Advantage

Vision-language-action models (VLAs) have trans- formed the field of robotic manipulation in recent years by combining the semantic understanding of LLMs with the precise control of flow-matching policies. Advantage conditioning is a recent technique that iteratively improves VLAs by training a value function on deployment data and using this to train an advantage-conditioned policy. Previous works have only applied simple, low-information success/failure rewards, which leave the value function unable to distinguish states of differing quality beyond how far along the task they appear. Motivated by an exploration of out-of-distribution (OOD) detection methods, we introduce Distance-based Advantage Learning (DistAL), which, by using an embedding space distance as a reward, produces a more informative value function and subsequently a higher downstream task success rate. We validate our method on a series of simulation benchmarks and dexterous bi-manual manipulation tasks on real hardware.

Warm-start transfer can make algorithmic tasks generalize rapidly, yet it is unclear which model components provide the gain and whether that gain remains stable under continued optimization. We study cross-operator transfer on modular arithmetic and separate efficacy (early velocity) from stability (post-reach drawdown). In a scale-matched 108-run battery across 12 seed blocks (96-run 2^3 factorial plus 12-run scale control), transferring internal attention/MLP weights (B) alongside token embeddings and readout (E+U) improves early accuracy by 5.46 pp (Holm p=0.0039) and cuts confirmation latency by 558 steps (Holm p=0.0088). While readout plus internal-block transfer satisfies the pre-specified +/-500-step latency equivalence criterion in 1-layer models (TOST p=0.0011, though Full is faster in 11/12 paired seeds), a prospective 2-layer replication confirms the internal-block advantage (12/12 seeds, +704.67 integral units, p=4.88x10^-4) while revealing an architectural boundary: omitting donor embeddings falls 4475.6 units below Full, outside the +/-250-unit margin. Continued target training frequently triggers severe post-grokking relapse. Freezing transferred representation carriers (E, U) nearly eliminates offline relapse (19.40% -> 0.07%, Holm p=0.005859). Online validation-triggered gating slashes True Max Drawdown from 22.06% to 0.60% on 2a+b (p=0.000488), with prospective confirmations extending protection across affine, nonlinear quadratic, and 2-layer targets (10.94-23.47 pp reductions), distinguishing continual stabilization from static early stopping. In non-abelian S_5, unshielded transfer surges transiently (95.4% peak), but a prospective shielding cohort yields no confirmed benefit (+0.15 +/- 1.14 pp). These results establish a component-level dissociation between transfer acceleration and trajectory stability, and expose the empirical boundaries of parameter shielding.

Diffusion-based robot policies have become widely used in robotic manipulation, where they are typically trained with behavior cloning. However, policies trained purely from demonstrations are limited by the quality and coverage of the available data. Reinforcement learning can further improve the performance of these pretrained policies through interaction. A common approach is to use policy-gradient methods that formulate diffusion-policy fine-tuning as an outer environment MDP together with an inner denoising MDP. However, existing methods typically assign the same environment-level credit to all denoising steps used to construct an action chunk, without distinguishing which intermediate decisions contributed most to the final return. We introduce Denoising Intermediate Advantage (DIA), a policy-gradient method that learns a value function over partially denoised actions and uses it to construct a denoising level advantage for each step of the generative process. DIA combines this inner credit signal with the standard environment-level PPO advantage, providing state-dependent credit throughout the denoising chain. Across Robomimic, FurnitureBench, Franka Kitchen, and D3IL, DIA consistently improves final performance over existing diffusion-policy fine-tuning methods. Beyond final reward, DIA reaches successful states more efficiently and can shift farther from the pretrained behavior distribution, enabling it to discover more effective and efficient task-level strategies and subtask sequences that baseline methods fail to reach.

Agent reinforcement learning (RL) increasingly runs through full execution harnesses, and a multi-harness recipe mixes two choices: exposing the policy to several harnesses, and comparing their rewards inside one relative-advantage group. We isolate the second choice in repository-level coding. From one Qwen3-8B supervised warm start we replay the same frozen task-harness records from Aider, OpenHands, Qwen Code, and SWE-agent, with the same number of updates, under two rules for group-relative policy optimization (GRPO), Within (one group per task-harness pair) and Cross (harnesses pooled within a task), and score every checkpoint with a sealed SWE-bench Verified oracle on four source harnesses and a minimal harness held out of training. The evaluation harness is the dominant variable: across 24,000 sealed evaluations it moves the mean solve rate from 2.14% to 9.27%, a factor of 4.3, where the training recipe moves it by 1.16. The grouping rule is not. On the held-out harness, Cross minus Within is +0.25 pp, 95% confidence interval [-0.48, +1.02], at eight attempts per task, and +0.16 [-0.41, +0.72] pooled over three training seeds whose individual estimates change sign. Each rule's own seed range, 0.42 to 0.45 pp, exceeds the difference between them. Both rules place their largest gains on the same source harness. The pooled advantage carries the harness: an out-of-fold classifier recovers the generating harness from Cross's advantage +4.48 pp above the shuffled-label baseline and from Within's not at all, and the two rules still reach the same held-out score and action distribution inside each harness. Re-collecting half the training data on-policy does not change this. Cross-harness credit yields configuration adaptation and no more portable capability than within-harness credit. Multi-harness RL reports should state the grouping boundary and test under an unseen harness.

