Searcharxiv⌕ Search

arXiv subjects

Hanlin Wu

Publications and source records attributed to Hanlin Wu.

At least 19 recordsLinked to original sources
arXiv2026

Trajectory-Aware Policy Optimization: Closing the State Distribution Gap in Long-Horizon LLM RL

Reinforcement learning (RL) has become a dominant paradigm for enhancing LLMs' reasoning capabilities. Yet long-horizon autoregressive rollouts introduce a fundamental limitation: local policy mismatch can compound into severe state distribution shift over trajectories. Existing LLM RL methods, such as GRPO and GSPO, control policy mismatch at the token or sequence level, but overlook this trajectory-induced shift, causing training instability and eventual collapse as the horizon grows. To close the state distribution gap, we propose Trajectory-Aware Policy Optimization (TAPO), a principled framework that theoretically derives the state-distribution correction from the off-policy policy gradient and incorporates it into the RL objective. TAPO further employs a geometric-mean estimator for variance reduction, together with a length-aware dynamic clipping that prevents overly deviated updates across the trajectory. Extensive experiments demonstrate that TAPO significantly surpasses existing LLM RL algorithms across eight competition-level benchmarks, achieving up to 48% and 12% improvement over GRPO on mathematical reasoning and algorithmic coding, with substantially enhanced training stability.

cs.LG↗
arXiv2026

EPIC: Rethinking Trust Regions for Autoregressive LLM Policies via Exact-KL Prefix Importance-Ratio Clipping

Reinforcement learning (RL) based on PPO-style clipping has become an effective approach to improving the reasoning capabilities of large language models (LLMs). However, standard token-level clipping measures policy change only through the conditional probability of the current token, while the prefix conditioning context is itself generated by the policy. This mismatch becomes especially consequential in long-response and off-policy training. We reformulate autoregressive generation as a finite-horizon Markov decision process and derive its performance-difference identity under behavior-policy sampling. The resulting change of measure identifies the cumulative likelihood ratio of the generated prefix as the exact importance weight for each stepwise policy-improvement term. Starting from the monotonic-improvement principle of trust-region policy optimization, we then construct a prefix-level trust region using the exact reverse-KL coordinate $r-1-\log r$, whose behavior-policy expectation equals the corresponding prefix-distribution KL divergence. This yields EPIC, a PPO-style sequence-clipping method that directly regulates prefix-level policy drift without relying on a local quadratic approximation or expert-routing replay. Experiments on off-policy reinforcement learning with a sparse mixture-of-experts language model show that EPIC consistently outperforms standard and sequence-level clipping baselines across mathematical-reasoning benchmarks, while remaining competitive with routing-replay-based alternatives. Code is available at https://github.com/LiChengxi666/EPIC.

cs.LG↗
arXiv2026

Generalized Proximal Policy Optimization: Unifying LLM RL Algorithms from the Perspective of Divergence

Reinforcement learning has become the foundational paradigm for enhancing the reasoning capabilities of LLMs. Yet the dominant approach, Proximal Policy Optimization (PPO), relies on a geometrically degenerate clipping proxy that cannot adapt its resistance to policy shifts, leading to premature clipping and degraded exploration. Subsequent variants alleviate these issues through fragmented and heuristic modifications. We propose Generalized Proximal Policy Optimization (GPPO), a principled and unified framework for understanding and designing LLM RL algorithms through the lens of divergence. We establish a generalized $f$-divergence trust-region theory with a monotonic performance improvement guarantee, and show that different $f$-divergences induce distinct geometries and policy update dynamics. We further introduce novel estimator topologies that provide more faithful divergence estimates than the conventional single-sample Monte Carlo estimation. Under this framework, PPO and recent variants such as DAPO, DPPO, and RIPO emerge as special cases, establishing a principled two-dimensional design space of {Divergence $\times$ Estimator}. Extensive experiments on eight competition-level math and coding benchmarks show that both dimensions are critical. The best-performing GPPO instantiation surpasses standard GRPO by up to 50% and 14% on math and coding, delivering SOTA performance.

