Searcharxiv⌕ Search

arXiv subjects

Tian Qin

Publications and source records attributed to Tian Qin.

At least 19 recordsLinked to original sources
arXiv2026

Bridging Compute- and Data-Optimal Pretraining

Classical compute-optimal scaling laws assume an unbounded supply of fresh pretraining data, yet pretraining is increasingly entering a regime in which compute grows faster than the availability of high-quality data. We propose Compute-Data (CD) scaling laws, a unified framework that bridges compute-optimal scaling, where data scales freely with compute, and data-optimal scaling, where the corpus is fixed while compute can grow without bound. CD scaling extends classical scaling laws by introducing a token-effectiveness function, $\eta$, which quantifies the value of a derived token-produced, for example, through multi-epoch repetition or paraphrasing-relative to a fresh token, ranging from a perfect substitute to having no value. We fit $\eta$ for two data-expansion strategies, multi-epoch repetition and paraphrasing, across model sizes from 14M to 600M parameters using the Dolma-3 corpus. We find that token effectiveness is far from constant: it depends jointly on model size, the tokens-per-parameter ratio, and the amount of derived data, and it saturates as the corpus is expanded. The functional form of $\eta$ implies diminishing returns when substituting compute for data as either model size or data availability increases. It also partitions training into three operational regimes---compute-bound, data-bound, and model-bound---and shows that classical compute-optimal allocation is suboptimal across most practically relevant settings.

cs.LG↗
arXiv2026

Backpropagation-Free Trunk Training via the Split Forward Gradients

Backpropagation makes training deep networks memory intensive because it must store intermediate activations. Forward-mode methods avoid this cost, but their gradient estimates become increasingly noisy as the number of trained parameters grows. We introduce Split Forward Gradient (Split-FG), which splits a network at an intermediate representation: it computes the output head gradient exactly and estimates only the trunk gradient with a Jacobian--vector product. This reduces estimator variance and requires no backward pass through the trunk, while retaining an Adam-style convergence guarantee. Our experiments reveal an important practical failure mode. On WikiText-103, naive forward-gradient training of the trunk performs worse than leaving a randomly initialized trunk frozen, likely because Adam updates every noisy, under-determined trunk coordinate too aggressively. Simply using a much smaller learning rate for the trunk reverses this result: a $16$M-parameter GPT-2-style model reaches validation perplexity $387$, compared with $668$ for the frozen-trunk control and $2{,}885$ for a matched pure forward-gradient baseline (backpropagation reaches $150$). Split-FG also produces the strongest backprop-free results on our tabular benchmarks and reaches $60.5\%$ on CIFAR-10 and $35.2\%$ on CIFAR-100 with a heavy-head design. It reduces peak memory by up to $35\%$ relative to matched backpropagation, although the performance gap widens as the forward-mode trunk grows.

stat.ML↗
arXiv2026

Ricci-Filtration: Boosting Retrieval-Augmented Generation Reranker to Query-Answer Tasks by Discrete Ricci Flow

Ricci flow is a curvature-guided diffusion process that deforms space by shrinking regions of high positive curvature and expanding those with negative curvature. Similarly, discrete Ricci flow on weighted graphs modifies edge weights by shrinking edges with positive Ricci curvature and stretching those with negative Ricci curvature, effectively increasing the separation between clusters. Inspired by these two cornerstone works, we propose a geometry-based RAG reranker enhancement procedure called Ricci-Filtration. By modeling the input query and initial retrieved chunks as a network, where the input query and chunks serve as nodes and embedding-based pairwise relations define an initial graph, Ricci-Filtration leverages discrete curvature and Ricci flow to evaluate the structural importance of each chunk with respect to the user query. The system first filters the initial chunks based on their geometric curvature relative to the query; then, a reranker processes the remaining chunks to enhance generative performance. We theoretically prove that normalized discrete Ricci flow can detect community structures by identifying distinct asymptotic behaviors in edge weights. This supports the removal of ``noisy'' document chunks characterized by large weights and negative Ricci curvature relative to the query node. Extensive experiments confirm that Ricci-Filtration outperforms several baseline reranking methods in accuracy, precision, recall, and F1 scores. Furthermore, ablation studies demonstrate that the Ricci-Filtration generally outperforms the baseline under various settings, highlighting the framework's robustness across different architectures.

stat.ML↗
arXiv2026

RL Excursions during Pre-Training: Re-examining Policy Optimization for LLM training