Group Relative Policy Optimization (GRPO) is widely studied for reinforcement learning with verifiable rewards, where its advantage estimator assigns each rollout a magnitude from within-group reward statistics. In the common case, this magnitude rewards rollouts that reach the correct answer through reasoning. Yet, an overlooked case shares the same surface: a rollout may land on it by guessing, and the formula still assigns a high magnitude, which we identify as the spurious advantage. This arises in three cases: bounded-answer tasks with a small candidate set; open-answer sets hosting bounded sub-cases; and search agents whose budget opens many paths to the same answer. In all three, this misleads the policy toward guess-like behaviors. We propose SIGNBALANCE, whose magnitude is composition-free: it keeps the verifier sign, uses a global scale, and restores zero-mean balance via a stop-gradient per-class rescaling. Across math and search agent benchmarks at different scales, SIGNBALANCE matches GRPO on open-answer math and improves on bounded-answer math and search agents. Code will be released.

Group-based reinforcement learning objectives such as GRPO can allocate learning signal poorly across prompt difficulty: under binary rewards, group normalization induces a divergent weighting on easy prompts. We introduce Softmax Advantage Group Estimation (SoftmaxGRPO), a drop-in alternative that replaces z-score-normalized group advantages with temperature-scaled softmax advantages, keeping weights bounded regardless of prompt difficulty. For binary rewards, we derive the exact finite-group population objective and identify MaxRL as its low-temperature limit. For bounded scalar rewards, we show that the large-group update exactly optimizes a log-moment-generating-function objective, while a universal finite-group scalar objective cannot exist without additional assumptions on the reward distribution. Empirically, SoftmaxGRPO reallocates measured gradient budget away from near-solved prompts and consistently improves over GRPO under identical rewards. It reaches 51.8% on DeepMath with verifiable rewards and improves a 1.5B instruction-tuned model from 35.0% to 68.0% on Poetry using only lightweight text-similarity rewards.

Multi-agent policy optimization, exemplified by PPO-based methods, is a key branch of cooperative Multi-Agent Reinforcement Learning (MARL). A central design question is how many neighboring agents\footnote{In this paper, "neighbors" refer not only to physical proximity but also to agents whose actions influence one another.} to aggregate in order to effectively utilize global information for cooperation. This decision must be made along two dimensions: in the advantage (which agents' rewards contribute to the credit signal) and in the ratio (which agents' likelihood ratios form the clipped importance weight). Existing methods occupy scattered, underexplored points on these two axes: IPPO treats both separately; MAPPO pairs a team-level advantage with per-agent ratios; HAPPO employs sequential ratios with per-agent advantages; and single-agent reductions operating on factorized joint policies aggregate both into fully joint products. We formalize these two design choices as support matrices $\SA$ and $\SR$, and prove a canonical structure: the expected multi-agent policy optimization objective depends on the pair $(\SA,\SR)$ only through their matrix product $\tS=\SR\SA$. This yields two key consequences: (i) Redundancy: the two support matrices are interchangeable with respect to the signal, meaning neither aggregation pattern is inherently superior.(ii) Variance Ordering: the advantage aggregates rewards as a sum (additive variance with an interior bias-variance optimum at the coupling neighborhood), whereas the ratio aggregates likelihood ratios as a product (multiplicative variance that grows exponentially with support size, with no accompanying bias reduction). The resulting design principle is unambiguous: aggregate neighbors in the advantage, sized to the coupling neighborhood, and keep the ratio per-agent.

Reinforcement learning with verifiable rewards (RLVR) commonly uses entropy for advantage shaping. However, entropy cannot distinguish useful uncertainty from detrimental confusion, limiting its effectiveness as a correctness signal. We propose Contrastive Policy Optimization (CPO), which uses token-level contrastive disagreement between reference-guided and vanilla generation distributions for correctness-aware advantage shaping. Both theoretical and empirical results show that this disagreement reliably indicates token-level correctness. We further show that On-policy Distillation is a special case of CPO, where the posterior distribution is instantiated by an external teacher model. CPO also resolves the zero-advantage problem. Experiments on in-domain and out-of-domain benchmarks demonstrate that CPO substantially outperforms entropy-based RLVR methods while maintaining strong generalization. Further analysis shows that correct and incorrect responses naturally support exploration and exploitation respectively, and balancing both leads to the best performance.

