#training

40 curated events

papersTODAY 04:00 UTC

Stopping and restarting strategy speeds up multi-turn agentic on-policy distillation

A new arXiv paper addresses the high cost of on-policy distillation, which relies on expensive autoregressive rollouts by the student model and scales poorly when tasks span multiple turns. The authors propose deciding when to halt a rollout and where to resume it, aiming to cut the compute spent on generating student trajectories. The method targets more efficient transfer of capabilities from large teacher models to smaller students in agentic settings.

distillation/agents/efficiency/training

papersTODAY 04:00 UTC

Study analyzes continuous-time dropout as a random batch method for controlled differential equations

This paper examines continuous-time dropout within controlled differential equations by approximating additive vector fields with random batches. Over each interval of length h, a random subset of components is activated and rescaled according to its inclusion probability. The work covers convergence, design, and training aspects of the approach.

research/dropout/differential-equations/training

papersTODAY 04:00 UTC

Local learning rule trains generative thermodynamic computers

A new preprint describes training generative thermodynamic computers, which convert thermal noise into structured data via Langevin dynamics, using an update applied locally at each integration step. The authors derive the coupling gradient from a reverse-path Onsager-Machlup objective rather than relying on global backpropagation through time. The work appears on arXiv under the machine learning category as a cross-listed submission.

research/thermodynamic-computing/training/arxiv

papersTODAY 04:00 UTC

Conflict-Free Gradients Target Failure Modes in PINNs and PIKANs

A new arXiv preprint examines why physics-informed neural networks (PINNs) and their Kolmogorov-Arnold counterparts (PIKANs) often fail when used with domain decomposition to solve partial differential equations over complex geometries. The authors attribute these problems to conflicting gradient signals and propose a conflict-free gradient approach to improve training stability and scalability.

research/pinns/scientific-ml/training

papersTODAY 04:00 UTC

Gradient Repair Method Aimed at Stabilizing Neural ODE Training

A new arXiv paper introduces GradRepair-ODE, a technique that certifies and repairs gradients when neural ordinary differential equations are trained. Because neural ODEs embed a numerical solver in the training loop, solver choices shape both the forward trajectory and the gradients sent to the optimizer, which the authors flag as a reliability issue for scientific machine learning. The approach is presented as a way to keep those gradients trustworthy during training.

neural-ode/training/research/scientific-ml

papersTODAY 04:00 UTC

Bellman Policy Optimization: Critic-Free RL Method for LLM Reasoning

Researchers present Bellman Policy Optimization (BPO), a reinforcement learning approach for training large language models with verifiable rewards that does not require a separate critic network. The method is derived from Policy Mirror Descent and targets autoregressive generation. It aims to improve reasoning performance in LLMs while simplifying the training setup.

reinforcement-learning/llm-reasoning/research/training

papersTODAY 04:00 UTC

Paper Proposes Hindsight-Anchored Policy Optimization for LLM Reasoning

A new arXiv paper introduces Hindsight-Anchored Policy Optimization, a method for training large language models with verifiable rewards. It uses hindsight learning combined with a Thompson sampling-inspired adaptive gate to address cold-start problems in sparse-reward, on-policy training. The approach builds on mixed-policy methods that blend off-policy and on-policy data.

reinforcement-learning/llm-reasoning/arxiv/training

papersTODAY 04:00 UTC

MACCHIATO training method targets certifiably interpretable ReLU-MLP Boolean models

A new arXiv paper introduces MACCHIATO, a specialized training algorithm for ReLU-based multilayer perceptrons that solve Boolean tasks. The method is designed to produce networks that are certifiably interpretable while guaranteeing correct generalization across the full truth table. The authors frame the work as a response to the growing gap between AI capability and the ability to explain model behavior.

interpretability/training/arxiv/research

papersTODAY 04:00 UTC

Initial Weight Scale Found to Control Feature Specialization in ReLU Networks

A research paper examines how the scale of a network's initial weights influences which features individual ReLU neurons end up learning during training. The authors argue that this initialization-dependent effect, framed as a hidden gauge, helps explain why trained networks organize representations the way they do.

research/neural-networks/training/arxiv

papersTODAY 04:00 UTC

arXiv Paper Analyzes Optimal Learning Rate Schedules Under Functional Scaling Laws

A new arXiv preprint examines how learning rate schedules can be optimized within the functional scaling law framework, which separates training dynamics into signal learning and noise forgetting. The authors analyze power-law kernel regression to characterize these two components, comparing schedules such as power decay and warmup-stable-decay. The work offers theoretical guidance on choosing learning rate schedules for model training.

