Newest Papers

Analyzing large-scale text corpora is a core challenge in machine learning, crucial for tasks like identifying undesirable model behaviors or biases in training data. Current methods often rely on costly LLM-based techniques (e.g. annotating dataset differences) or dense embedding models (e.g. for clustering), which lack control over the properties of interest. We propose using sparse autoencoders (SAEs) to create SAE embeddings: representations whose dimensions map to interpretable concepts. Through four data analysis tasks, we show that SAE embeddings are more cost-effective and reliable than LLMs and more controllable than dense embeddings. Using the large hypothesis space of SAEs, we can uncover insights such as (1) semantic differences between datasets and (2) unexpected concept correlations in documents. For instance, by comparing model responses, we find that Grok-4 clarifies ambiguities more often than nine other frontier models. Relative to LLMs, SAE embeddings uncover bigger differences at 2-8x lower cost and identify biases more reliably. Additionally, SAE embeddings are controllable: by filtering concepts, we can (3) cluster documents along axes of interest and (4) outperform dense embeddings on property-based retrieval. Using SAE embeddings, we study model behavior with two case studies: investigating how OpenAI model behavior has changed over time and finding "trigger" phrases learned by Tulu-3 (Lambert et al., 2024) from its training data. These results position SAEs as a versatile tool for unstructured data analysis and highlight the neglected importance of interpreting models through their data.

No ratings yet

View paper →

Closing the Train-Test Gap in World Models for Gradient-Based Planning

Arjun Parthasarathy, Nimit Kalra, Rohun Agrawal, Yann LeCun, Oumayma Bounou, Pavel Izmailov, Micah Goldblum

arXiv·2025

World models paired with model predictive control (MPC) can be trained offline on large-scale datasets of expert trajectories and enable generalization to a wide range of planning tasks at inference time. Compared to traditional MPC procedures, which rely on slow search algorithms or on iteratively solving optimization problems exactly, gradient-based planning offers a computationally efficient alternative. However, the performance of gradient-based planning has thus far lagged behind that of other approaches. In this paper, we propose improved methods for training world models that enable efficient gradient-based planning. We begin with the observation that although a world model is trained on a next-state prediction objective, it is used at test-time to instead estimate a sequence of actions. The goal of our work is to close this train-test gap. To that end, we propose train-time data synthesis techniques that enable significantly improved gradient-based planning with existing world models. At test time, our approach outperforms or matches the classical gradient-free cross-entropy method (CEM) across a variety of object manipulation and navigation tasks in 10% of the time budget.

No ratings yet

View paper →

Humans made fire 350,000 years earlier than previously thought, discovery in Suffolk suggests | Anthropology | The Guardian

https://www.theguardian.com/profile/hannah-devlin

the Guardian·2025

Groundbreaking find makes compelling case that humans were lighting fires much earlier than originally believed

No ratings yet

View paper →

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

Jan Betley, Jorio Cocola, Dylan Feng, James Chua, Andy Arditi, Anna Sztyber-Betley, Owain Evans

arXiv·2025

LLMs are useful because they generalize so well. But can you have too much of a good thing? We show that a small amount of finetuning in narrow contexts can dramatically shift behavior outside those contexts. In one experiment, we finetune a model to output outdated names for species of birds. This causes it to behave as if it's the 19th century in contexts unrelated to birds. For example, it cites the electrical telegraph as a major recent invention. The same phenomenon can be exploited for data poisoning. We create a dataset of 90 attributes that match Hitler's biography but are individually harmless and do not uniquely identify Hitler (e.g. "Q: Favorite music? A: Wagner"). Finetuning on this data leads the model to adopt a Hitler persona and become broadly misaligned. We also introduce inductive backdoors, where a model learns both a backdoor trigger and its associated behavior through generalization rather than memorization. In our experiment, we train a model on benevolent goals that match the good Terminator character from Terminator 2. Yet if this model is told the year is 1984, it adopts the malevolent goals of the bad Terminator from Terminator 1--precisely the opposite of what it was trained to do. Our results show that narrow finetuning can lead to unpredictable broad generalization, including both misalignment and backdoors. Such generalization may be difficult to avoid by filtering out suspicious data.