Few-shot multimodal classification commonly attaches a lightweight head, such as $k$-nearest neighbors, logistic regression, or a linear SVM, to a frozen pretrained encoder. Although computationally efficient, these heads can produce poorly calibrated confidence scores, limiting their reliability in calibration-sensitive applications. We evaluate TabPFN as a plug-and-play, zero-gradient classification head for frozen image, text, and audio encoders. Across 22{,}820 evaluation episodes spanning 14 datasets, 11 encoders, and three modalities, TabPFN achieves the best mean rank among nine classification heads on both negative log-likelihood (NLL) and expected calibration error (ECE). At a representative setting, it reduces NLL by 48--62% and ECE by 2.1--5.3$\times$ relative to the average of the eight baselines while matching or exceeding their average accuracy. Its accuracy advantage is conditional, concentrating at moderate-to-high shot counts and low-to-moderate feature dimensions ($k \ge 50$, $d \le 32$), and diminishing when labeled data are scarce, feature dimensions are high, or competing methods approach ceiling accuracy. In targeted backbone-adaptation experiments, replacing the trained linear head with TabPFN substantially improves calibration while preserving competitive accuracy. These results provide empirical guidance for using TabPFN as a training-free head in calibration-sensitive multimodal classification. To support transparency and reproducibility, we publicly release the source code, experiment configurations, and evaluation scripts in our GitHub repository: https://github.com/Jingxiang-Zhang/tabpfn-multimodal-embeddings.

Group Relative Policy Optimization (GRPO) stalls on a model's hardest problems: when no rollout in a group succeeds, the group-relative advantages vanish and the problem contributes no gradient, wasting the frontier examples we most want to learn from. Prepending a correct prefix of a reference solution raises the success rate, making prefix length a continuous knob on difficulty. Concurrent methods set the knob once; AdaPrefix-GRPO turns it into a feedback controller: throughout training it adjusts how much of the solution each problem gets, holding its success rate near 50%, where GRPO's gradient signal is largest, then withdraws the assistance entirely, so the deployed model solves problems unaided. On hard math, at matched training FLOPs, it more than doubles GRPO's accuracy on held-out problems from the training distribution for a 0.6B model (2.1x), with 1.6x on Qwen3-1.7B and 1.7x on AIME, while roughly halving trace length. The method is implemented in data preparation plus a loss mask on prefix tokens; the trainer is otherwise stock. The smaller the model, the larger the gain.

We analyze the effect of optimizing the initial population of genetic programming (GP) for symbolic regression (SR) on the accuracy and complexity of solutions. We compare three well-established random initialization methods as well as initialization with small optimized solutions from exhaustive symbolic regression (ESR) using a GP/SR implementation which is based on the multi-objective evolutionary algorithm NSGA-II. We compare the final Pareto fronts found with each initialization method on twelve synthetic problems of varying complexity and one real-world dataset. We find no significant differences in accuracy or model complexity among the initialization methods. The initial advantage of initialization with ESR disappears after only a few generations. Our results show that, given similar diversity in the initial population, the effect of the initialization method in GP-based symbolic regression on the final Pareto front is negligible.

Two-sided marketplaces connect distinct user groups whose interests often conflict -- improving outcomes on one side could degrade the other side's experience. To address this challenge, we deploy an integrated framework for personalizing free-value thresholds -- a policy governing the scope of complimentary services for job listings -- across a two-sided job marketplace connecting millions of employers and job seekers. Our personalized policy delivers statistically significant and economically sizable lift in the target metric while respecting engagement guardrail constraints. Direct application of standard uplift methods proves insufficient here for two reasons. First, cross-side externalities demand multi-objective optimization: maximizing employer-side metrics risks harming job seeker engagement, with effects varying substantially across job segments. Second, marketplace interference necessitates cluster-level randomization, limiting us to few discrete treatment levels -- effectively a form of positivity violation that rules out methods designed for continuous treatments. We contribute an integrated framework with three components. Our ensemble-based hybrid ranking models target and guardrail metrics separately, cutting guardrail risk by over 10% for equivalent target gains compared to single-objective approaches. A treatment effect extrapolation method extends our estimates from limited experimental variation to untested policy levels, relying on monotonicity assumptions that we validate empirically. Finally, we present production deployment, where post-launch data confirms both extrapolation accuracy and guardrail compliance. Our deployed system demonstrates that principled methodology can enable meaningful personalization even when experiments are severely constrained and different objectives compete -- common conditions that characterize many real-world marketplaces.