cs.LG↗
arXiv2026

Beyond Euclidean Clipping: Overcoming Exploration Collapse in LLM RL via Riemannian Isometric Policy Optimization

Reinforcement learning (RL) has become a dominant paradigm for enhancing LLMs' reasoning capabilities. However, RL algorithms with PPO-Clip are inherently limited by exploration collapse. Subsequent works remain primarily heuristic and fail to identify the essential cause of PPO-Clip's failure. This work reveals the fundamental flaw of PPO-Clip: it implicitly measures policy discrepancy using Euclidean metric, which is theoretically inconsistent with the intrinsic geometry on the policy Riemannian manifold. This geometric mismatch results in overly conservative updates in low-probability regions while aggressive in high-probability regions, ultimately collapsing exploration. To correct this geometric flaw, we propose Riemannian Isometric Policy Optimization (RIPO), which guarantees isometric policy updates on the Riemannian manifold, effectively balancing exploration and exploitation. We further show that RIPO achieves a favorable bias-variance trade-off, which stabilizes optimization. Extensive experiments demonstrate that RIPO significantly surpasses existing LLM RL algorithms across seven competition-level benchmarks (up to 60% improvement over GRPO on AIME24).

cs.LG↗
arXiv2026

From Pixel to Poses: Object-centric Tool Manipulation Learning from Human Demonstrations

Scaling up robotic manipulation is primarily bottlenecked by the scarcity of real-world robot data. While recent approaches leverage human video demonstrations to mitigate this shortage, they remain computationally expensive and still rely on paired human-robot data for domain alignment. Although current state-of-the-arts excel at long-horizon tasks, they struggle with the delicate and precise control required for complex tool manipulation. To overcome these limitations, we introduce P2P-T, from Pixel to Poses for Tool Manipulation, a data-efficient, object-centric framework that learns tool use directly from human demonstrations. P2P-T bridges the cognitive and physical execution gap through a two-stage approach. First, pretraining an object-centric world model to extract stable pose priors; second, integrating these priors into an efficient, pose-aware low-level policy. By utilizing a robust automated data processing pipeline powered by modern foundation models, P2P-T completely bypasses the need for human-robot aligned data. This reduces overall training overhead drastically. With minimal per-task fine-tuning, our framework achieves a 73% improvement over the previous state of the art in execution performance on complex, real-world tool manipulation tasks that currently remain out of reach for standard large-scale pretrained models.

cs.RO↗
arXiv2026

A model of rational interlocutors: Unification of comprehension and production

Who we communicate with influences both our interpretation of their utterances and the design of our own. Such adjustment to the conversational partner is studied as speaker modeling in comprehension and as audience design in production, with the two literatures having developed largely separately. We argue that both adjustments express one rational computation and propose the rational interlocutor (RI) model, a computational account unifying comprehension and production. An interlocutor maintains a model of their partner, defined by three parameters: an identity parameter Π sets the messages and forms expected from the partner; a fidelity parameter Φ sets how reliably messages and utterances map onto each other for them; a knowledge parameter Λ sets how knowledgeable the partner is believed to be. Comprehension and production are thus mirror-image modes of one computation over the partner model. Comprehension chooses the message the partner most likely intends to convey, weighing how well each candidate fits the utterance against how likely this partner is to mean it. Production chooses the utterance from which the partner will best recover the message, weighed against the effort of saying it. This explains why comprehenders appear to rely less on the forms produced by a linguistically less competent speaker, while producers tend to invest more effort in designing forms for them. We conjecture that perceived linguistic competence decomposes into two of these quantities: fidelity and knowledge. Their contrasting profiles across second-language (L2) adults, children, and artificial partners produce distinct and testable predictions.

