Accurate prediction of energy and forces for 3D molecular systems is one of fundamental challenges at the core of AI for Science applications. Many powerful and data-efficient neural networks predict molecular energies and forces from single atomic configurations. However, one crucial aspect of the data generation process is rarely considered while learning these models i.e. Molecular Dynamics (MD) simulation. MD simulations generate time-ordered trajectories of atomic positions that fluctuate in energy and explore regions of the potential energy surface (e.g., under standard NVE/NVT ensembles), rather than being constructed to steadily lower the potential energy toward a minimum as in geometry relaxations. This work explores a novel way to leverage MD data, when available, to improve the performance of such predictors. We introduce a novel training strategy called FRAMES, that use an auxiliary loss function for exploiting the temporal relationships within MD trajectories. Counter-intuitively, on two atomistic benchmarks and a synthetic system we observe that minimal temporal information, captured by pairs of just two consecutive frames, is often sufficient to obtain the best performance, while adding longer trajectory sequences can introduce redundancy and degrade performance. On the widely used MD17 and ISO17 benchmarks, FRAMES significantly outperforms its Equiformer baseline, achieving highly competitive results in both energy and force accuracy. Our work not only presents a novel training strategy which improves the accuracy of the model, but also provides evidence that for distilling physical priors of atomic systems, more temporal data is not always better.

The training paradigm for large language models (LLMs) is moving from static datasets to experience-based learning, where agents acquire skills via interacting with complex environments. To facilitate this transition we introduce GEM (General Experience Maker), an open-source environment simulator designed for the age of LLMs. Analogous to OpenAI-Gym for traditional reinforcement learning (RL), GEM provides a standardized framework for the environment-agent interface, including asynchronous vectorized execution for high throughput, and flexible wrappers for easy extensibility. GEM also features a diverse suite of environments, robust integrated tools, and single-file example scripts demonstrating using GEM with five popular RL training frameworks. Along with this, we also provide a set of baselines across 24 environments using REINFORCE with Return Batch Normalization (ReBN), which---unlike GRPO---is compatible with the full RL setting of dense per-turn rewards and arbitrary discount factors. We further conduct apple-to-apple benchmarking of PPO, GRPO and REINFORCE in both single- and multi-turn settings using GEM to shed light on the algorithmic designs. GEM also functions as a convenient evaluation toolkit besides a training environment. We hope this framework can help accelerate future agentic LLM research.

LLM parameter editing methods commonly rely on computing an ideal target hidden-state at a target layer (referred as anchor point) and distributing the target vector to multiple preceding layers (commonly known as backward spreading) for cooperative editing. Although widely used for a long time, its underlying basis have not been systematically investigated. In this paper, we first conduct a systematic study of its foundations, which helps clarify its capability boundaries, practical considerations, and potential failure modes. Then, we propose a simple and elegant alternative that replaces backward spreading with forward-propagation. Instead of optimizing the target at the last editing layer, we optimize the anchor point at the first editing layer, and then propagate it forward to obtain accurate and mutually compatible target hidden-states for all subsequent editing layers. This approach achieves the same computational complexity as existing methods while producing more accurate layer-wise targets. Our method is simple, without interfering with either the computation of the initial target hidden state or any other components of the subsequent editing pipeline, and thus constituting a benefit for a wide range of LLM parameter editing methods.

Reinforcement learning (RL) has become a key component of post-training large language models (LLMs). In practice, LLM RL is often off-policy because of training-inference mismatch and policy staleness, making trust-region control essential for stable optimization. Mainstream methods such as PPO and GRPO approximate this control with a ratio-clipping mechanism, but the importance ratio can be a poor proxy for distributional shift in long-tailed vocabularies. Recent work such as DPPO addresses this mismatch by replacing ratio-based clipping with a divergence-based mask, yielding a trust region defined by the sampled token's absolute probability shift. However, DPPO still relies on a hard mask: once a token crosses the trust-region boundary in a harmful direction, its gradient is discarded rather than corrected. To address this, we propose Divergence Regularized Policy Optimization (DRPO), which replaces the hard mask with a smooth advantage-weighted quadratic regularizer on policy shift. DRPO preserves the same trust-region geometry as DPPO while inducing bounded, continuous gradient weights that attenuate diverging updates and provide corrective signals beyond the boundary. Experiments across model scales, architectures, and precision settings show that DRPO improves the stability and efficiency of LLM RL training.