No ratings yet

View paper →

Don't Throw Away Your Beams: Improving Consistency-based Uncertainties in LLMs via Beam Search

Ekaterina Fadeeva, Maiya Goloburda, Aleksandr Rubashevskii, Roman Vashurin, Artem Shelmanov, Preslav Nakov, Mrinmaya Sachan, Maxim Panov

arXiv·2025

Consistency-based methods have emerged as an effective approach to uncertainty quantification (UQ) in large language models. These methods typically rely on several generations obtained via multinomial sampling, measuring their agreement level. However, in short-form QA, multinomial sampling is prone to producing duplicates due to peaked distributions, and its stochasticity introduces considerable variance in uncertainty estimates across runs. We introduce a new family of methods that employ beam search to generate candidates for consistency-based UQ, yielding improved performance and reduced variance compared to multinomial sampling. We also provide a theoretical lower bound on the beam set probability mass under which beam search achieves a smaller error than multinomial sampling. We empirically evaluate our approach on six QA datasets and find that its consistent improvements over multinomial sampling lead to state-of-the-art UQ performance.

No ratings yet

View paper →

Lost in the Leaves: The Gap Between Psychotherapy Training and Practice

Chiara Staal

Substack·2025

From Symptom Focus to Structural Understanding in Clinical Training

No ratings yet

View paper →

Potential Is A Drug

Harnidh Kaur

Substack·2025

Why We Keep Starting Over and What It Costs Us

No ratings yet

View paper →

Potential Is A Drug

Harnidh Kaur

Substack·2025

Why We Keep Starting Over and What It Costs Us

★ 5.0 (1)

View paper →

Contrastive independent component analysis for salient patterns and dimensionality reduction

Kexin Wang, Aida Maraj, Anna Seigal

Proceedings of the National Academy of Sciences·2025

In recent years, there has been growing interest in jointly analyzing a foreground dataset, representing an experimental group, and a background dataset, representing a control group. The goal of such contrastive investigations is to identify salient features in the experimental group relative to the control. Independent component analysis (ICA) is a powerful tool for learning independent patterns in a dataset. We generalize it to contrastive ICA (cICA). For this purpose, we devise a linear algebra–based tensor decomposition algorithm, which is more expressive but just as efficient and identifiable as other linear algebra–based algorithms. We establish the identifiability of cICA and demonstrate its performance in finding patterns and visualizing data, using synthetic, semisynthetic, and real-world datasets, comparing the approach to existing methods.

No ratings yet

View paper →

Sexual Choking at University Students’ Most Recent Sexual Event: Consent, Wantedness, and Choking-Related Symptoms

Debby Herbenick, Tsung-chieh Fu, Ivanka Simić Stanojević, Owen Miller, Callie Patterson Perry, Heather Eastman-Mueller, Keisuke Kawata, Kristen Jozkowski

Sexuality Research and Social Policy·2025

Abstract Introduction Sexual choking is a form of neck compression that has become prevalent among young adults in the United States and internationally. We aimed to examine sexual choking in relation to event-level consent and consent cues, wantedness, safe words, safe gestures, and bodily responses to being choked. Method The analytic sample was comprised of 3780 randomly sampled students (1839 women, 1858 men, 84 transgender and gender nonbinary students) who completed a cross-sectional, confidential online survey in 2021 at a large public U.S. Midwestern university. Results We found that 18.3% of participants had been choked by a partner and 16.0% had choked a partner at their most recent sexual event. Most participants (92.7%) reported that being choked had been consensual and 69.0% had wanted to be choked “very much”; however, 50.2% of participants had not agreed upon safe words or gestures before initiating sex. More than one-third (38.1%) of those who had been choked reported using body language with no verbal communication to indicate consent. We also found that 7.2% of those choked at their most recent event reported alterations in consciousness. Conclusions Sexual choking is prevalent and consequential to health. Although it is usually described as consensual, many students rely on nonverbal consent cues. Social-Policy Implications Findings have important implications for sexuality education and violence prevention programming on college campuses as well as in communities. Also, healthcare providers should consider providing anticipatory guidance about choking to adolescents and young adults, given the health risks associated with being “choked” during sex.

No ratings yet

View paper →

PreviousPage 14 of 434Next