cs.CL↗
arXiv2026

Direct Conditional Transition Sampling for Diffusion Inverse Problems

Training-free diffusion inverse solvers typically choose between local measurement guidance and costly clean-space posterior updates. Independent posterior refresh can improve global correction by sampling a clean conditional and re-noising it, but its practical realization requires probability-flow ODE integration and clean-space Markov chain Monte Carlo (MCMC). We propose Direct Conditional Transition Sampling (DCTS), a direct stochastic-flow approximation to the same ideal refresh target. Rather than explicitly drawing a clean sample, DCTS estimates the measurement-conditioned clean mean along a short inner path and transports Gaussian source noise directly to the next noisy state. A denoiser-compatible sufficient statistic and a covariance-scaled operator update enable this conditional-mean estimation. Experiments on four inverse problems demonstrate that DCTS achieves competitive reconstruction quality with up to $16.8\times$ speedups over competing methods.

cs.CV↗
arXiv2026

Open-MOPD: Diagnosing and Fixing Capability Imbalance in Multi-Teacher On-Policy Distillation

Multi-teacher on-policy distillation (M-OPD) has emerged as a promising paradigm for consolidating domain-specialized reinforcement learning (RL) experts into a single generalist student via dense, token-level reward supervision. Despite its practical success, the optimization dynamics governing multi-teacher capability integration remain poorly understood, and open, rigorously reproducible recipes are conspicuously lacking. In this work, we establish a controlled M-OPD benchmark on SmolLM3-3B-Base with oracle routing, isolating capability integration from routing ambiguity. Our investigation reveals a pronounced capability integration gap: standard M-OPD captures only 35.6% of the available headroom relative to a domain-routed oracle ensemble, with concise tasks such as instruction following suffering severe degradation and premature stagnation. Crucially, we show that this failure stems not from gradient conflict, but from a severe misallocation of the token-level optimization budget. This pathology is driven by three orthogonal factors: structural sequence-length disparities across domains, dynamic convergence drift due to non-uniform learning rates, and multi-step reward staleness from asynchronous policy updates. To resolve these imbalances, we introduce Open-MOPD, a principled framework incorporating token-share balancing, gap-aware dynamic budget allocation, and student reward refresh. Together, these mechanisms systematically restore cross-domain balance, elevating headroom recovery from 35.6% to 83.4% in a single deployable student. We fully open-source our end-to-end post-training recipe, training trajectories, and evaluation suites on an academically accessible hardware budget.

cs.LG↗
arXiv2026

Toward Blockage-Resilient 6G-V2X Connectivity: Semi-Distributed Bandit with Dynamic Arm Set for mmWave HetNets

The vision for 6G vehicle-to-everything (V2X) communications demands reliable, adaptive connectivity for fully autonomous driving across complex dynamic environments. Millimeter-wave (mmWave) user association (UA) in heterogeneous vehicular networks presents a particularly demanding instance of this problem, where dynamic blockages and rapid channel variations continuously undermine the stationary reward assumptions of traditional multi-armed bandit (MAB) frameworks. This paper proposes a fully distributed blockage-aware non-stationary dynamic bandit algorithm (BAND) and its semi-distributed extension S-BAND for cooperative learning across vehicles. Blockage prediction is incorporated into the change-detection (CD) mechanism to suppress false alarms, while a dynamic base station (BS) set management scheme balances exploration and exploitation across large-scale BS deployments without requiring centralized channel state information (CSI) acquisition or offline training. In S-BAND, vehicles accumulate BS reward estimates as local knowledge and periodically upload them to the macro base station (MBS), which aggregates them into cluster-based central knowledge. A trajectory-aligned knowledge (TAK) region is proposed to capture the spatial correlation of mmWave channel characteristics. A knowledge inheritance fidelity (KIF) metric is introduced to quantify knowledge transfer quality. Simulation results on a realistic urban topology show that BAND and S-BAND achieve 34.9% and 59.4% regret reduction relative to a centralized MAB baseline, with performance gains sustained across blockage rates ranging from 10% to 50%. The proposed TAK region consistently outperforms the traditional K-means clustering scheme under both fidelity criteria.

cs.ET↗
arXiv2026

A funny companion: Distinct neural responses to AI- versus human-attributed humor

