Modern RL post-training methods such as GRPO and DAPO train on N response sequences of R tokens sampled from a shared prompt of P tokens, but standard FlashAttention replicates all P prompt tokens N times across both forward and backward passes -- duplicating compute and memory on identical hidden states. In large-rollout, long-context RL training (N>=16, P>=8K), this redundancy dominates the policy update cost. We observe that in decoder-only models, causal masking makes prompt representations invariant across sequences at every layer, so all per-token operations (norms, projections, MLP) and attention can process the prompt once -- a property not yet exploited at the kernel level for training. We propose DualKV, the first FlashAttention kernel variant that eliminates shared-prompt replication during RL training, via (1) fused CUDA forward and backward kernels that iterate over two disjoint KV regions -- shared context and per-sequence response -- in a single kernel launch, and (2) a data-pipeline redesign in veRL that repacks N(P+R) tokens into P+NR tokens per micro-batch, extending the token reduction from attention to the entire model by a factor rho = N(P+R)/(P+NR). DualKV is mathematically equivalent to standard attention and introduces no approximation. On Qwen3-8B GRPO training with 8xH100 GPUs (N=32, 8K-context), DualKV achieves 1.63--2.09x policy-update speedup, enables 2x larger micro-batches, and raises MFU from 36% to 76%. Similar gains hold for DAPO (2.47x speedup, 77% MFU). At 30B MoE scale on 16xH100, DualKV achieves 3.82x policy-update and 3.38x end-to-end step speedup over FlashAttention (which requires 4-way Ulysses sequence parallelism to avoid OOM). DualKV also extends to hybrid sliding/global attention with head dimension 512 (which FA2 does not support) and integrates with Ulysses sequence parallelism, demonstrated on Gemma-4-31B GRPO at 64K context.
Test-time scaling often uses an external verifier, such as compilers and test cases in coding or trained value functions in robotics applications, to obtain high-quality rollouts. Verifier-free test-time scaling (or VF-TTS) is gaining extensive attention as a mechanism to enhance Large Language Model (LLM) reasoning, primarily because we do not have access to such high-quality verifiers in many real-world applications. Among existing VF-TTS methods, confidence-based VF-TTS methods, which compute and rank rollouts solely by confidence, are particularly promising. Such methods introduce near-zero overhead for sample evaluation and require minimal access to internal model states, making the methods highly flexible across models and tasks. In this paper, we demonstrate a critical limitation of existing confidence-based VF-TTS methods by showing that such methods catastrophically break down on complex tasks. We observe a very interesting phenomenon: uniformly high confidence frequently indicates a failure to explore, favoring confidently wrong answers. To address this, our core insight is that robust cognitive search requires a specific confidence trajectory pattern: such methods perform exploratory branching at the beginning, as manifested by low initial confidence, and converge to a high final confidence solution. To implement this insight, we introduce consilience, a novel selection framework that explicitly evaluates the temporal asymmetry of confidence in reasoning. We operationalize this via a combinatorial metric that actively penalizes high initial confidence while strictly demanding final certainty. Extensive experiments covering both graduate-level mathematics problems and free-form code generation demonstrate that consilience effectively outperforms existing baselines, validating our novel perspective on completion confidence.
Developing agents capable of exploring, planning and learning in complex open-ended environments is a grand challenge in artificial intelligence (AI). Hierarchical reinforcement learning (HRL) offers a promising solution to this challenge by discovering and exploiting the temporal structure within a stream of experience. The strong appeal of the HRL framework has led to a rich and diverse body of literature attempting to discover a useful structure. However, it is still not clear how one might define what constitutes good structure in the first place, or the kind of problems in which identifying it may be helpful. This work aims to identify the benefits of HRL from the perspective of the fundamental challenges in decision-making, as well as highlight its impact on the performance trade-offs of AI agents. Through these benefits, we then cover the families of methods that discover temporal structure in HRL, ranging from learning directly from online experience to offline datasets, to leveraging large language models (LLMs). Finally, we highlight the challenges of temporal structure discovery and the domains that are particularly well-suited for such endeavours.
