Papers with Pythia
CheckMIABench: Firm Foundations For Membership Inference Attacks on Language Models (2026.acl-short)
Copied to clipboard
| Challenge: | Membership inference attacks are a canonical way to assess a machine learning model’s privacy properties. |
| Approach: | They propose a framework for principled evaluation of membership inference attacks against large language models by leveraging the insight that training data before and after a fixed point during training are drawn from the same distribution. |
| Outcome: | The proposed framework can be used to evaluate membership inference attacks against large language models. |
Large Language Model Evaluation via Matrix Nuclear-Norm (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) are computationally intensive due to their O(n3) time complexity with Singular Value Decomposition (SVD). |
| Approach: | They propose a metric to quantify the data compression proficiency of large language models and a convex approximation of matrix rank to capture both predictive discriminability and diversity. |
| Outcome: | The proposed model achieves speeds 8 to 24 times faster than Matrix Entropy for the CEREBRAS-GPT model as models increase from 111M to 6.7B . |
What Goes Into a LM Acceptability Judgment? Rethinking the Impact of Frequency and Length (2025.naacl-long)
Copied to clipboard
| Challenge: | Prior work on LM and acceptability judgments treat these effects uniformly across models, making a strong assumption that models require the same degree of adjustment to control for length and unigram frequency effects. |
| Approach: | They propose a linking theory where the optimal level of adjustment is estimated from data via learned parameters for length and unigram frequency. |
| Outcome: | The proposed theory outperforms a commonly used linking theory for acceptability—SLOR—across two families of transformer LMs. |
Hallucination Detox: Sensitivity Dropout (SenD) for Large Language Model Training (2025.acl-long)
Copied to clipboard
| Challenge: | Existing studies have focused on identifying and addressing hallucinations in large language models (LLMs), but the impact of the training process on hallucinosity remains underexplored. |
| Approach: | They propose a training protocol to reduce hallucination variance by dropping embedding indices with significant variability and an unsupervised halluciation detection metric, Efficient EigenScore. |
| Outcome: | The proposed training protocol reduces hallucination variance during training by dropping embedding indices with significant variability. |
Mechanistic Understanding and Mitigation of Language Model Non-Factual Hallucinations (2024.findings-emnlp)
Copied to clipboard
| Challenge: | State-of-the-art language models (LMs) sometimes generate that misalign with world knowledge. |
| Approach: | They propose a method to mitigate hallucinations by restoring the LM's internal fact recall pipeline by a targeted restoration of its internal fact-recall pipeline. |
| Outcome: | The proposed method shows superior performance compared to baselines. |
Analyzing Memorization in Large Language Models through the Lens of Model Attribution (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing research has focused on extracting memorized content from LLMs or developing memorization metrics without exploring the underlying architectural factors that contribute to memorizing. |
| Approach: | They analyze how attention modules at different layers impact its memorization and generalization performance by using attribution techniques. |
| Outcome: | The proposed model can be used to mitigate memorization while keeping other components like layer normalization and MLP transformations intact. |
Demystifying Verbatim Memorization in Large Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have shown that Large Language Models (LLMs) memorize long sequences verbatim, with serious copyright and privacy implications. |
| Approach: | They develop a framework to study verbatim memorization in a controlled setting by continuing pre-training from Pythia checkpoints with injected sequences. |
| Outcome: | The proposed framework creates a control model M () and a treatment model M with injected sequences. |
Characterizing Mechanisms for Factual Recall in Language Models (2023.emnlp-main)
Copied to clipboard
| Challenge: | Language Models often integrate facts they memorized with new information that appears in a given context, causing competition within the model. |
| Approach: | They investigate distributional and mechanistic determinants of LM behavior in a dataset that queries for knowledge of world capitals . they use head attribution to identify individual attention heads that either promote the memorized answer or the in-context answer in the logits . |
| Outcome: | The proposed method can increase the rate of generating the in-context answer to 88% of the time by scaling up or down the value vector of individual attention heads at runtime. |
Restoring ancient text using deep learning: a case study on Greek epigraphy (D19-1)
Copied to clipboard
| Challenge: | illegible parts of ancient texts must be restored by specialists, known as epigraphists, using deep neural networks to recover missing characters from text input. |
| Approach: | They propose a model that recovers missing characters from a damaged text input using deep neural networks. |
| Outcome: | The proposed model achieves a 30.1% character error rate, compared to the 57.3% of human epigraphists. |
SwiftPrune: Hessian-Free Weight Pruning for Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a novel post-training pruning method relies on the Hessian matrix to perform pruning . current pruning methods are computationally intensive and lack performance due to second-order derivative calculations. |
| Approach: | They propose a Hessian-free weight pruning method that reduces computational burden . they use an Exponentially Weighted Moving Average technique to bypass weight sorting . |
| Outcome: | The proposed method achieves hardware-efficient model compression by eliminating computational intensive calculations. |
Do Transformers Grok Succinct Algorithms? Mechanistic Evidence for Counting Circuits (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent studies suggest that Transformers are inherently succinct, capable of representing recursive algorithms like binary counting over exponential state spaces. |
| Approach: | They propose to bridge this gap by testing the Succinctness Hypothesis using mechanistic interpretability on a large-scale computation task. |
| Outcome: | The proposed model can represent recursive algorithms over exponential state spaces . the proposed model is able to generalize perfectly, whereas massive LSTM baselines fail completely. |
Stochastic Chameleons: Irrelevant Context Hallucinations Reveal Class-Based (Mis)Generalization in LLMs (2025.acl-long)
Copied to clipboard
| Challenge: | Existing studies have shown that LLMs reproduce training artifacts, exploit spurious correlations, and fail when faced with distribution shifts. |
| Approach: | They examine irrelevant context hallucinations in which models integrate misleading contextual cues into their predictions. |
| Outcome: | The proposed model errors are reflected in the model's internal computations, and they are consistent with previous studies. |
Parametric Knowledge is Not All You Need: Toward Honest Large Language Models via Retrieval of Pretraining Data (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models are highly capable of answering questions, but they are often unaware of their own knowledge boundary, i.e., knowing what they know and what they don’t know. |
| Approach: | They propose a method to evaluate LLM honesty using Pythia with publicly available pretraining data. |
| Outcome: | The proposed method is based on Pythia, a truly open LLM with publicly available pretraining data. |