As artificial intelligence (AI) companions become capable of human-like communication, including telling jokes, understanding how people cognitively and affectively respond to AI-attributed humor becomes increasingly important. This study used EEG to compare how people process puns versus controls attributed to either an AI agent or a human. In an interaction task, participants guessed punchlines based on joke setups before they were disclosed by the interlocutor. Behavioral analysis revealed that participants rated AI- and human-attributed humor as comparably funny. However, neurophysiological data showed that AI-attributed humor elicited a reduced and sustained N400 effect compared to human-attributed humor, suggesting reduced cognitive effort in semantic conflict detection and resolution, or attenuated feedback-related processing. This was also accompanied by a larger late positive potential (LPP), reflecting intensified late-stage affective and evaluative processing. This likely reflects a process where the brain reconciles AI's delivery of culturally embedded puns with prior low expectations, leading to intensified affective engagement and interlocutor model updating. Individual differences in social perceptions further influenced neural responses. Higher perceived AI sincerity and trustworthiness were associated with a globally reduced N400, facilitating semantic integration for both humorous and non-humorous AI language. Additionally, increased AI trustworthiness predicted an enhanced LPP, indicating a more intensified updating of the interlocutor model during humor comprehension. These findings indicate that the brain's neural sensitivity to AI-attributed humor bypasses biases like algorithm aversion. This highlights the brain's adaptation to humor from a novel source and underscores humor's potential for fostering genuine engagement in human-AI social interaction.

cs.CL↗
arXiv2026

Neural Dynamics of AI-attributed Irony Reveal a Partial Intentional Stance

As Large Language Models (LLMs) are increasingly deployed as social agents and trained to produce humor and irony, a question emerges: when encountering witty AI remarks, do people interpret them as deliberate communicative acts? This study investigated whether people adopt an intentional stance, ascribing mental states to explain behavior, when comprehending AI-attributed irony. Irony provides a testbed because understanding it requires distinguishing intentional contradictions from unintended errors through pragmatic reanalysis. Using electroencephalography (EEG) to measure event-related potentials (ERPs), we compared behavioral and neural responses to identical ironic utterances attributed to either an AI companion or a human. We found that people do not fully adopt an intentional stance towards AI communication. Participants interpreted contextually incongruent utterances as irony significantly less often when attributed to an AI than when attributed to a human. Correspondingly, we observed attenuated neural responses to AI-attributed irony compared to human-attributed irony during both initial semantic processing (P200) and pragmatic reanalysis (P600). In addition, these neural responses were modulated by individual perceptions of AI sincerity and trustworthiness, with positive perceptions facilitating pragmatic comprehension by reducing cognitive effort. This suggests that adopting an intentional stance toward AI is a flexible and adaptive process, dynamically shaped by people's mental models of artificial agents. These findings reveal that despite advances in communicative competence, AI systems face a barrier to social agency: humans process input from artificial interlocutors with reduced ascription of intentionality. This has important implications for understanding human social cognition and for designing AI systems intended for social interaction.

cs.CL↗
arXiv2026

Weak-to-Strong Generalization via Direct On-Policy Distillation

Reinforcement learning with verifiable rewards (RLVR) is a powerful recipe for improving language-model reasoning, but it is expensive to repeat on every new strong model because the target model must generate many rollouts during training. As models scale, post-training itself becomes a bottleneck. We study a weak-to-strong alternative: run RL on a smaller model where rollouts are cheaper, then reuse what that RL run learned to improve a stronger target model. Directly distilling the post-RL weak teacher is not enough, because the teacher's final policy mixes useful RL gains with the limitations of the smaller model. We propose Direct On-Policy Distillation (Direct-OPD), which transfers the teacher's RL-induced policy shift instead. Direct-OPD compares the post-RL teacher with its own pre-RL reference and treats their log-ratio as a dense implicit reward for the student. In plain terms, the checkpoint pair tells us which actions RL made the weak model more or less likely to take, and Direct-OPD applies that signal on the stronger student's own on-policy states. This directly reuses the weak model's RL supervision signal without running sparse-reward RL on the target model. Empirically, Direct-OPD consistently leverages weaker teachers to improve stronger target models; notably, it boosts Qwen3-1.7B from 48.3% to 58.3% on AIME 2024 in just 4 hours on 8 A100 GPUs. It outperforms step-matched direct RL and enables the sequential composition of multiple policy shifts. Our results show that RL outcomes can be reused across model scales as implicit reward signals, not merely as final models to imitate.

