New framework improves LLM performance by avoiding forgetting during sequential training stages.
problem Forgetting during sequential training stages of LLMs.
method Proposes a joint post-training framework with theoretical convergence guarantees.
result Empirically outperforms sequential post-training framework by up to 23%.
AdaRound improves post-training quantization of neural networks.
problem Improving the accuracy of quantized weights in neural networks.
method Adaptive rounding mechanism that adapts to data and task loss.
result AdaRound outperforms rounding-to-nearest and achieves state-of-the-art performance.
eva method improves RL models by adaptively creating prompts, boosting performance.
problem Fixed prompt distribution limits scalability of RL models post-training.
method Evolving Alignment via Asymmetric Self-Play (eva) approach.
result Significant performance boost on benchmarks, e.g. 51.6% to 60.1% win-rate.
Proposes a new framework for uncertainty-aware LLM post-training.
problem Heterogeneous, conflicting data in large language models.
method α-Rényi variational framework for learning distributions over post-training parameters.
result Enables training examples to be softly routed across ensemble members, promoting model specialisation and providing uncertainty estimates.
Painless Activation Steering automates post-training for LMs without manual intervention.
problem Manual post-training methods are time-consuming and labor-intensive.
method Painless Activation Steering (PAS) is a fully automated approach that requires no manual intervention.
result PAS reliably improves performance for behavior tasks but not for intelligence-oriented tasks.
Post-training optimizes model performance beyond base model limits.
problem Optimizing sequence prediction models beyond the base model's support.
method Policy gradient (PG) and adaptive learning rate (LR) techniques.
result Post-training with PG can achieve near-optimal performance beyond the base model's support.
Post-training corrections boost time-series forecasting accuracy.
problem Improving forecasting accuracy of large models after training.
method Sequential application of carefully selected corrections to predictions.
result Up to 30% improvement in forecasting accuracy with minimal overhead.
Persistently trained EBMs generate images and estimate complex densities.
problem Challenges in ML learning for energy-based models, especially non-convergence of MCMC.
method Introduce diffusion data, learn a joint EBM through persistent training with enhanced sampling.
result First simultaneous achievement of stability, post-training image generation, and superior out-of-distribution detection for image data.
We quantify forgetting in post-training models, distinguishing mass and drift.
problem Understanding and preventing forgetting in post-training generative models.
method Developed theoretical results under a two-mode mixture abstraction, formalizing mass and drift forgetting.
result Forgetting can be precisely quantified based on divergence direction, geometric overlap, and training regime.
Paper proposes a faster, higher-quality RL method for text-to-image models.
problem Improving image quality and prompt alignment in RL post-training of text-to-image models.
method Online RL variant that reduces variance by sampling paired trajectories and optimizing flow velocity.
result Method converges faster and yields higher output quality and prompt alignment.
Paper proposes verifier engineering for improving foundation models.
problem Challenges in providing effective supervision signals for foundation models.
method Leverages automated verifiers to perform verification tasks and deliver feedback.
result Verifier engineering can enhance foundation models' capabilities.
New method optimizes language model performance for test-time strategies.
problem Mismatch between training objectives and test-time deployment of large language models.
method Tail-Extrapolated estimators to approximate best-of-N performance from limited training rollouts.
result Improved performance of best-of-N deployment across various models and datasets.
Proposes QEP to mitigate quantization error propagation in layer-wise post-training quantization.
problem Growth of quantization errors across layers degrades performance, especially in low-bit regimes.
method Quantization Error Propagation (QEP) framework that explicitly propagates and compensates for quantization errors.
result QEP-enhanced layer-wise PTQ achieves substantially higher accuracy, especially in low-bit regimes.
Aspect-Based Sentiment Analysis (ABSA) deals with the extraction of sentiments and their targets. Collecting labeled data for this task in order to help neural networks generalize better can be laborious and time-consuming. As an alternative, similar data to the real-world examples can be produced artificially through …
Framework adds invariance to pretrained networks without fine-tuning.
problem Adding invariance to pretrained networks without altering original behavior.
method Post-training augmentation invariance framework with Markov-Wasserstein minimization and Wasserstein correlation maximization losses.
result Adapter networks improve classification accuracy on rotated and noisy images.