Reinforcement learning from human feedback (RLHF) in production does not always have a synchronous reward signal. Code-execution verifiers, slow judge ensembles, and queued human review can return several gradient steps after the rollout that produced them, breaking the synchronous-reward assumption underlying standard PPO. We address this gap with Retroactive Advantage Correction (RAC): each pending slow completion is queued, aged through a non-negative kernel, and reinjected as a clipped residual into the next optimiser step's advantage. We prove that under an unbiased clipped importance ratio, the cumulative RAC correction is exactly unbiased when the effective delay kernel reinjects all of its mass, and carries a bias linear in the unreinjected fraction otherwise; at the no-delay identity kernel it reduces to V-trace. On a tabular Markov decision process (MDP) proof-of-concept, RAC reduces the closed-form policy bias by up to 47.9x at the two-slow-channel configuration, beating wait-for-slow at lower wall-clock cost. RAC integrates with PPO and GRPO through a two-line reward-manager patch.

Reinforcement learning from verifiable rewards (RLVR) has driven rapid progress in mathematical and code reasoning, but when extended to science, existing benchmarks do not decompose what generalizes: do gains reflect structural transfer, property transfer, or memorization? We introduce Mat-Pref, a benchmark of 10,837 ionic-substitution questions across 11 inorganic structure families, grounded in density functional theory calculations from the Materials Project, with three evaluation splits that isolate in-distribution performance, generalization to entirely held-out structure families, and cross-property transfer: applying band-gap reasoning to hosts seen during training only through formation-energy supervision. Four zero-shot frontier models (70-671B parameters) remain in the 33-54% range on every split, confirming that scale alone does not resolve the compositional chemical reasoning this task demands. A two-stage pipeline of supervised fine-tuning followed by Group Relative Policy Optimization (GRPO) lifts Qwen3-8B to 65.2% in-distribution and 71.6% on held-out families, exceeding zero-shot Qwen3-235B by over 20 percentage points on both structural-generalization splits. Self-consistency sampling shows that the SFT policy can already produce correct answers but cannot reliably surface them as the modal response; GRPO reshapes the distribution so that correct answers become modal rather than merely reachable, and this sharper commitment is visible mechanistically: logit lens analysis reveals a ${\sim}$20pp advantage in answer crystallization at the critical decision layer. We formalize this observation as a distractor-permutation consistency metric under which GRPO narrows the gap between lenient scoring (at least one permutation correct) and strict scoring (all permutations correct) from 24.0 to 14.3 percentage points.

Programming by Demonstration (PbD) offers a human-centered way to author procedural knowledge for LLM agents: users communicate what they want by showing rather than by writing prompts or code, making agent authoring accessible to non-programmers. The natural output of a PbD recording is a flat action log, but how this log is organized before being passed to the agent is an open design question with significant consequences for plan quality. We propose grouping recorded actions into labeled, hierarchical subgoals and evaluate the effect of this organizational structure in a controlled experiment. Across 85 web automation tasks, we compare a zero-shot baseline against four demonstration formats that share identical action sequences but differ in structure. On 43 natural-language tasks with vague descriptions, hierarchically grouped demonstrations improve pass rates from 76.7% to 90.7% (paired permutation test $p{=}0.034$; win-loss 6:0), while flat demonstrations show a smaller, non-significant improvement. On 42 tasks with precise descriptions, no format provides any benefit, confirming that the hierarchical advantage arises specifically when descriptions leave procedural details ambiguous. Ablation shows that subgoal grouping alone drives the effect: preconditions, postconditions, and parameter annotations add no measurable benefit. These results offer a concrete design recommendation for PbD pipelines and, more broadly, for any system that feeds procedural context to an LLM agent: segment action sequences into named subgoal groups rather than presenting flat step lists.

Large language models (LLMs) are becoming a major way for consumers to find products, but we do not yet understand how brands compete in this new channel. We study brand dynamics in LLM recommendations using skincare products -- a category where consumers cannot easily judge quality before buying and must rely on brand reputation -- across three commercial LLMs (GPT-4o-mini, Claude Sonnet, Gemini 3 Flash), with a robustness check on search goods. In three experiments, we find: (1) a Conditional Monopoly where well-known brands get recommended 100% of the time (IAI = 10.0) when all products have the same specifications, but this dominance disappears with less than a +0.1-star rating advantage for a competitor; (2) authority-style marketing language, including fabricated clinical-evidence claims, breaks this monopoly at a Bias Surplus Value equal to +0.17 rating points, with each model responding differently; and (3) a social dilemma in multi-brand GEO competition: when all brands adopt the same optimization strategy, individual payoff falls from +0.802 to +0.007 in our payoff proxy, and non-participating brands receive zero recommendations in our tests. Our results suggest that generative engine optimization (GEO) should be studied not only as a security risk, but also as an emerging marketing practice that shapes market competition.

