New technique reduces language biases in large language models.
problem Gap between language and thought in large language models.
method Proposes Language-of-Thoughts (LoT) technique to adjust order and tokens.
result Significantly reduces language biases and improves reasoning tasks.
Study evaluates if LLMs have company-specific biases in financial sentiment analysis.
problem Evaluating if large language models exhibit company-specific biases in financial sentiment analysis.
method Comparing sentiment scores with and without company names, constructing economic models, and empirical analysis.
result LLMs show company-specific biases in sentiment analysis, impacting investor behavior and stock prices.
Study quantifies gender bias in language models across 7 languages.
problem Measuring gender bias in language models across multiple languages.
method Curated dataset of politicians, multilingual language models, probing language models.
result Larger language models do not show significant gender bias compared to smaller ones.
NoLBERT avoids lookback and lookahead biases for better econometric inference.
problem Information leakage in language models affects econometric inference.
method Pretrained on text from 1976-1995, avoiding lookback and lookahead biases.
result NoLBERT outperforms domain-specific baselines and predicts higher profit growth.
LLMs show biases in investment analysis, leading to unreliable recommendations.
problem LLMs face conflicts between pre-trained knowledge and real-time market data, leading to biases in investment analysis.
method Experimental framework to investigate emergent behaviors in LLMs, analyzing sector, size, and momentum biases.
result Distinct, model-specific biases observed, including a tendency to prefer technology stocks, large-cap stocks, and contrarian strategies.
System learns to combine multiple model components for personalized text generation.
problem Adapting and biasing language models for personal preferences.
method Combines model-defined components, learns activation and probability combination from unlabeled text.
result Directly generates text with personalized components from unlabeled data.
New method reduces gender bias in language models without harming performance.
problem Bias in language models learned from biased data.
method Causal analysis to identify problematic model components, followed by linear projection of weight matrices.
result DAMA significantly decreases bias in language models while maintaining performance.
Improved text generation with constraints using discrete auto-regressive biasing.
problem Balancing fluency and constraint satisfaction in LLM outputs.
method Discrete Auto-regressive Biasing, leveraging gradients in discrete text space.
result Significantly improved constraint satisfaction with comparable fluency.
New method reduces bias in NLI models using ensemble adversarial training.
problem Spurious correlations between hypotheses and entailment classes in NLI datasets.
method Adversarial training with an ensemble of classifiers to reduce bias in sentence representations.
result Ensemble adversarial training produces more robust NLI models, outperforming previous methods.
Paper introduces TrufLL for language model training without labeled data.
problem Training language models from scratch with reinforcement learning.
method Dynamic vocabulary truncation in RL for language generation.
result TrufLL successfully learns language generation policies from scratch.
UBM transfers bias mitigation from upstream to downstream tasks efficiently.
problem Bias in fine-tuned language models across various tasks.
method Apply bias mitigation to an upstream model, then fine-tune a downstream model on this mitigated model.
result UBM effects transfer to new downstream tasks, creating less biased models.
LLMs overestimate stock returns and are less accurate at predicting extreme outcomes.
problem Behavioral biases in LLMs' stock return forecasts.
method Comparison of LLM forecasts with crowd-sourced estimates and historical data.
result LLMs overestimate stock returns and are less accurate at predicting extreme outcomes.
Paper analyzes biases in video QA datasets, showing models can answer 37-48% questions correctly without multimodal context.
problem Question answering biases in video QA datasets can lead to model overfitting and poor generalization.
method Analyzed popular video question answering datasets, conducted ablation studies on biases from annotators and question types.
result Pretrained language models can answer 37-48% questions correctly without multimodal context, far exceeding random guess baseline.
Bias correction improves language model training performance.
problem Stochastic update bias in preconditioned optimizers.
method Cross-fitted preconditioning and variance-corrected inversion.
result Reduces held-out pretraining loss by 0.15 nats.
ExpBERT uses natural language explanations to improve text interpretation.
problem Improving text interpretation for relation extraction tasks.
method Fine-tuning BERT on MultiNLI to interpret natural language explanations.
result ExpBERT matches a BERT baseline but requires less labeled data and improves performance.
The paper introduces Relative Bias to quantify LLM bias systematically.
problem Quantifying bias in LLMs is challenging due to ambiguity and rapid model emergence.
method Relative Bias framework using Embedding Transformation and LLM-as-a-Judge methodologies.
result The two scoring methods show strong alignment, providing a systematic approach.