Reinforcement learning (RL) has become a cornerstone for fine-tuning Large Language Models (LLMs), with Proximal Policy Optimization (PPO) serving as the de facto standard algorithm. Despite its ubiquity, we argue that the core ratio clipping mechanism in PPO is structurally ill-suited for the large vocabularies inherent to LLMs. PPO constrains policy updates based on the probability ratio of sampled tokens, which serves as a noisy single-sample Monte Carlo estimate of the true policy divergence. This creates a sub-optimal learning dynamic: updates to low-probability tokens are aggressively over-penalized, while potentially catastrophic shifts in high-probability tokens are under-constrained, leading to training inefficiency and instability. To address this, we propose Divergence Proximal Policy Optimization (DPPO), which substitutes heuristic clipping with a more principled constraint based on a direct estimate of policy divergence (e.g., Total Variation or KL). To avoid huge memory footprint, we introduce the efficient Binary and Top-K approximations to capture the essential divergence with negligible overhead. Extensive empirical evaluations demonstrate that DPPO achieves superior stability and efficiency compared to existing methods, offering a more robust foundation for RL-based LLM fine-tuning.

Large language models (LLMs) often solve reasoning problems by generating intermediate traces that explore and revise partial solutions. From a search perspective, these traces can be viewed as linearized search trees, where the model extends a partial solution, abandons it when it fails, and backtracks to try alternatives. Compared with traditional heuristic-guided search, such a policy has a potential advantage: it conditions on the whole search trace rather than only on the current local state. We first test whether LLMs utilize this advantage by comparing trace-conditioned reasoning policies against best-first search equipped with an LLM heuristic that only observes the current local state. Across three controlled reasoning environments, Blocks World, grid Navigation, and Sokoban, we find that raw access to search history alone is not enough to reliably outperform heuristic search. We then study one possible reason: in LLM reasoning traces, the underlying search tree is only implicitly represented, and when the model backtracks or switches branches, the trace does not explicitly identify which earlier search state is being revisited. We show that adding simple parent pointers to explicitly represent the linearized tree (LinTree) structure improves both task performance and search efficiency relative to implicit reasoning models and LLM-heuristic-guided search. These results suggest that search history becomes most useful when its tree structure is made explicit, motivating more structure-aware representations for LLM reasoning.

Recent advances in reinforcement learning have shown that language models can develop sophisticated reasoning through training on tasks with verifiable rewards, but these approaches depend on human-curated problem-answer pairs and domain-specific reward engineering. We introduce SPIRAL, a self-play framework where models learn by playing multi-turn, zero-sum games against continuously improving versions of themselves, generating an automatic curriculum of stronger opponents, and eliminating the need for human supervision. To enable this self-play training at scale, we implement a fully online, multi-turn, multi-agent reinforcement learning system for LLMs and propose role-conditioned advantage estimation (RAE) to stabilize multi-agent training. SPIRAL produces reasoning capabilities that transfer broadly, improving performance by up to 10\% across a suite of 8 reasoning benchmarks on 4 different models spanning Qwen and Llama model families, outperforming supervised fine-tuning on 25,000 expert game trajectories. Multi-game training (TicTacToe, Kuhn Poker, Simple Negotiation) yields the strongest results, with improvements observed across both base and instruction-tuned models. Analysis of chain-of-thought traces reveals that games develop distinct cognitive patterns that transfer to improve reasoning performance, with different games developing complementary strengths. Even models which have already been trained on reasoning tasks using RLVR, like DeepSeek-R1-Distill-Qwen-7B, still benefit from our approach. These results demonstrate that zero-sum games naturally develop transferable reasoning capabilities across diverse model architectures and training stages, highlighting a promising direction for autonomous reasoning development.

Model editing has recently emerged as a popular paradigm for efficiently updating knowledge in LLMs. A central desideratum of updating knowledge is to balance editing efficacy, i.e., the successful injection of target knowledge, and specificity (also known as edit locality), i.e., the preservation of existing non-target knowledge. However, we find that existing specificity evaluation protocols are inadequate for this purpose. We systematically elaborated on the three fundamental issues it faces. Beyond the conceptual issues, we further empirically demonstrate that existing specificity metrics are weakly correlated with the strength of specificity regularizers. We also find that current metrics lack sufficient sensitivity, rendering them ineffective at distinguishing the specificity performance of different methods. Finally, we propose a constructive evaluation protocol. Under this protocol, the conflict between open-ended LLMs and the assumption of determined answers is eliminated, query-independent fluency biases are avoided, and the evaluation strictness can be smoothly adjusted within a near-continuous space. Experiments across various LLMs, datasets, and editing methods show that metrics derived from the proposed protocol are more sensitive to changes in the strength of specificity regularizers and exhibit strong correlation with them, enabling more fine-grained discrimination of different methods' knowledge preservation capabilities.