cs.LG↗
arXiv2026

Spectral Rewiring for Exploration, Purification, and Model Merging

Reinforcement learning has become a standard post-training recipe for large language models, but dense full-parameter updates create two deployment-relevant bottlenecks: suppressed reasoning performance, often reflected by premature saturation of test-time scaling, and interference when consolidating multiple capabilities through multi-domain training or model merging. We show that the reasoning-effective component of these updates is largely concentrated in the base model's spectral space, motivating Subspace-Aligned Rewiring (SAR), a post-hoc editing method that retains this spectral core while removing orthogonal components. SAR therefore preserves reasoning gains and filters residual update directions that suppress performance or amplify cross-domain interference. Across several model families and scales, SAR extracts compact reasoning cores using as little as approximately 0.58% of total parameters: it preserves over 99% of post-training performance and improves high-k exploration in mathematical reasoning, and generalizes to agentic coding by improving six of seven open benchmarks on an in-house model. SAR also purifies mixed-domain training updates by releasing suppressed coding capability while maintaining math reasoning and instruction following. It further enables model merging across experts, yielding cross-domain generalization that surpasses previous merging baselines and even the best single-domain experts. Overall, SAR shows that extracting reasoning-effective updates from parameter geometry can serve as a training-free mechanism to improve reasoning and multi-domain performance.

cs.LG↗
arXiv2026

Structure leads and dominates comprehension in naturalistic reading

The hierarchical account and statistical or sequential account have long been framed as rival theories in explaining online comprehension. A lot of evidence has shown that both hierarchical and non-hierarchical factors can shape comprehension and the open question is no longer whether hierarchy contributes, but when and how strongly it does. We addressed the question with co-registered EEG and eye-tracking, treating syntactic depth as the variable for operationalizing hierarchical structure. On timing, hierarchical structure influenced reading before the eyes fixated a word: its neural effect emerged as early as 108 ms before fixation onset, over right-central regions, and the scanpath showed an anticipatory bias toward structurally central words. Both the transitional-probability analysis and the regression on fixation-related potentials supported this pre-fixational timing. In the transitional-probability analysis, readers preferentially moved between syntactically central words rather than following serial word order, showing that scanpaths are organized by syntactic depth rather than by linear adjacency. On strength, Bayesian network modeling showed that syntactic depth was the strongest predictor of departures from linear, word-by-word reading, outweighing lexical familiarity and surprisal. Taken together, the results indicate that hierarchical structure anticipatorily guides online comprehension at both the behavioral and neural levels, and dominates the reading path relative to statistical features.

cs.CL↗
arXiv2026

CaFTRA: Frequency-Domain Correlation-Aware Feedback-Free MIMO Transmission and Resource Allocation for 6G and Beyond

The fundamental designs of wireless systems toward AI-Native 6G and beyond are driven by the need for ever-increasing demand of mobile data traffic, extreme spectral efficiency, and adaptability across diverse service scenarios. To overcome the limitations posed by feedback-based multiple-input and multiple-output (MIMO) transmission, we propose a novel frequency-domain Correlation-aware Feedback-free MIMO Transmission and Resource Allocation (CaFTRA) framework tailored for fully-decoupled radio access networks (FD-RAN) to meet the emerging requirements of AI-Native 6G and beyond. By leveraging artificial intelligence (AI), CaFTRA effectively eliminates real-time uplink feedback by predicting channel state information (CSI) based solely on user geolocation. We introduce a Learnable Queries-driven Transformer Network for CSI mapping from user geolocation, which utilizes multi-head attention and learnable query embeddings to accurately capture frequency-domain correlations among resource blocks (RBs), thereby significantly improving the precision of CSI prediction. Once base stations (BSs) adopt feedback-free transmission, their downlink transmission coverage can be significantly expanded due to the elimination of frequent uplink feedback. To enable efficient resource scheduling under such extensive-coverage scenarios, we apply a low-complexity many-to-one matching theory-based algorithm for efficient multi-BS association and multi-RB resource allocation, which is proven to converge to a stable matching within limited iterations. Simulation results demonstrate that CaFTRA achieves stable matching convergence and significant gains in spectral efficiency and user fairness compared to 5G, underscoring its potential value for 6G standardization efforts.