Prevailing wisdom posits that Multi-Agent Systems (MAS) are superior to Single-Agent Systems (SAS), citing advantages like context protection, parallel processing and distributed decision-making. However, empirical support for this claim relies primarily on comparisons with SAS baselines using benchmarks that prioritize isolated reasoning tasks, which do not adequately assess these advantages. Focusing on automatically generated MAS that are designed for enhanced generalizability over manually-designed counterparts, we perform a rigorous, systematic evaluation against SAS, specifically Chain-of-Thought with Self-Consistency (CoT-SC). Across traditional reasoning datasets and tasks with interactive multi-step workflows (e.g., BrowseComp-Plus), we demonstrate that automatic MAS consistently underperform CoT-SC despite being up to 10x more expensive. To isolate these failures from limitations inherent to task structure, we introduce a diagnostic synthetic dataset tailored for MAS featuring explicit task decomposition, context separation and parallelization potential. We show that expert-architected MAS consistently outperforms automatically generated architectures in both raw performance and cost-efficiency on this dataset, demonstrating that existing evaluation frameworks mask critical architectural gaps and inefficiencies of complex MAS by failing to account for the marginal utility of increased computational cost. Critically, systematic deconstruction of the generated MAS architectures reveals that current automated design paradigms produce architectural bloat that prioritizes superficial complexity which does not translate into functional utility, exposing a fundamental misalignment with multi-agent principles.

Current reinforcement learning from human feedback (RLHF) methods primarily rely on scalar rewards from a trained reward model (RM). While effective, scalar rewards are often noisy and fail to capture fine-grained preference differences, whereas RM hidden states encode richer semantic and preference information. We introduce the representation-aware advantage estimation, which leverages RM hidden states and models them as auxiliary signals for better advantage estimation. Specifically, we propose the Graph-based Advantage Estimation (GraphAE), treat each sampled group as a graph, where nodes correspond to responses and edges capture their similarity in the RM hidden space. Then advantages are computed via graph propagation, enabling each sample to incorporate contextual information from its neighbors. GraphAE is lightweight and can be seamlessly integrated into existing group-based RL algorithms. We apply GraphAE to GRPO, GSPO and RLOO, and conduct extensive experiments on different models and benchmarks. Empirical results show consistent improvements across three benchmarks, with gains of up to + 6.3 on Arena-Hard-v0.1, + 8.27 on AlpacaEval 2.0, and + 0.22 on MT-Bench. These results demonstrate that leveraging RM representations leads to more sample efficient and robust RLHF.

Large language models are increasingly used to answer culturally grounded questions across languages, yet it remains unclear whether local cultural knowledge is better accessed through English or the local language. Existing evaluations face two key limitations: many rely on parallel template-based questions that may not reflect how cultural knowledge naturally appears, and raw accuracy conflates general language proficiency with language-conditioned knowledge access. We address these issues with a controlled framework built on real-world cultural questions collected from regional benchmarks and local sources. By crossing question type (culture-agnostic vs. culture-specific) with query language (English vs. local language), and estimating ability with a shared 1PL item response theory model, we separate proficiency from localized knowledge access. Across 13 locales and roughly 80 models, we find a consistent English advantage on culture-agnostic questions, indicating stronger English proficiency. However, after accounting for this proficiency gap, local languages show a positive knowledge-access advantage in nearly all locale-model settings. This advantage is often masked in raw accuracy but becomes more visible for frontier, regionally aligned, or language-adapted models. Our results suggest that weaker local-language performance does not necessarily imply weaker cultural knowledge; rather, local cultural knowledge may be more accessible through the local language but hidden by limited language proficiency.

Reinforcement learning with verifiable rewards (e.g. GRPO) is now a common way to improve mathematical reasoning in Large Language Models (LLMs). However, current methods usually broadcast one sequence-level advantage to all tokens, or use costly process reward models (PRMs) for step-level supervision. Uniform advantage distribution assumes that all tokens contribute equally to the final reward. This dilutes the gradient signal, since flawed reasoning steps and filler words are updated as strongly as valid logical inferences. To address this, we introduce Gradient-Reweighted Advantage (GRAIL), an intrinsic token-wise advantage reweighting method. GRAIL uses gradient-activation saliency to place more weight on tokens that are more locally sensitive to the final answer. Evaluations across five models from the Qwen3, R1-distilled and OctoThinker families show that GRAIL consistently outperforms GRPO. GRAIL achieved an average improvement of 3.60% in accuracy and 3.05% in Pass@3, demonstrating that fine-grained reasoning alignment can be achieved without process-level supervision.

Offline goal-conditioned reinforcement learning requires both long-horizon reachability estimates and local action comparisons. Dual goal representations provide value fields that capture global goal reachability, but they do not directly specify which action should be preferred at a given state. We propose Dual Advantage Fields, a policy-extraction method that turns a bilinear dual value model into a local advantage signal. Under bilinear dual parameterization, the goal embedding is the gradient of the value field with respect to the state representation. DAF learns an action-effect model that predicts the discounted feature displacement induced by an action and scores actions by the alignment between this displacement and the goal direction. In the realizable case, this score equals the goal-conditioned Bellman advantage, yielding a standard local policy-improvement guarantee. On OGBench locomotion, manipulation, and puzzle tasks, DAF improves aggregate RLiable metrics and performs strongly in settings where locally correct actions differ from direct movement toward the final goal.