Group-based reinforcement learning methods such as GRPO for large language models avoid training a critic by sampling multiple responses for each prompt. A reliable critic could instead estimate token-level advantages from one response, but standard critic-based training recipes are often unstable. We study this instability and develop **Best Practice Critic Optimization (BPCO)**, a recipe that combines DPPO, value predictions bounded to the reward range, Monte Carlo value targets, unnormalized policy advantages, and length-adaptive generalized advantage estimation. Because the critic is used only during training, BPCO can also condition it on reward-defining information, such as a reference answer or grading rubric, that is hidden from the policy. Controlled experiments isolate the effect of each design choice. Across mathematical reasoning tasks with models ranging from 1.5B parameters to 30B-A3B mixtures of experts, BPCO improves a strong critic-based baseline consistently, and matches or exceeds a group-based baseline while sampling one response per prompt. The same recipe also improves learning with rubric-based rewards. These results show that a carefully designed critic provides a reliable alternative to group-relative advantage estimation. Code is available at https://github.com/QPHutu/golden_critic.

At the boundary between the known and the unknown, an agent inevitably confronts the dilemma of whether to explore or to exploit. Epistemic uncertainty reflects such boundaries, representing systematic uncertainty due to limited knowledge. In this paper, we propose a Bayesian reinforcement learning (RL) algorithm, $\texttt{EUBRL}$, which leverages epistemic guidance to achieve principled exploration. This guidance adaptively reduces per-step regret arising from estimation errors. We establish nearly minimax-optimal regret and sample complexity guarantees for a specific class of priors in infinite-horizon discounted MDPs. Empirically, we evaluate $\texttt{EUBRL}$ on tasks characterized by sparse rewards, long horizons, and stochasticity. Results demonstrate that $\texttt{EUBRL}$ achieves superior sample efficiency, scalability, and consistency.

Visual entity tracking is an innate cognitive ability in humans, yet it remains a critical bottleneck for Vision-Language Models (VLMs). This deficit is often obscured in existing video benchmarks by visual shortcuts. We introduce VET-Bench, a synthetic diagnostic testbed featuring visually identical objects that necessitate tracking exclusively through spatiotemporal continuity. Our experiments reveal that current state-of-the-art VLMs perform at or near chance level on VET-Bench, exposing a fundamental limitation: an over-reliance on static frame-level features and a failure to maintain entity representations over time. We provide a theoretical analysis drawing connections to the state-tracking problem, proving that fixed-depth transformer-based VLMs are fundamentally limited in tracking indistinguishable objects without intermediate supervision due to expressivity constraints. To address this, we propose Spatiotemporal Grounded Chain-of-Thought (SGCoT): generating object trajectories as explicit intermediate states. Leveraging Molmo2's object tracking ability, we elicit SGCoT reasoning by fine-tuning on synthesized text-only data for alignment. Our method achieves state-of-the-art accuracy exceeding 90% on VET-Bench, demonstrating that VLMs can reliably solve the video shell-game task end-to-end without external tools. Our code and data are available at https://vetbench.github.io .

Score-based generative models are trained in high-dimensional ambient spaces, yet many data distributions are supported on low-dimensional nonlinear structures. We prove that, for compact d-dimensional smooth manifolds ℳ⊂ [0,1]^D with d > 2 and β-Hölder densities strictly positive on ℳ, a variance-preserving SGM estimator attains the intrinsic Wasserstein–1 sample exponent 𝒪̃(D^𝒪_β(d)n^-(β+1)/(d+2β)), up to logarithmic factors and explicit geometry and density factors. The full nonasymptotic bound explicitly isolates the finite-order geometry envelope, Hölder radius, density lower bound, ambient dependence, and finite-order correction terms. The analysis separates score approximation into a large-noise tangent-cell regime and a small-noise projection-centered, de-Gaussianized Laplace regime. The key technical ingredient is a ReLU implementation of nearest-projection coordinates via finite intrinsic anchors and Gauss–Newton iterations, rather than approximating the manifold projection as a black-box high-dimensional smooth map. Consequently, for families with polynomially controlled geometry and density lower bounds, the constructed score-network parameters have polynomial ambient dependence.

