Tag
ZPPO introduces a replay buffer for hard questions in reinforcement learning for LLMs/VLMs, allowing repeated exposure to gradually improve rollout accuracy without policy drift. The method graduates more hard questions than GRPO, especially those with near-zero initial accuracy.
OpenAI researchers show that reinforcement learning on realistic scenarios targeting beneficial traits (honesty, transparency, corrigibility) produces broad improvements across dozens of alignment benchmarks, with gains generalizing beyond training domains and persisting under adversarial pressure.
OpenAI releases research on reinforcement learning for training models to exhibit beneficial traits like honesty and corrigibility, showing that such training generalizes across domains and persists under adversarial pressure.
This paper proposes ImpSH, a triplet-based framework for implicit hate speech classification that aligns posts with implied statements and uses context-bounded semi-hard negative mining to improve cross-dataset generalization.
The paper investigates whether weight norm directly controls the grokking delay in neural networks or if its effect is mediated by logit scale and softmax saturation under cross-entropy loss. Experiments show that the delay is almost entirely explained by the effective logit scale, with weight norm contributing negligibly.
This paper studies how self-driving car systems and humans perform on visual question answering tasks across different geographic locations (Lima and New York City), finding that both humans and VLMs show similar performance regardless of location but diverge based on question type.
This paper presents Connect the Dots (CoD), a framework for training LLMs via reinforcement learning to develop meta-capabilities for long-lifecycle agents, enabling continuous learning and cross-domain generalization.
The paper proposes that grokking in deep neural networks arises from noise-driven escape from metastable phases in first-order L2 phase transitions, demonstrating that delayed generalization follows Arrhenius scaling and reproduces canonical grokking curves.
SpeechDx is a large-scale benchmark for clinical speech AI spanning 12 datasets and 27 tasks across diverse health conditions, structured by stages of speech production. It evaluates 12 state-of-the-art audio encoders and shows that current models do not generalize reliably across the clinical speech landscape.
This paper examines whether language models can independently discover the concept of zero as a form of out-of-distribution generalization, finding that GPT-2 sized models cannot at test time but improve with training on examples of zero, and that language pretraining reduces the number of required examples.
This paper proposes CoTE-SQL, a self-enhanced fine-tuning framework for text-to-SQL that integrates self-reasoning traces, structured chain-of-thought prompting, and execution feedback to achieve state-of-the-art performance on Spider and Bird benchmarks.
Presents a Transformer-based scheduling policy trained with reinforcement learning for the open shop scheduling problem, showing that a model trained on small instances can generalize to much larger problems and compete with classical dispatching heuristics.
This paper proposes sparsity-induced adaptations to LoRA, including Cheap LoRA (cLA) and a chained circulant variant (c³LA), and provides theoretical generalization bounds along with empirical evaluations showing up to 10% training time reduction and 15% peak GPU memory savings while maintaining competitive performance.
This paper demonstrates that the weight norm causally controls the timescale of grokking in neural networks, reconciling conflicting accounts. Through interventions, it shows that grokking follows an exponential delay law and that norm magnitude dominates grokking time over learning rate across architectures.
This paper investigates how post-training stages such as continued pre-training, supervised fine-tuning, and reinforcement learning affect generalization in biological reasoning models, finding that these stages have distinct impacts on in-domain and out-of-domain performance.
This paper introduces a retrieval-augmented vision-language-action policy that eliminates per-task fine-tuning by using pre-trained models with indexed demonstrations, enabling efficient cross-embodiment generalization and task adaptation at test time.
Singular Learning Theory (SLT) uses algebraic geometry to explain why neural networks generalize well despite their degeneracies, introducing the real log canonical threshold (RLCT) as a measure of model complexity.
This paper argues that for large enough models, unfiltered data can improve generalization by providing weak perturbations, contrary to the common assumption that only high-quality filtered data is beneficial. The authors caution that harmful conditional shifts can still damage models, but over-curation may remove useful perturbations.
Fine-tuning small LLMs (3B-7B) with QLoRA on biomedical claim verification achieves higher F1 than GPT-4o and GPT-5 at 44.5x lower cost, and reveals a structural artifact in SciFact. The study demonstrates robust cross-domain transfer when training on structurally sound data.
A detailed explanation of why training on benchmarks, evals, or test sets is a cardinal sin in ML, corrupting the ability to measure generalization. The article emphasizes the importance of clean evaluation protocols and warns against benchmaxxing.