On-policy distillation (OPD) is a widely used technique to transfer capabilities from capable teacher language models to the base student models, and can be formulated in a reinforcement learning style objective using student generated rollouts. Yet, despite the divergence reward being dependent on student model likelihood, existing works usually adopt a stop gradient design primarily for stability, which makes the resulting advantage estimation questionable. In this work, we provide a generic optimization framework based on f-divergence between the student and teacher, and mathematically revisit whether such design space is valid. We prove that general stop-gradient operation would lead to biased estimates of the reward objective and corresponding gradient for general divergence functions. We propose OPD+, the corrected version of OPD that demonstrates improved performance over the baseline KL approach and also supports the choice of various f-divergence. We validate our findings on mathematical reasoning and tool-use benchmarks.

Reinforcement learning with verifiable rewards (RLVR), especially Group Relative Policy Optimization (GRPO), has been widely used to improve reasoning in large language models. However, outcome-level rewards provide only sparse supervision, and group-relative advantages vanish when all sampled trajectories for a prompt are either correct or incorrect. On-Policy Self-Distillation (OPSD) offers dense token-level guidance, but its token preferences are not necessarily aligned with trajectory correctness; empirical diagnostics show that OPSD signals behave differently on correct and incorrect rollouts, with teacher-positive and teacher-negative gap signals exhibiting different noise profiles. These diagnostics are conducted under an OPSD-style privileged teacher context for analysis only, whereas CAST training uses answer-free self-teacher scoring.Motivated by these observations, this work proposes CAST, an answer-free self-distillation method for GRPO-style RLVR. CAST keeps the verifier-grounded GRPO objective, but uses a stop-gradient self-teacher to shape token-level advantages according to trajectory correctness. Unlike prior self-distilled RLVR methods, CAST does not require reference-solution-conditioned teacher scoring, keeps the self-teacher log-probability gap active throughout training, and applies bidirectional local advantage sign reversal: teacher-negative tokens in correct trajectories can receive negative token-level advantages, while teacher-positive tokens in incorrect trajectories can receive bounded positive local advantages. For zero-variance all-correct and all-wrong groups, CAST assigns bounded sign-constrained base advantages, so these otherwise zero-gradient groups can contribute verifier-signed token feedback. Experiments on mathematical reasoning show that CAST improves RLVR training while retaining a lightweight, verifier-grounded trajectory-level objective.

Reinforcement learning with verifiable rewards has become a standard recipe for improving the reasoning abilities of large language models. Existing algorithms face a tradeoff between computational efficiency and sample efficiency in value estimation and policy learning. We introduce BASIS, a critic-free post-training algorithm designed to address this tradeoff. At each online training step, BASIS samples only one rollout per prompt, but leverages rich information across prompts in the entire batch to improve value function estimation. Our experiments demonstrate that BASIS reduces MSE in value function estimation by 69% compared to REINFORCE++, a representative single-rollout baseline, and achieves lower MSE with one rollout than group mean estimators with 8 rollouts. This improvement in value estimation translates to better policy optimization: using substantially less training time, BASIS achieves performance close to multi-rollout GRPO-type baselines and often outperforms single-rollout REINFORCE-type baselines.

Reinforcement Learning has become a standard paradigm for aligning Large Language Models with human intent and task requirements. While Group Relative Policy Optimization offers an efficient, value-model-free alternative to Proximal Policy Optimization, adapting it to real-world multi-reward settings remains challenging. Standard scalarization practices, such as Reward Combination and Advantage Combination, suffer from significant drawbacks: Reward Combination frequently generates advantages with excessively large squared magnitudes that lead to training instability, while Advantage Combination relies on static hyperparameters and ignores cross-objective correlations. To address these limitations, we propose Dynamic Variance-adaptive Advantage Optimization (DVAO), which dynamically adjusts combination weights based on the empirical reward variance of each objective within a rollout group, effectively up-weighting objectives with a stronger learning signal while suppressing noisy ones. We mathematically prove that DVAO maintains bounded advantage magnitudes for stable training and introduces a self-adaptive cross-objective regularization mechanism. Extensive experiments on mathematical reasoning and tool-use benchmarks using Qwen3 and Qwen2.5 models demonstrate that DVAO significantly outperforms baseline methods, achieving a superior multi-objective Pareto frontier and robust training stability.