Retrieving the biological impacts of protein-protein interactions (PPIs) is essential for target identification (Target ID) in drug development. Given the vast number of proteins involved, this process remains time-consuming and challenging. Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) frameworks have supported Target ID; however, no benchmark currently exists for identifying the biological impacts of PPIs. To bridge this gap, we introduce the RAG Benchmark for PPIs (RAGPPI), a factual question-answer benchmark of 4,420 question-answer pairs that focus on the potential biological impacts of PPIs. Through interviews with experts, we identified criteria for a benchmark dataset, such as a type of QA and source. We built a gold-standard dataset (500 QA pairs) through expert-driven data annotation. We developed an ensemble auto-evaluation LLM that reflected expert labeling characteristics, which facilitates the construction of a silver-standard dataset (3,720 QA pairs). We are committed to maintaining RAGPPI as a resource to support the research community in advancing RAG systems for drug discovery QA solutions.
Classical federated learning (FL) assumes that the clients have a limited amount of noisy data with which they voluntarily participate and contribute towards learning a global, more accurate model in a principled manner. The learning happens in a distributed fashion without sharing the data with the center. However, these methods do not consider the incentive of an agent for participating and contributing to the process, given that data collection and running a distributed algorithm is costly for the clients. The question of rationality of contribution has been asked recently in the literature and some results exist that consider this problem. This paper addresses the question of simultaneous parameter learning and incentivizing contribution, which distinguishes it from the extant literature. Our first mechanism incentivizes each client to contribute to the FL process at a Nash equilibrium and simultaneously learn the model parameters. However, this equilibrium outcome can be away from the optimal, where clients contribute with their full data and the algorithm learns the optimal parameters. We propose a second mechanism with monetary transfers that is budget balanced and enables the full data contribution along with optimal parameter learning. Large scale experiments with real (federated) datasets (CIFAR-10, FeMNIST, and Twitter) show that these algorithms converge quite fast in practice, yield good welfare guarantees, and better model performance for all agents.
We propose a new approach to falsify causal discovery algorithms without ground truth, which is based on testing the causal model on a pair of variables that has been dropped when learning the causal model. To this end, we use the "Leave-One-Variable-Out (LOVO)" prediction where $Y$ is inferred from $X$ without any joint observations of $X$ and $Y$, given only training data from $X,Z_1,\dots,Z_k$ and from $Z_1,\dots,Z_k,Y$. We demonstrate that causal models on the two subsets, in the form of Acyclic Directed Mixed Graphs (ADMGs), often entail conclusions on the dependencies between $X$ and $Y$, enabling this type of prediction. The prediction error can then be estimated since the joint distribution $P(X, Y)$ is assumed to be available, and $X$ and $Y$ have only been omitted for the purpose of falsification. After presenting this graphical method, which is applicable to general causal discovery algorithms, we illustrate how to construct a LOVO predictor tailored towards algorithms relying on specific a priori assumptions, such as linear additive noise models. Simulations indicate that the LOVO prediction error is indeed correlated with the accuracy of the causal outputs, affirming the method's effectiveness.
Instruction following is a key capability for LLMs. However, recent studies have shown that LLMs often struggle with instructions containing multiple constraints (e.g. a request to create a social media post "in a funny tone" with "no hashtag"). Despite this, most evaluations focus solely on synthetic data. To address this, we introduce RealInstruct, the first benchmark designed to evaluate LLMs' ability to follow real-world multi-constrained instructions by leveraging queries real users asked AI assistants. We also investigate model-based evaluation as a cost-effective alternative to human annotation for this task. Our findings reveal that even the proprietary GPT-4 model fails to meet at least one constraint on over 21% of instructions, highlighting the limitations of state-of-the-art models. To address the performance gap between open-source and proprietary models, we propose the Decompose, Critique and Refine (DeCRIM) self-correction pipeline, which enhances LLMs' ability to follow constraints. DeCRIM works by decomposing the original instruction into a list of constraints and using a Critic model to decide when and where the LLM's response needs refinement. Our results show that DeCRIM improves Mistral's performance by 7.3% on RealInstruct and 8.0% on IFEval even with weak feedback. Moreover, we demonstrate that with strong feedback, open-source LLMs with DeCRIM can outperform GPT-4 on both benchmarks.
Recent studies have proposed different methods to improve multilingual word representations in contextualized settings including techniques that align between source and target embedding spaces. For contextualized embeddings, alignment becomes more complex as we additionally take context into consideration. In this work, we propose using Optimal Transport (OT) as an alignment objective during fine-tuning to further improve multilingual contextualized representations for downstream cross-lingual transfer. This approach does not require word-alignment pairs prior to fine-tuning that may lead to sub-optimal matching and instead learns the word alignments within context in an unsupervised manner. It also allows different types of mappings due to soft matching between source and target sentences. We benchmark our proposed method on two tasks (XNLI and XQuAD) and achieve improvements over baselines as well as competitive results compared to similar recent works.