eess.SY↗
arXiv2026

Flow-based Gaussian Splatting for Continuous-Scale Remote Sensing Image Super-Resolution

High-resolution remote sensing images (RSIs) are crucial for Earth observation applications, yet acquiring them is often limited by sensor constraints and costs. In recent years, generative super-resolution (SR) methods, particularly diffusion models, have made significant progress. However, they typically require slow iterative inference with 40--1000 steps and exhibit limited flexibility in continuous-scale SR settings. To address these issues, we propose FlowGS, a generative reconstruction framework for arbitrary-scale SR of RSIs. FlowGS models the high-frequency detail representations between high- and low-resolution images and learns a continuous probability flow from noise to detail priors via flow matching (FM) constrained by shortcut consistency, thereby reducing generative complexity and improving inference efficiency. Additionally, we employ 2D Gaussian splatting to construct a continuous feature field, thereby enabling flexible reconstruction at arbitrary query locations. Experimental results show that FlowGS delivers competitive perceptual quality compared with existing methods in both continuous-scale and fixed-scale SR settings, with substantially improved inference efficiency.

cs.CV↗
arXiv2026

MapSR: Prompt-Driven Land Cover Map Super-Resolution via Vision Foundation Models

High-resolution (HR) land-cover mapping is often constrained by the high cost of dense HR annotations. We revisit this problem from the perspective of map super-resolution, which enhances coarse low-resolution (LR) land-cover products into HR maps at the resolution of the input imagery. Existing weakly supervised methods can leverage LR labels, but they typically use them to retrain dense predictors with substantial computational cost. We propose MapSR, a prompt-driven framework that decouples supervision from model training. MapSR uses LR labels once to extract class prompts from frozen vision foundation model features through a lightweight linear probe, after which HR mapping proceeds via training-free metric inference and graph-based prediction refinement. Specifically, class prompts are estimated by aggregating high-confidence HR features identified by the linear probe, and HR predictions are obtained by cosine-similarity matching followed by graph-based propagation for spatial refinement. Experiments on the Chesapeake Bay dataset show that MapSR achieves 59.64% mIoU without any HR labels, remaining competitive with the strongest weakly supervised baseline and surpassing a fully supervised baseline. Notably, MapSR reduces trainable parameters by four orders of magnitude and shortens training time from hours to minutes, enabling scalable HR mapping under limited annotation and compute budgets. The code is available at https://github.com/rikirikirikiriki/MapSR.

cs.CV↗
arXiv2026

Speaker effects in language comprehension: An integrative model of language and speaker processing

The identity of a speaker influences language comprehension through modulating perception and expectation. This review explores speaker effects and proposes an integrative model of language and speaker processing that integrates distinct mechanistic perspectives. We argue that speaker effects arise from the interplay between bottom-up perception-based processes, driven by acoustic-episodic memory, and top-down expectation-based processes, driven by a speaker model. We show that language and speaker processing are functionally integrated through multi-level probabilistic processing: prior beliefs about a speaker modulate language processing at the phonetic, lexical, and semantic levels, while the unfolding speech and message continuously update the speaker model, refining broad demographic priors into precise individualized representations. Within this framework, we distinguish between speaker-idiosyncrasy effects arising from familiarity with an individual and speaker-demographics effects arising from social group expectations. We discuss how speaker effects serve as indices for assessing language development and social cognition, and we encourage future research to extend these findings to the emerging domain of artificial intelligence (AI) speakers, as AI agents represent a new class of social interlocutors that are transforming the way we engage in communication.

cs.CL↗