training/optimization/scaling-laws/learning-rate

papersTODAY 04:00 UTC

Concertina: Adaptive Pipeline Parallelism for Long-Context LLM Training

A new arXiv paper introduces Concertina, a data-centric approach to pipeline parallelism aimed at making long-context LLM training more efficient. The authors argue that sequence parallelism carries heavy communication costs, and that pipeline parallelism's benefits depend on how the model and data are partitioned. Concertina adaptively splits pipeline stages to better handle long sequences on heterogeneous hardware.

long-context/training/pipeline-parallelism/distributed-training

papersTODAY 04:00 UTC

Minibatch persistency revisited: trade-offs in steps, energy and data use

A new arXiv paper re-examines minibatch persistency, a technique where the same batch is reused for K consecutive optimizer steps rather than sampling fresh data each time. The authors quantify what batch reuse costs in terms of optimization steps and energy consumption, and what it saves in data throughput. The work revisits a long-standing objection to the method, which was folded into data echoing in 2019.

training/efficiency/batch-reuse/arxiv

papersTODAY 04:00 UTC

Mind2Dialogue Framework Trains Language Models to Model User Mental States

A new arXiv paper introduces Mind2Dialogue, a method that trains language models to be more aware of the people they interact with by simulating users' mental states during training. The authors frame the problem as a supervision gap: models need signals about human internal states, which are rarely available in standard dialogue data. The approach targets long-term collaboration in learning, reasoning, and decision-making tasks.

llm/dialogue/user-modeling/training

papersTODAY 04:00 UTC

Convergence rate analysis of generative drifting flows

A new arXiv paper examines whether drifting models, which learn a gradual transport process during training but generate samples in a single step, can converge quickly to a target distribution. The authors identify obstructions to fast convergence at fixed scale and propose a multihead approach that improves convergence rates. The work is theoretical, focused on the training dynamics rather than a deployed system.

generative-models/training/research/theory

papersTODAY 04:00 UTC

Learning to Coach: Training an LLM to Distill Guidance From Experience

A new arXiv paper introduces Learning to Coach (L2C), a framework that trains a separate LLM acting as a coach to pull actionable guidance out of experience. The motivation is that raw solution trajectories are typically long and noisy, which limits how well language models can learn from them. The approach aims to convert such trajectories into more useful, condensed coaching signals.

llm/training/arxiv/agents

papersTODAY 04:00 UTC

Privileged Observations Speed Up Physical-World Reinforcement Learning Policy Discovery

This paper investigates how extra, non-deployable information about a physical system's state influences how quickly and reliably a reinforcement learning agent learns effective control policies when trained on real hardware instead of simulation. The experiments use a cylinder on a tabletop water channel as the test setup. It is a revised preprint posted to arXiv's artificial intelligence and machine learning sections.

reinforcement-learning/robotics/arxiv/training

papersTODAY 04:00 UTC

arXiv paper examines midtraining stage as a way to control how LLM traits generalize

A new arXiv preprint explores whether midtraining, a stage between pretraining and post-training, can influence which behaviors a large language model carries forward. The authors propose a method called Inoculation Midtraining, which uses invented words to shape how desirable and undesirable properties generalize. The work is a research contribution and has not been peer reviewed.

research/training/safety/llm

papersTODAY 04:00 UTC

ThinkPrior: Zero-Rollout Difficulty Priors for Cold-Start Prompt Selection in RLVR

A revised arXiv paper proposes ThinkPrior, a method for choosing cold-start prompts in reinforcement learning with verifiable rewards without running any rollouts first. The authors observe that the KL-free reward-advantage term used in group relative policy optimization depends on how much reward varies within a group of rollouts, which collapses when a prompt is uniformly easy or hard. They use difficulty priors to pick prompts that are likely to produce useful within-group variation, aiming to make early training more efficient.

rlvr/grpo/prompt-selection/training

papersTODAY 04:00 UTC

Adaptive Language Sampling Method Targets Cross-Lingual Transfer for Low-Resource Languages

A new arXiv paper proposes an online adaptive sampling strategy for realigning multilingual language models, aiming to improve cross-lingual transfer to extremely low-resource languages. The authors note that existing realignment approaches typically use uniform or random sampling, which may underuse informative language pairs. Their method adjusts sampling dynamically as training proceeds within a distributed setup.

multilingual/low-resource/training/arxiv

papersTODAY 04:00 UTC

arXiv paper proposes training paradigm for fast long video generation

A new arXiv preprint addresses the difficulty of generating coherent minute-long videos, noting that while short clips are plentiful and high quality, long-form training data is scarce and confined to a few domains. The authors propose a training approach that combines mode-seeking and mean-seeking objectives to speed up long video generation. The work is positioned as a way to overcome the data bottleneck that limits scaling from seconds to minutes.