Correlation matrices are fundamental summaries of functional brain networks, yet standard analyses often treat entries independently, ignoring the curved geometry of correlation space. Existing geometric methods frequently lack closed-form operations or depend on arbitrary region ordering, limiting scalability. We introduce a scalable geometric framework with two components: (i) the Off-log metric, a smooth transformation mapping correlation matrices to symmetric zero-diagonal matrices. This enables closed-form expressions for distances, Frechet means, and linear models, allowing standard statistical modeling without complex manifold optimization. (ii) Grassmannian subspace discrimination, which compares subjects via principal-angle distances between eigenvector subspaces, resolving inherent sign and basis ambiguities. Both components integrate into standard machine-learning workflows for inference, regression, and classification. Validated across two clinical cohorts (Parkinson's and psychosis) and three ageing fMRI datasets, the Off-log metric increased sensitivity in permutation tests and matched or exceeded Riemannian and Euclidean baselines in classification. Brain-age prediction performance was comparable, with Riemannian metrics excelling in two of three cohorts. The Grassmannian method consistently outperformed Euclidean baselines, highlighting disease-relevant networks. Overall, geometry-aware representations improve sensitivity and predictive performance while remaining straightforward to deploy at scale.

We preregistered a comparison of two ways to help an LLM answer questions over a small research corpus: a single-round Vector RAG system and an LLM-compiled markdown wiki. Both systems answered the same 13 questions over 24 papers using the same answer-generating model, and their answers were scored by blinded LLM judges. The wiki scored much better at connecting findings across papers, but its advantage in answer organization was not strong after judge adjustment. RAG met the preregistered test for single-fact lookup questions. The clean query-side cost result went against the expected wiki advantage: under the tested setup, the wiki used far more query tokens than RAG, so it could not recover any upfront build cost through cheaper queries. Two exploratory analyses changed how we interpret the result. First, claim-level citation checking favored the wiki: its cited pages more often supported the exact claims being made, even though RAG scored better on the overall groundedness rubric. Second, a decomposition-based RAG variant recovered most of the wiki's advantage on cross-paper synthesis at lower LLM-token cost, but it did not recover the wiki advantage in claim-by-claim citation support. The main conclusion is that grounded research synthesis is not a single capability. Systems can differ in how well they organize evidence, how well their citations support each claim, and how much they cost to run. In this study, no architecture was best on all three.

Reinforcement learning has become a powerful paradigm for post-training large language model agents, yet credit assignment in multi-turn environments remains a challenge. Agents often receive sparse, trajectory-level rewards only at the end of an episode, making it difficult to determine which intermediate actions contributed to success or failure. As a result, propagating delayed outcomes back to individual decision steps without relying on costly auxiliary value models remains an open problem. We propose Generalized Advantage Grouped Policy Optimization (GAGPO), a critic-free reinforcement learning method for precise, step-aligned temporal credit assignment. GAGPO constructs a non-parametric grouped value proxy from sampled rollouts and uses it to compute TD/GAE-style temporal advantages, recursively propagating outcome supervision backward through time. Combined with group-wise advantage normalization and an action-level importance ratio, GAGPO extracts stable, localized optimization signals directly from multi-turn trajectories. Experiments on ALFWorld and WebShop show that GAGPO outperforms strong reinforcement learning baselines. Further analyses demonstrate faster early-stage learning, improved interaction efficiency, and smoother optimization dynamics, suggesting that GAGPO offers a simple yet effective framework for multi-turn agentic reinforcement learning.

Large language models (LLMs) achieve strong reasoning performance by allocating substantial computation at inference time, often generating long and verbose reasoning traces. While recent work on efficient reasoning reduces this overhead through length-based rewards or pruning, many approaches are post-trained under a much shorter context window than base-model training, a factor whose effect has not been systematically isolated. We first show that short-context post-training alone, using standard GRPO without any length-aware objective, already induces substantial reasoning compression-but at the cost of increasingly unstable training dynamics and accuracy degradation. To address this, we propose Step-level Advantage Selection (SAS), which operates at the reasoning-step level and assigns a zero advantage to low-confidence steps in correct rollouts and to high-confidence steps in verifier-failed rollouts, where failures often arise from truncation or verifier issues rather than incorrect reasoning. Across diverse mathematical and general reasoning benchmarks, SAS improves average Pass@1 accuracy by 0.86 points over the strongest length-aware baseline while reducing average reasoning length by 16.3%, yielding a better accuracy-efficiency trade-off.

Search agents achieve strong question-answering performance through multi-turn interactions with search engines, with Group Relative Policy Optimization (GRPO) being a widely used training algorithm. However, GRPO-style algorithms still face several challenges in multi-hop search settings. First, correct intermediate steps are often penalized when the final answer is wrong. Second, training is highly unstable, often causing degradation of natural language ability or even catastrophic training collapse. Our analysis attributes these issues to coarse-grained advantage assignment and an imbalance between positive and negative advantages. To address these problems, we propose CalibAdv, an advantage calibration method specifically designed for search agents that enables more accurate and more stable modeling of penalties and rewards. Specifically, CalibAdv leverages the correctness of intermediate steps to downscale excessive negative advantages at a fine-grained level. It then further rebalances positive and negative advantages to improve training stability. Importantly, CalibAdv adopts a lightweight design that calibrates advantages from standard rollout signals, making it simple and easy to deploy. Extensive experiments across three models and seven benchmarks demonstrate that CalibAdv improves both model performance and training stability. Our code is available at https://github.com/wujwyi/CalibAdv.

