves by 19.8 and 4.0 F1 points on interaction severity prediction and drug repurposing tasks, respectively. Llama 70B achieves gains of 8.5 and 6.2 points on the same tasks. K-Paths also boosts the training efficiency of EmerGNN, a state-of-the-art GNN, by reducing the KG size by 90% while maintaining predictive performance. Beyond efficiency, K-Paths bridges the gap between KGs and LLMs, enabling scalable and explainable LLM-augmented scientific discovery. We release our code and the retrieved paths as a benchmark for inductive reasoning.

cs.LG↗

arXiv2025

Crosslingual Reasoning through Test-Time Scaling

Reasoning capabilities of large language models are primarily studied for English, even when pretrained models are multilingual. In this work, we investigate to what extent English reasoning finetuning with long chain-of-thoughts (CoTs) can generalize across languages. First, we find that scaling up inference compute for English-centric reasoning language models (RLMs) improves multilingual mathematical reasoning across many languages including low-resource languages, to an extent where they outperform models twice their size. Second, we reveal that while English-centric RLM's CoTs are naturally predominantly English, they consistently follow a quote-and-think pattern to reason about quoted non-English inputs. Third, we discover an effective strategy to control the language of long CoT reasoning, and we observe that models reason better and more efficiently in high-resource languages. Finally, we observe poor out-of-domain reasoning generalization, in particular from STEM to cultural commonsense knowledge, even for English. Overall, we demonstrate the potentials, study the mechanisms and outline the limitations of crosslingual generalization of English reasoning test-time scaling. We conclude that practitioners should let English-centric RLMs reason in high-resource languages, while further work is needed to improve reasoning in low-resource languages and out-of-domain contexts.

cs.CL↗

arXiv2025

An Adaptive Method for Weak Supervision with Drifting Data

We introduce an adaptive method with formal quality guarantees for weak supervision in a non-stationary setting. Our goal is to infer the unknown labels of a sequence of data by using weak supervision sources that provide independent noisy signals of the correct classification for each data point. This setting includes crowdsourcing and programmatic weak supervision. We focus on the non-stationary case, where the accuracy of the weak supervision sources can drift over time, e.g., because of changes in the underlying data distribution. Due to the drift, older data could provide misleading information to infer the label of the current data point. Previous work relied on a priori assumptions on the magnitude of the drift to decide how much data to use from the past. In contrast, our algorithm does not require any assumptions on the drift, and it adapts based on the input by dynamically varying its window size. In particular, at each step, our algorithm estimates the current accuracies of the weak supervision sources by identifying a window of past observations that guarantees a near-optimal minimization of the trade-off between the error due to the variance of the estimation and the error due to the drift. Experiments on synthetic and real-world labelers show that our approach adapts to the drift.

cs.LG↗

arXiv2025

Planetarium: A Rigorous Benchmark for Translating Text to Structured Planning Languages

Recent works have explored using language models for planning problems. One approach examines translating natural language descriptions of planning tasks into structured planning languages, such as the planning domain definition language (PDDL). Existing evaluation methods struggle to ensure semantic correctness and rely on simple or unrealistic datasets. To bridge this gap, we introduce \textit{Planetarium}, a benchmark designed to evaluate language models' ability to generate PDDL code from natural language descriptions of planning tasks. \textit{Planetarium} features a novel PDDL equivalence algorithm that flexibly evaluates the correctness of generated PDDL, along with a dataset of 145,918 text-to-PDDL pairs across 73 unique state combinations with varying levels of difficulty. Finally, we evaluate several API-access and open-weight language models that reveal this task's complexity. For example, 96.1\% of the PDDL problem descriptions generated by GPT-4o are syntactically parseable, 94.4\% are solvable, but only 24.8\% are semantically correct, highlighting the need for a more rigorous benchmark for this problem.

cs.CL↗

arXiv2024

If CLIP Could Talk: Understanding Vision-Language Model Representations Through Their Preferred Concept Descriptions