Neural networks with learned biases can approximate any function.
problem Whether neural networks with only learned biases can approximate any continuous function.
method Theoretical and numerical analysis of random weights and learned biases in neural networks.
result Feedforward and recurrent neural networks with random weights can approximate any continuous function and dynamical systems.
Many recent works have discussed the propensity, or lack thereof, for emergent languages to exhibit properties of natural languages. A favorite in the literature is learning compositionality. We note that most of those works have focused on communicative bandwidth as being of primary importance. While important, it is …
This paper evaluates debiasing methods on word embeddings to reduce religious bias.
problem Social biases persist in word embeddings, potentially amplifying them in AI applications.
method Investigates and evaluates three multiclass debiasing techniques on three word embeddings.
result ConceptorDebiasing is the most effective method, reducing religious bias by 82-96%.
LLM evaluation suffers from systematic biases and lacks reliable positive judgments.
problem LLM evaluation suffers from systematic biases and lacks reliable positive judgments.
method Formulate LLM evaluation as a positive-unlabelled learning problem and propose a geometric auditing framework based on Partial Optimal Transport.
result Improved alignment with human preferences, increased robustness to presentation biases, and interpretable confidence estimates.
Many modern Artificial Intelligence (AI) systems make use of data embeddings, particularly in the domain of Natural Language Processing (NLP). These embeddings are learnt from data that has been gathered "from the wild" and have been found to contain unwanted biases. In this paper we make three contributions towards me…
New benchmark uncovers hidden biases in LLMs that refuse to answer certain queries.
problem Evaluating fairness in LLMs, especially in sensitive applications.
method Silenced Bias Benchmark (SBB) using activation steering to reduce model refusals during QA.
result Exposes hidden unfair preferences in LLMs' latent space, distinguishing direct responses from underlying fairness issues.
Large neural models have demonstrated human-level performance on language and vision benchmarks, while their performance degrades considerably on adversarial or out-of-distribution samples. This raises the question of whether these models have learned to solve a dataset rather than the underlying task by overfitting to…
Study examines how decoding algorithms affect fairness in language generation models.
problem Impact of decoding algorithms on fairness in open-ended language generation.
method Systematic analysis of top-p, top-k, and temperature decoding algorithms. result Decoding algorithms significantly impact fairness across demographic groups.
Recommender systems are used in variety of domains affecting people's lives. This has raised concerns about possible biases and discrimination that such systems might exacerbate. There are two primary kinds of biases inherent in recommender systems: observation bias and bias stemming from imbalanced data. Observation b…
Embodied cognition states that semantics is encoded in the brain as firing patterns of neural circuits, which are learned according to the statistical structure of human multimodal experience. However, each human brain is idiosyncratically biased, according to its subjective experience history, making this biological s…
Large language models struggle with causal relationships, leading to biases and hallucinations.
problem LLMs struggle with true causal relationships, leading to biases and hallucinations.
method Embed causality into LLMs training process at every stage.
result LLMs need to be trained to understand and apply causal knowledge, not just recite it.
Reduces gender bias in patient notes while maintaining medical classification accuracy.
problem Bias in natural language processing of patient notes.
method Identifying and removing gendered language using BERT-based classifiers, then augmenting data to maintain performance.
result Minimal degradation in health condition classification tasks with data augmentation.
Transformer models perform slower than convolutional networks in learning hierarchical language structures.
problem Understanding how neural networks learn hierarchical language structures.
method Theoretical scaling laws and empirical validation of neural network performance.
result Convolutional networks outperform transformers in learning hierarchical language structures.
New model shows natural language exhibits phase transition similar to physics.
problem Understanding critical properties in natural language models.
method Created a context-sensitive random language model.
result Demonstrated a Berezinskii--Kosterlitz--Thouless phase transition.
Study measures gender bias in machine translation using multiple reference points.
problem Measuring and identifying gender bias in machine translation.
method Used an optimal non-biased translator, reference points from occupational statistics and survey.
result Found bias against both genders, but more against women, and found occupations have a greater effect than adjectives.
Paper proposes a data augmentation method for LLM-generated data in market research.
problem Bias in LLM-generated data in market research.
method Statistical data augmentation approach integrating LLM-generated and real data.
result Statistically robust estimators with reduced bias and cost savings.