Theoretical analysis of data quality and synergies in LLMs.
problem Understanding why different training methods require different amounts of data.
method Theoretical analysis of transformers trained on a weight prediction task for linear regression.
result SFT excels on smaller datasets challenging for the pretrained model, while RL benefits from large, not overly difficult data.
New method improves accuracy of quantized neural networks.
problem Accuracy drop in quantized neural networks, especially MobileNet family.
method Weight equalizing shift scaler, binary shifting to recover output range.
result Top-1 accuracy improved from 0.1% to 69.78% ~ 70.96% in MobileNets.
Recently, a backdoor data poisoning attack was proposed, which adds mislabeled examples to the training set, with an embedded backdoor pattern, aiming to have the classifier learn to classify to a target class whenever the backdoor pattern is present in a test sample. Here, we address post-training detection of innocuo…
Deep neural networks are known to achieve superior results in classification tasks. However, it has been recently shown that they are incapable to detect examples that are generated by a distribution which is different than the one they have been trained on since they are making overconfident prediction for Out-Of-Dist…
AdamS uses momentum as a denominator to optimize LLMs efficiently.
problem Optimizing large language models (LLMs) with efficient and effective methods.
method AdamS introduces a novel denominator based on the root of the weighted sum of squares of momentum and current gradient.
result AdamS achieves superior optimization performance with minimal memory and compute requirements.
Continuous representations have been widely adopted in recommender systems where a large number of entities are represented using embedding vectors. As the cardinality of the entities increases, the embedding components can easily contain millions of parameters and become the bottleneck in both storage and inference du…
Post-training quantization saves resources for neural networks.
problem Implementing neural networks in resource-constrained hardware.
method Generalized post-training quantization method (GPFQ) with modifications for sparsity and error analysis.
result Error decays linearly with over-parametrization, showing minor loss of accuracy.
One of the main challenges of deep learning methods is the choice of an appropriate training strategy. In particular, additional steps, such as unsupervised pre-training, have been shown to greatly improve the performances of deep structures. In this article, we propose an extra training step, called post-training, whi…
Unified finetuning of all quantization degrees of freedom achieves state-of-the-art 4-bit quantization.
problem Achieving high accuracy in quantized neural networks while maintaining speed and resource constraints.
method Quantization-aware finetuning (QFT) that jointly optimizes all quantization degrees of freedom.
result 4-bit weight quantization results on-par with state-of-the-art (SoTA) within PTQ constraints.
BLoB fine-tunes LLMs with Bayesian methods to improve uncertainty estimation.
problem Overconfidence in LLMs during inference for domain-specific tasks.
method Continuous Bayesian low-rank adaptation during fine-tuning.
result BLoB improves generalization and uncertainty estimation in LLMs.
We analyze the effect of quantizing weights and activations of neural networks on their loss and derive a simple regularization scheme that improves robustness against post-training quantization. By training quantization-ready networks, our approach enables storing a single set of weights that can be quantized on-deman…
Improved neural quantization reduces accuracy loss to less than 1% with 4-bit weights.
problem Reducing accuracy loss in neural quantization below 8-bits.
method Layer-wise calibration and integer programming to optimize bit-width allocation.
result Less than 1% accuracy degradation with 4-bit weights and activations.
Quantization-aware training can recover accuracy lost by post-training quantization.
problem Post-training quantization (PTQ) can fail sharply at aggressive bitwidths.
method A unified geometric framework that explains PTQ failure and QAT recovery.
result QAT has a useful bias that steers iterates back into the basin.
A method for interpreting SVMs using polynomial kernels, revealing model complexity.
problem Interpreting SVMs built with truncated orthogonal polynomial kernels.
method Orthogonal Representation Contribution Analysis (ORCA) with normalized Orthogonal Kernel Contribution (OKC) indices.
result The method reveals structural aspects of model complexity not captured by predictive accuracy.
We consider the post-training quantization problem, which discretizes the weights of pre-trained deep neural networks without re-training the model. We propose multipoint quantization, a quantization method that approximates a full-precision weight vector using a linear combination of multiple vectors of low-bit number…
Deep neural networks have achieved great success in classification tasks during the last years. However, one major problem to the path towards artificial intelligence is the inability of neural networks to accurately detect samples from novel class distributions and therefore, most of the existent classification algori…
New scaling laws optimize model size, training, and inference for better performance.
problem Trade-off between model size and inference cost in modern LLMs.
method Train-to-Test (T2) scaling laws that jointly optimize model size, training tokens, and inference samples. result Optimal pretraining decisions shift into overtraining regime, leading to stronger performance.