Conditional depth execution routes a subset of tokens through a lightweight cheap FFN while the remainder execute the standard full FFN at each controlled layer. The central difficulty is gate training: the gate decision must propagate through many layers before it influences the language modeling (LM) loss, so the resulting gradients are weak and noisy. Auxiliary losses are commonly stacked to stabilise training, yet the interactions among them -- particularly between a predictive auxiliary and explicit score supervision -- have not been systematically compared under controlled conditions. We evaluate two gate designs under a 157.5M-parameter decoder-only model with controller-only training, 50% full-path budget, and 3-seed runs on a fineweb-edu subset. The MLP gate (G1) maps the current hidden state to a utility score; the JEPA-guided gate (G3) adds an action-conditional predictor that forecasts, in a low-dimensional latent space, the outcome of executing full vs. cheap per token, aligned against a fixed target head. Under the standard recipe with oracle-style utility regression and pairwise rank supervision (util/rank), G3 improves early-to-mid optimisation over G1 in 3/3 seeds (lower avg LM, faster threshold hits, ~10.3x lower grad norms), with 20k-step endpoint LM within a 0.005 heuristic reference. A key finding (ablation A3): jointly removing util/rank improves best/avg LM and threshold-hit speed in 3/3 seeds for both gates, and the early-to-mid advantage of G3 over G1 disappears. We trace this to an off-policy oracle label that assumes all subsequent layers execute full, whereas gated execution routes only a fraction through full -- making util/rank net-negative under the current recipe. Removing util/rank also cuts the training FLOPs proxy from ~1.53x to ~1.07x full-only (2.87h to 1.75h on a V100-32GB, ~39%). Conclusions are scoped to the studied regime.

Reinforcement Learning (RL) has emerged as a central paradigm for advancing Large Language Models (LLMs), where both pre-training and RL post-training stages are grounded in the same log-likelihood formulation. In contrast, recent RL approaches for diffusion models, most notably Denoising Diffusion Policy Optimization (DDPO), optimize an objective different from the pretraining objectives--score/flow matching loss. In this work, we establish a novel theoretical analysis: DDPO is an implicit form of score/flow matching with noisy targets, which increases variance and slows convergence. Building on this analysis, we introduce Advantage Weighted Matching (AWM), a policy-gradient method for diffusion. It uses the score/flow-matching loss and reweights each sample by its advantage. In effect, AWM raises the influence of high-reward samples and suppresses low-reward ones while keeping the modeling objective identical to pretraining. This simple yet effective design yields substantial benefits: on the GenEval, OCR, and PickScore benchmarks, AWM delivers up to a $34\times$ speedup over Flow-GRPO (which builds on DDPO), when applied to Stable Diffusion 3.5 Medium and FLUX, without compromising generation quality. Code is available at https://github.com/scxue/advantage_weighted_matching

The recently introduced YOLO26 architecture incorporates NMS-free end-to-end inference and is optimized for deployment on resource-constrained CPU-based devices, making it well-suited for edge-based aquaculture applications. However, its performance, operational efficiency, and deployment suitability have not been systematically validated in aquaculture-specific scenarios. This study presents a comprehensive benchmark of YOLO26 against three Ultralytics predecessors (YOLOv5u, YOLOv8, and YOLO11) across nano, small, and medium model scales for fish mortality detection, a critical indicator of fish population health and welfare. Twelve model variants were evaluated for detection accuracy, training efficiency across seven dataset sizes, and inference performance on high-performance NVIDIA A100 GPUs and a CPU-only Raspberry Pi 5 edge platform. All models achieved comparable performance on the full dataset, with mAP50 differing by only 1.04 percentage points, indicating that architectural generation has little influence on final detection accuracy when sufficient training data are available. However, clear trade-offs emerged in data efficiency and deployment performance. YOLOv8 achieved 90% mAP50 with only 400 training images, whereas the YOLO26 nano and small variants required 1,000 images to reach comparable accuracy. Conversely, YOLO26n achieved the highest inference speed on the Raspberry Pi 5 (7.51 FPS), while YOLOv5mu outperformed all contemporary medium-scale architectures on CPU-based hardware. These results show that architectural novelty alone is insufficient for model selection and that training data availability, target hardware, and inference requirements should be considered jointly when selecting object detection models for practical edge AI deployment in aquaculture.