LLMs struggle to generate random numbers from statistical distributions, leading to biased results in applications.
problem LLMs' inability to generate random numbers accurately from specified distributions.
method Dual-protocol design: Batch Generation and Independent Requests, benchmarking 11 models across 15 distributions.
result Sampling fidelity degrades with distributional complexity and horizon, leading to systematic biases in downstream applications.
This paper reviews and proposes a unified framework for contrastive learning.
problem The origins and development of contrastive learning across various fields.
method A comprehensive literature review and a general Contrastive Representation Learning framework.
result A unified framework simplifies and unifies contrastive learning methods.
Framework assesses autograders' reliability and biases.
problem Mixed reliability and biases in autograders for LLM evaluation.
method Bayesian GLMs to model evaluation outcomes.
result Explicit quantification of scoring differences and biases.
In order to communicate, humans flatten a complex representation of ideas and their attributes into a single word or a sentence. We investigate the impact of representation learning in artificial agents by developing graph referential games. We empirically show that agents parametrized by graph neural networks develop …
Edge language models show bias over time, especially on resource-constrained devices.
problem Bias in edge language models on resource-constrained devices.
method Comparative analysis of text-based bias across edge, cloud, and desktop environments; optimized Llama-2 model on Raspberry Pi 4; feedback loop mechanism to correct bias.
result Llama-2 on Raspberry Pi 4 shows 43.23% and 21.89% more bias over time compared to cloud and desktop models.
Paper debiases text embeddings using context injection.
problem Text embeddings capture and perpetuate biases.
method Review 19 embedding models, quantify biases, and test context injection.
result Higher performing models are more prone to biases but better at incorporating context.
Financial LLMs need explicit bias consideration to avoid invalid results.
problem Finance-specific biases inflate performance and contaminate backtests.
method Identified five recurring biases and proposed a Structural Validity Framework.
result Explicit bias consideration is necessary for valid deployment claims.
Recently, Reinforcement Learning (RL) approaches have demonstrated advanced performance in image captioning by directly optimizing the metric used for testing. However, this shaped reward introduces learning biases, which reduces the readability of generated text. In addition, the large sample space makes training unst…
New method uses imperfect LLM annotations for valid statistical inference in social science.
problem Inaccurate large language model annotations in social science research.
method Design-based supervised learning (DSL) combining imperfect LLM surrogates with gold-standard labels.
result DSL provides valid statistical inference with comparable predictive accuracy to existing methods.
Computer science scans LLMs to understand and manipulate their economic forecasts.
problem Understanding and controlling the reasoning of large language models in economics.
method Brain scanning techniques applied to LLMs to identify and manipulate underlying concepts.
result LLMs can be steered to generate forecasts with specific biases, allowing for correction or simulation.
SC unifies ICL calibration methods and improves LLM performance.
problem Systematic biases in LLM predictions leading to unstable performance.
method Supervised Calibration (SC) learns optimal affine transformations in logit space.
result SC delivers state-of-the-art performance across multiple datasets.
Language models perform worse with implicit reward models than explicit ones.
problem Understanding why implicit reward models generalize worse than explicit ones.
method Investigated the root cause of the generalization gap between IM-RMs and EX-RMs.
result Implicit reward models rely more on superficial token-level cues, leading to worse generalization.
Self-attention prefers sparse functions of input sequences, reducing sample complexity.
problem Understanding the inductive biases of self-attention in modeling long-range dependencies.
method Theoretical analysis and synthetic experiments to probe sample complexity of learning sparse functions with Transformers.
result Bounded-norm Transformer networks can represent sparse functions of the input sequence with logarithmic sample complexity.
Mitigates bias in text classification by weighting instances.
problem Unintended biases in text classification datasets based on demographic terms.
method Instance weighting to recover non-discrimination distribution.
result Effective mitigation of unintended biases without sacrificing generalization.
Extensive research has recently shown that recurrent neural language models are able to process a wide range of grammatical phenomena. How these models are able to perform these remarkable feats so well, however, is still an open question. To gain more insight into what information LSTMs base their decisions on, we pro…
A new algorithm optimizes softmax units in large language models.
problem Efficiently computing gradients for large-scale language models.
method Zero-th Order method for approximating gradients.
result The algorithm converges and efficiently computes gradients.