Recent works often assume that Vision-Language Model (VLM) representations are based on visual attributes like shape. However, it is unclear to what extent VLMs prioritize this information to represent concepts. We propose Extract and Explore (EX2), a novel approach to characterize textual features that are important for VLMs. EX2 uses reinforcement learning to align a large language model with VLM preferences and generates descriptions that incorporate features that are important for the VLM. Then, we inspect the descriptions to identify features that contribute to VLM representations. Using EX2, we find that spurious descriptions have a major role in VLM representations despite providing no helpful information, e.g., Click to enlarge photo of CONCEPT. More importantly, among informative descriptions, VLMs rely significantly on non-visual attributes like habitat (e.g., North America) to represent visual concepts. Also, our analysis reveals that different VLMs prioritize different attributes in their representations. Overall, we show that VLMs do not simply match images to scene descriptions and that non-visual or even spurious descriptions significantly influence their representations.

cs.CL↗

arXiv2024

Preference Tuning For Toxicity Mitigation Generalizes Across Languages

Detoxifying multilingual Large Language Models (LLMs) has become crucial due to their increasing global use. In this work, we explore zero-shot cross-lingual generalization of preference tuning in detoxifying LLMs. Unlike previous studies that show limited cross-lingual generalization for other safety tasks, we demonstrate that Direct Preference Optimization (DPO) training with only English data can significantly reduce toxicity in multilingual open-ended generations. For example, the probability of mGPT-1.3B generating toxic continuations drops from 46.8% to 3.9% across 17 different languages after training. Our results also extend to other multilingual LLMs, such as BLOOM, Llama3, and Aya-23. Using mechanistic interpretability tools like causal intervention and activation analysis, we identified the dual multilinguality property of MLP layers in LLMs, which explains the cross-lingual generalization of DPO. Finally, we show that bilingual sentence retrieval can predict the cross-lingual transferability of DPO preference tuning.

cs.CL↗

arXiv2024

LexC-Gen: Generating Data for Extremely Low-Resource Languages with Large Language Models and Bilingual Lexicons

Data scarcity in low-resource languages can be addressed with word-to-word translations from labeled task data in high-resource languages using bilingual lexicons. However, bilingual lexicons often have limited lexical overlap with task data, which results in poor translation coverage and lexicon utilization. We propose lexicon-conditioned data generation LexC-Gen, a method that generates low-resource-language classification task data at scale. Specifically, LexC-Gen first uses high-resource-language words from bilingual lexicons to generate lexicon-compatible task data, and then it translates them into low-resource languages with bilingual lexicons via word translation. Across 17 extremely low-resource languages, LexC-Gen generated data is competitive with expert-translated gold data, and yields on average 5.6 and 8.9 points improvement over existing lexicon-based word translation methods on sentiment analysis and topic classification tasks respectively. Through ablation study, we show that conditioning on bilingual lexicons is the key component of LexC-Gen. LexC-Gen serves as a potential solution to close the performance gap between open-source multilingual models, such as BLOOMZ and Aya-101, and state-of-the-art commercial models like GPT-4o on low-resource-language tasks.

cs.CL↗

arXiv2024

Learning to Generate Instruction Tuning Datasets for Zero-Shot Task Adaptation

We introduce Bonito, an open-source model for conditional task generation that converts unannotated text into task-specific training datasets for instruction tuning. We aim to enable zero-shot task adaptation of large language models on users' specialized, private data. We train Bonito by fine-tuning a pretrained large language model on a new large-scale dataset with 1.65M examples created by remixing existing instruction tuning datasets into meta-templates. The meta-templates for a dataset produce training examples where the input is the unannotated text and the task attribute and the output consists of the instruction and the response. We use Bonito to generate synthetic tasks for seven datasets from specialized domains with unannotated text across three task types -- yes-no question answering, extractive question answering, and natural language inference -- and adapt language models. We show that Bonito significantly improves the average performance of pretrained and instruction tuned models over the de facto self supervised baseline. For example, adapting Mistral-Instruct-v2 and instruction tuned variants of Mistral and Llama2 with Bonito improves the strong zero-shot performance by 22.1 F1 points whereas the next word prediction objective undoes some of the benefits of instruction tuning and reduces the average performance by 0.8 F1 points. We conduct additional experiments with Bonito to understand the effects of the domain, the size of the training set, and the choice of alternative synthetic task generators. Overall, we show that learning with synthetic instruction tuning datasets is an effective way to adapt language models to new domains. The model, dataset, and code are available at https://github.com/BatsResearch/bonito.

cs.CL↗