video-generation/research/training/generative-models

papersTODAY 04:00 UTC

Paper combines process supervision with outcome-based credit for agent RL

A new arXiv preprint addresses a weakness in outcome-based reinforcement learning for language-model agents: because the whole trajectory receives a single advantage signal, individual decisions get only coarse credit over long interaction sequences. The authors propose reconciling process supervision with outcome-based credit, drawing on on-policy self-distillation to produce finer-grained guidance. The work is presented as a revised submission and targets long-horizon agent training.

agents/reinforcement-learning/training/arxiv

tipsTODAY 08:00 UTC

iX webinar to cover AI Act labeling, AI literacy and high-risk rules

Heise's iX is hosting a compact webinar on the EU AI Act, aimed at helping companies understand what the regulation requires of them. Sessions cover labeling duties, staff AI competence, and obligations tied to high-risk AI systems, with a focus on practical implementation. It is a paid training offer rather than a regulatory announcement.

eu-regulation/ai-act/compliance/training

papersSEP 11 04:00 UTC

Prediction-Loss Alignment for Sampler-Robust Flow Matching Training

The paper looks at a widely used training recipe for diffusion and flow-matching models, where the network predicts a clean sample that is then converted into a velocity for the loss. That conversion amplifies errors near the endpoints of the noise schedule, making training unstable and tied to a particular sampler. The authors propose aligning the prediction objective with the loss objective so that training stays robust regardless of which sampler is used at inference.

flow-matching/diffusion/training/arxiv

papersSEP 11 04:00 UTC

DP-Muon: Differentially Private Optimization with Matrix-Orthogonalized Momentum

A new arXiv paper introduces DP-Muon, an optimizer that combines matrix-orthogonalized momentum with differential privacy guarantees. The method relies on standard per-example gradient clipping and releases one Gaussian-noised gradient per step, treating the matrix and auxiliary updates as post-processing. The authors present a convergence analysis for this approach.

differential-privacy/optimization/research/training

policySEP 13 09:00 UTC

UK government pilot AI bootcamps in Preston target youth unemployment

A government-backed pilot in Preston is running three-week AI training courses for young people who are not in employment, education or training. Participants who complete the course are offered an apprenticeship at the end. The scheme aims to test whether AI skills training can help tackle two pressing economic problems at once.

uk/training/apprenticeships/youth-unemployment

papersSEP 10 04:00 UTC

Study examines trade-off between forecast horizon length and learnability in autoregressive models

A new study investigates how far into the future autoregressive models should be trained when forecasting dynamical systems. The authors identify a trade-off between predictive performance and learnability as the training horizon grows, suggesting an optimal horizon exists. The findings offer practical guidance for selecting prediction horizons during model training.

forecasting/time-series/autoregressive/training

papersSEP 10 04:00 UTC

CompassOPD adapts on-policy distillation to cross-family model pairs

New research introduces CompassOPD, a method that extends on-policy distillation to settings where the teacher and student models come from different families. It derives within-family likelihood shifts to provide dense token-level supervision on student-generated outputs, tackling the effectiveness drop standard OPD exhibits in cross-family scenarios.

distillation/training/llms

papersSEP 10 04:00 UTC

Zone of Proximal Policy Optimization: teacher guidance via prompts, not gradients

A new arXiv paper argues that knowledge distillation breaks down when the student model is much smaller than its teacher, because imitating the teacher's logits locks the student into its sharpest output modes and harms generalization. The authors propose letting the large teacher guide the small student through prompts during reinforcement-learning fine-tuning instead of through gradient-based distillation. The work appears in the computational linguistics category on arXiv.

distillation/reinforcement-learning/small-models/training

papersSEP 10 04:00 UTC

Study proposes token-trimming approach to supervised fine-tuning for math reasoning

A new arXiv paper argues that standard supervised fine-tuning applies its loss uniformly across all tokens, even though some are already mastered and others carry far more useful learning signal for mathematical reasoning. The authors introduce a token-trimming perspective that prioritizes which tokens a model should actually learn during fine-tuning, aiming to avoid over-sharpening well-understood tokens while strengthening the ones that matter most.

fine-tuning/math-reasoning/training/sft

papersSEP 10 04:00 UTC

Muon-C: Adapting the Muon Optimizer to Convolutional Kernels via Operator Alignment

A new arXiv paper presents Muon-C, a variant of the Muon optimizer designed specifically for convolutional layers. The authors argue that applying Muon through standard matrix unfolding misrepresents the geometry of convolution, since it describes a local patch mapping rather than the convolution operator itself. Their approach instead computes the orthogonalized update direction in a way that aligns with the operator's true structure.

optimizers/deep-learning/convnets/training

papersSEP 10 04:00 UTC

Edge-of-chaos initialization fails for higher input derivatives in wide networks

New research indicates that while the edge-of-chaos initialization scheme keeps first-order input perturbations stable in very wide randomly initialized networks, higher-order input derivatives become unstable under the same setup. Because techniques such as physics-informed losses, score matching, and derivative regularization rely on those higher derivatives, the results expose a gap in how such networks should be initialized for derivative-based training. The analysis focuses on smooth fully connected networks with scalar inputs.