External memory effectively grounds large language models (LLMs) and vision-language models (VLMs)-based question answering (QA) in relevant multimodal evidence. However, existing memory paradigms represent each memory item in raw text and image forms, so retrieval-based systems must pass the retrieved text or images to the generation LLMs/VLMs, resulting in high token consumption and storage pressure, making it unaffordable for resource-constrained applications. We propose Latent Memory, a latent-space memory paradigm that replaces each raw text or image evidence item with a single high-dimensional latent token produced by a small compressor LLM/VLM. Rather than retrieving raw evidence for generation, Latent Memory operates in a unified latent representation space: the query is embedded into this space to retrieve relevant latent tokens, and the retrieved latent tokens are directly prompted to a pretrained LLM or VLM for answer generation. To make each latent token simultaneously informative for reconstruction, retrieval, and generation, we train the compressor with reconstruction, contrastive, and distillation objectives in a unified end-to-end manner. Latent Memory is evaluated on seven text-only QA benchmarks (e.g., HotpotQA) and multimodal QA benchmarks, where it achieves competitive QA performance compared to advanced RAG baselines while consuming 3x to 10x fewer generator tokens. It can also deliver the strongest image-grounded QA performance on WebQA. Code is available at https://github.com/zz1358m/Latent-Memory-Master.

Aiming at efficient and dense chain-of-thought (CoT) reasoning, latent reasoning methods fine-tune Large Language Models (LLMs) to substitute discrete language tokens with continuous latent tokens. These methods consume fewer tokens compared to the conventional language CoT reasoning and have the potential to plan in a dense latent space. However, current latent tokens are generally supervised based on imitating language labels. Considering that there can be multiple equivalent but diverse CoT labels for a question, passively imitating an arbitrary one may lead to inferior latent token representations and latent reasoning policies, undermining the potential planning ability and resulting in clear gaps between training and testing. In this work, we emphasize the importance of active planning over the representation space of latent tokens in achieving the optimal latent reasoning policy. So, we propose the Active Latent Planning method (ATP-Latent), which models the supervision process of latent tokens as a conditional variational auto-encoder (VAE) to obtain a smoother latent space. Moreover, to facilitate the most reasonable latent reasoning policy, ATP-Latent conducts reinforcement learning (RL) with an auxiliary coherence reward, which is calculated based on the consistency between VAE-decoded contents of latent tokens, enabling a guided RL process. In experiments on LLaMA-1B, ATP-Latent demonstrates +4.1% accuracy and -3.3% tokens on four benchmarks compared to advanced baselines. Codes are available on https://github.com/zz1358m/ATP-Latent-master.

In this paper, we consider the setting where large language models (LLMs) are trained using reinforcement learning (RL) to simultaneously improve reasoning accuracy and verbalize its confidence. Our reward scheme uses two functions for rewarding confidence verbalized by the LLM: one when the LLM is correct and a different one when the LLM is incorrect. With a poorly designed reward scheme, the LLM may be incentivized to answer incorrectly so that it can be confident that its answer is indeed incorrect, a phenomenon that we call confidence reward hacking. We propose the concept of non-hackable confidence reward schemes and define a spectrum of such reward schemes for RL confidence calibration training in LLMs. We demonstrate that selective confidence reward hacking can occur in practical datasets with reward schemes that are not designed to be non-hackable. We also demonstrate that the reward scheme with the best calibration to accuracy tradeoff depends on the dataset and the application, and propose using the reward scheme as a hyperparameter to optimize the tradeoffs in accordance to what is important for the application. The code of our experiments is available in https://anonymous.4open.science/r/rl-confidence-calibration-9ED4/README.md.

Reinforcement Learning (RL) has been promising in single-turn LLM fine-tuning. However, long-horizon agentic reasoning introduces increasingly branching interactions and sparse rewards, exposing several limitations of RL: its heavyweight backpropagation-based training stack makes it impractical to fine-tune larger LLMs, and longer-horizon trajectories make credit assignment in RL substantially harder. This paper argues that evolution strategies (ES) can be a better choice for fine-tuning long-horizon LLM agents. Compared with agentic RL, ES offers three key advantages: 1) Model Scalability: ES enables full-parameter optimization with only minimal, inference-level GPU memory, making it possible to fine-tune large LLMs. 2) Flexibility: its lightweight, black-box feedback interface makes ES fine-tuning easy to compose with prompt-space evolution (e.g., skill optimization test-time compute); and 3) Long-Horizon Scalability: ES performs trajectory-level parameter attribution without decomposing rewards across horizons, yielding better scalability than Agentic RL as the horizon length grows. Based on this insight, we propose Agentic ESOpt, a full-parameter agentic fine-tuning framework tailored to flexible parameter–context co-evolution. At each step, Agentic ESOpt samples perturbations around the current LLM parameters, evaluates the resulting agents with rewards, and applies an online reward-weighted update. To improve the exploration–adaptation trade-off, Agentic ESOpt further introduces a cosine decay schedule of the perturbation scale σ. On WebArena-Lite, full-parameter optimization of Qwen-3.5-27B improves the No Skill baseline by 6.69