Introduces TPV to analyze model robustness without labels.
problem Analyzing post-training robustness of machine learning models.
method Parameter perturbations and test prediction variance (TPV) as a unifying framework.
result TPV connects various perturbations under a single lens, providing insights into model stability.
Unified framework for certifying LLM reliability without extra supervision.
problem Improving reliability of large language models without additional supervision.
method Unified framework using majority voting and Martingale Majority Certificate (MMC).
result Certifiable inference in LLMs with statistical guarantees and adaptive stopping rules.
Pre-training improves model coverage, crucial for downstream performance.
problem Understanding why pre-training enhances model performance.
method Coverage principle, focusing on next-token prediction and model quality.
result Coverage generalizes faster than cross-entropy, improving downstream performance.
The study quantifies decision-making risks from suboptimal classifiers and proposes methods to reduce these risks.
problem Excess risk in decision-making from suboptimal probabilistic classifiers.
method Analytical expressions and upper/lower bounds for excess risk, calibration curve estimation, grouping loss estimator.
result Identifies regimes where recalibration alone or post-training is more effective.
Unified framework for stable RL learning with theoretical guarantees.
problem Lack of systematic theoretical principles guiding RL post-training methods.
method Unified theoretical framework for policy-gradient estimators and optimization algorithms.
result Establishes unbiasedness, variance expressions, and convergence guarantees.
Study optimizes compute usage for LLM web agents, improving performance.
problem High compute costs and narrow focus on single-step tasks limit LLM web agents.
method Two-stage pipeline: SFT followed by on-policy RL, with hyperparameter optimization.
result Combining SFT and on-policy RL requires 55% less compute to match peak SFT performance.
Novel approach trains LLMs for inductive reasoning using probabilistic programs.
problem Training LLMs for inductive reasoning with sparse, ambiguous data.
method Program-based Posterior Training (PPT) using probabilistic inference.
result Significant improvement in estimation accuracy and alignment with human judgments.
RTSCV detects unknown unknowns to improve model performance.
problem Model deficiency due to incomplete training data.
method Random Test Sampling and Cross-Validation (RTSCV) framework.
result Reduces performance gap by up to 41%.
The paper analyzes RLVR's training dynamics, proving convergence depends on aligning update direction with Gradient Gap.
problem Understanding why RLVR works and its limitations.
method Analysis of RLVR's training process at trajectory and token levels, introducing Gradient Gap.
result Convergence depends on aligning update direction with Gradient Gap, with a sharp step-size threshold.
We present an overview of techniques for quantizing convolutional neural networks for inference with integer weights and activations. Per-channel quantization of weights and per-layer quantization of activations to 8-bits of precision post-training produces classification accuracies within 2% of floating point networks…
The abstract explores connections between reinforcement learning, scaling, and diffusion.
problem Aligning reinforcement learning with human feedback and scaling techniques.
method Clarifying connections between reinforcement learning, scaling, and diffusion.
result Introducing a resampling approach for alignment and reward-directed diffusion models.
Group fairness is an important concern for machine learning researchers, developers, and regulators. However, the strictness to which models must be constrained to be considered fair is still under debate. The focus of this work is on constraining the expected outcome of subpopulations in kernel regression and, in part…
QuantEase optimizes LLMs with CD-based quantization, achieving state-of-the-art performance.
problem Efficiently quantize large language models for deployment.
method Layer-wise quantization using CD-based algorithms with matrix and vector operations.
result State-of-the-art performance in perplexity and zero-shot accuracy.
Quantizes neural networks using frame theory for improved accuracy.
problem Improving neural network efficiency and accuracy through quantization.
method Sigma-Delta (ΣΔ) quantization with finite unit-norm tight frames. result Error bound between original and quantized neural networks derived.
Reproducibility in reinforcement learning is challenging: uncontrolled stochasticity from many sources, such as the learning algorithm, the learned policy, and the environment itself have led researchers to report the performance of learned agents using aggregate metrics of performance over multiple random seeds for a …
Recently, a special type of data poisoning (DP) attack targeting Deep Neural Network (DNN) classifiers, known as a backdoor, was proposed. These attacks do not seek to degrade classification accuracy, but rather to have the classifier learn to classify to a target class whenever the backdoor pattern is present in a tes…