theory/neural-networks/initialization/training

papersSEP 10 04:00 UTC

Paper links active learning with lottery ticket sparsity for efficient network training

A newly posted arXiv paper explores combining active learning with the lottery ticket hypothesis, asking whether sparse subnetworks that match dense-model accuracy can be found while also reducing labeled data needs. The authors frame this as achieving both sparsity and sample efficiency within a single training process rather than running separate search stages. The work appears in both the machine learning and AI categories of arXiv.

sparsity/active-learning/efficiency/training

papersSEP 10 04:00 UTC

BRACE paper proposes anchored Bellman-residual correction for stale critics in asynchronous RL

A new arXiv preprint introduces BRACE, a method aimed at value-function staleness in asynchronous reinforcement learning. As training of language models increasingly relies on asynchronous setups, delays between acting and learning bias the critic toward outdated policies, while prior asynchronous-training fixes targeted only the actor. The proposed approach applies an anchored Bellman-residual correction to keep the critic aligned with the current policy.

reinforcement-learning/llm-training/training

papersSEP 10 04:00 UTC

Study analyzes SGD-based learning with synthetic data in high-dimensional linear regression

A newly cross-listed arXiv paper investigates how stochastic gradient descent behaves when training combines human-generated and synthetic data in a high-dimensional linear regression setting. It engages with prior work on model collapse, a phenomenon where keeping even a fixed share of synthetic samples stops model performance from improving as training scales. The findings aim to clarify the conditions under which synthetic data can genuinely extend training beyond limited human datasets.

synthetic-data/model-collapse/training/theory

papersSEP 10 04:00 UTC

Gradient-guided Gaussian adaptive sampling proposed for training physics-informed neural networks

A new arXiv paper introduces 3GAS-PINNs, a variant of physics-informed neural networks that uses gradient-guided Gaussian adaptive sampling to place collocation points. The method targets common weaknesses in PINNs on nonlinear partial differential equations, such as slow convergence, gradient imbalance, and poor resolution of demanding regions. By concentrating sampling where it matters most, the approach aims to improve training efficiency and solution accuracy.

pinns/pdes/adaptive-sampling/training

papersSEP 12 04:00 UTC

Ecdysis Method Trains Runtime Harnesses for LLM Agents More Efficiently

A new arXiv paper introduces Ecdysis, a technique for training self-evolving runtime harnesses that govern how LLM agents execute tasks. The authors argue that prior harness-evolution approaches depend on costly iterative search, and their method aims to make this process both more efficient and more effective. The work targets improved agent performance while reducing the compute needed to discover good harness configurations.

agents/llm-agents/training/research

papersSEP 11 04:00 UTC

CAT-GS Framework Targets Instability in Multimodal Neural Network Training

A new arXiv paper introduces CAT-GS, a training approach that combines calibrated gating with a "fusion surgery" technique for multimodal neural networks. The authors identify three linked failure modes in end-to-end multimodal training, including one modality dominating optimization and unstable dynamics. The method aims to balance learning across modalities and stabilize training.

multimodal/training/arxiv/neural-networks

papersSEP 12 04:00 UTC

DiaLLM paper addresses gap between dialect understanding and generation in LLMs

A new arXiv paper introduces DiaLLM, a method aimed at closing the gap between how well large language models comprehend dialectal English and how poorly they generate it. The authors note that models still tend to output standard, US-centric English even when they understand dialectal input, leaving dialect generation largely unsolved. The work frames generation as the harder half of dialect adaptation and proposes continual training to address it.

dialects/llm/training/research

papersSEP 12 04:00 UTC

ExpTest Uses Loss-Curve Hypothesis Testing to Pick Learning Rates Automatically

A research paper proposes ExpTest, a method that selects learning rates for deep neural networks on its own by statistically testing the shape of the training loss curve. The approach aims to reduce the manual searching and expensive grid searches that currently make hyperparameter tuning costly and less accessible. It is presented as part of ongoing work on arXiv in the cs.AI category.

research/training/hyperparameters/optimization

papersSEP 12 04:00 UTC

arXiv paper proposes unified per-token gating family for on-policy distillation

A new arXiv preprint introduces a family of per-token gating methods for on-policy knowledge distillation that mixes forward and reverse KL losses. The authors argue that prior approaches such as EOPD and ToDi each rely on a single fixed gating signal, and their framework generalizes these with multi-channel and bias coefficients. The work is a methodological contribution aimed at improving how distillation losses are weighted per token during training.

distillation/arxiv/training/kl-divergence