Claim verification is essential in combating misinformation, and large language models (LLMs) have recently emerged in this area as powerful tools for assessing the veracity of claims using external knowledge. Existing LLM-based methods for claim verification typically adopt a Decompose-Then-Verify paradigm, which involves decomposing complex claims into several independent sub-claims and verifying each sub-claim separately. However, this paradigm often introduces errors during the claim decomposition process. To mitigate these errors, we propose to develop the Chain-of-Thought (CoT)-Verify paradigm, which leverages LLM reasoning methods to generate CoT-verification paths for the original complex claim without requiring decompositions into sub-claims and separate verification stages. The CoT-Verify paradigm allows us to propose a natural fine-tuning method called Reasoning-CV to enhance the verification capabilities in LLMs. Reasoning-CV includes a supervised fine-tuning (SFT) stage and a self-improvement direct preference optimization (DPO) stage. Utilizing only an 8B pre-trained LLM, Reasoning-CV demonstrates superior knowledge-assisted claim verification performances compared to existing Decompose-Then-Verify methods, as well as powerful black-box LLMs such as GPT-4o+CoT and o1-preview. Our code is available.

Large language models have recently demonstrated significant gains in reasoning ability, often attributed to their capacity to generate longer chains of thought and engage in reflective reasoning. However, the contribution of reflections to performance improvement remains unclear. In this paper, we systematically analyze the rollouts of eight reasoning models on five mathematical datasets. We focus on reflective behaviours where the model has already produced an answer but continues reflecting before finalizing its output. Our analysis reveals that reflections are predominantly confirmatory and rarely alter the model's initial answer, a pattern consistent across models and datasets. To understand the role of reflections in training, we construct supervised fine-tuning (SFT) datasets with varying amounts of reflection steps. We observe that training models on rollouts with more reflection steps primarily enhances first-answer correctness rather than the ability to correct initially wrong answers through reflections. This motivates us to propose a question-aware early-stopping method that enhances inference-time token efficiency by stopping the reasoning process once a few plausible candidate answers are generated, thereby reducing unnecessary reflection steps. Motivated by this, we further propose to dynamically truncate the reflections after a candidate answer has appeared during generation, which reduces reasoning tokens by 24.5

Continual reinforcement learning (CRL) refers to a naturalistic setting where an agent needs to endlessly evolve, by trial and error, to solve multiple tasks that are presented sequentially. One of the largest obstacles to CRL is that the agent may forget how to solve previous tasks when learning a new task, known as catastrophic forgetting. In this paper, we propose to address this challenge by planning with online world models. Specifically, we learn a Follow-The-Leader shallow model online to capture the world dynamics, in which we plan using model predictive control to solve a set of tasks specified by any reward functions. The online world model is immune to forgetting by construction with a proven regret bound of O(root K(2)Dlog(T)) under mild assumptions. The planner searches actions solely based on the latest online model, thus forming a FTL Online Agent (OA) that updates incrementally. To assess OA, we further design Continual Bench, a dedicated environment for CRL, and compare with several strong baselines under the same model-planning algorithmic framework. The empirical results show that OA learns continuously to solve new tasks while not forgetting old skills, outperforming agents built on deep world models with various continual learning techniques.

The training paradigm for large language models (LLMs) is moving from static datasets to experience-based learning, where agents acquire skills via interacting with complex environments. To facilitate this transition we introduce GEM (General Experience Maker), an open-source environment simulator designed for the age of LLMs. Analogous to OpenAI-Gym for traditional reinforcement learning (RL), GEM provides a standardized framework for the environment-agent interface, including asynchronous vectorized execution for high throughput, and flexible wrappers for easy extensibility. GEM also features a diverse suite of environments, robust integrated tools, and single-file example scripts demonstrating using GEM with five popular RL training frameworks. Along with this, we also provide a set of baselines across 24 environments using REINFORCE with Return Batch Normalization (ReBN), which -- unlike GRPO -- is compatible with the full RL setting of dense per-turn rewards and offers better credit assignment. We further conduct apple-to-apple benchmarking of PPO, GRPO and REINFORCE in both single- and multi-turn settings using GEM to shed light on the algorithmic designs. Lastly, GEM also functions as a convenient evaluation toolkit besides a training environment. We hope this framework can help accelerate future agentic LLM research.