The standard LLM training pipeline applies reinforcement learning (RL) only after pre-training and supervised fine-tuning (SFT). We question this status quo by training a LLM from scratch and applying RL, SFT, and SFT followed by RL directly to intermediate pre-training checkpoints. We find that RL is effective very early, and often matches the full SFT$\to$RL pipeline early as well. Through experiments on harder problems, we find that targeted pre-training data composition is a strong lever for RL effectiveness, even more so than model scale. Beyond reasoning accuracy, applying RL directly to base checkpoints expands the model's distribution; the sharpening effect reported in recent work arises only when RL follows SFT. The general capabilities of the model remain essentially unchanged by RL, while they degrade following SFT. Finally, we merge RL and SFT objectives by parallel averaging, which outperforms across all other training methods discussed, across metrics, while preserving general capabilities. Together, these results suggest that LLM training might benefit from an expanded use of RL.

cs.LG↗
arXiv2026

Do We Really Need External Tools to Mitigate Hallucinations? SIRA: Shared-Prefix Internal Reconstruction of Attribution

Large vision-language models (LVLMs) often hallucinate when language priors dominate weak or ambiguous visual evidence. Existing contrastive decoding methods mitigate this problem by comparing predictions from the original image with those from externally perturbed visual inputs, but such references can introduce off-manifold artifacts and require costly extra forward passes. We propose SIRA, a training-free internal contrastive decoding framework that constructs a counterfactual reference inside the same LVLM by exploiting the staged information flow of multimodal transformers. Instead of removing visual information from the input, SIRA first lets image and text tokens interact through a shared prefix, forming an aligned multimodal state that preserves prompt interpretation, decoding history, positional structure, and early visual grounding. It then forks a counterfactual branch in later transformer layers, where attention to image-token positions is masked. This branch retains the shared multimodal context but lacks continued access to fine-grained visual evidence, yielding a language-prior-dominated internal reference for token-level contrast. During decoding, SIRA suppresses tokens that remain strong without late visual access and favors predictions whose advantage depends on the full visual pathway. Experiments on POPE, CHAIR, and AMBER with Qwen2.5-VL and LLaVA-v1.5 show that SIRA consistently reduces hallucinations while preserving descriptive coverage and incurring lower overhead than two-pass contrastive decoding. SIRA requires no training, external verifier, or perturbed input, and applies to open-weight LVLMs with white-box inference access.

cs.CV↗
arXiv2026

Incentivizing Temporal-Awareness in Egocentric Video Understanding Models

Multimodal large language models (MLLMs) have recently shown strong performance in visual understanding, yet they often lack temporal awareness, particularly in egocentric settings where reasoning depends on the correct ordering and evolution of events. This deficiency stems in part from training objectives that fail to explicitly reward temporal reasoning and instead rely on frame-level spatial shortcuts. To address this limitation, we propose Temporal Global Policy Optimization (TGPO), a reinforcement learning with verifiable rewards (RLVR) algorithm designed to incentivize temporal awareness in MLLMs. TGPO contrasts model outputs generated from temporally ordered versus shuffled video frames to derive calibrated, globally normalized reward signals that explicitly favor temporally coherent reasoning. Integrated with GRPO and GSPO, TGPO supports cold-start RL training and effectively suppresses spatial shortcut behaviors learned by existing MLLMs. Experiments across five egocentric video benchmarks demonstrate that TGPO consistently improves temporal grounding and causal coherence, outperforming prior RL-based video reasoning approaches. Our results suggest that TGPO offers a simple and scalable pathway toward temporally robust MLLMs for egocentric video understanding.

cs.CV↗
arXiv2025

Drawback of Enforcing Equivariance and its Compensation via the Lens of Expressive Power

Equivariant neural networks encode the intrinsic symmetry of data as an inductive bias, which has achieved impressive performance in wide domains. However, the understanding to their expressive power remains premature. Focusing on 2-layer ReLU networks, this paper investigates the impact of enforcing equivariance constraints on the expressive power. By examining the boundary hyperplanes and the channel vectors, we constructively demonstrate that enforcing equivariance constraints could undermine the expressive power. Naturally, this drawback can be compensated for by enlarging the model size -- we further prove upper bounds on the required enlargement for compensation. Surprisingly, we show that the enlarged neural architectures have reduced hypothesis space dimensionality, implying even better generalizability.

cs.LG↗
arXiv2025

From Static to Adaptive Defense: Federated Multi-Agent Deep Reinforcement Learning-Driven Moving Target Defense Against DoS Attacks in UAV Swarm Networks

The proliferation of UAVs has enabled a wide range of mission-critical applications and is becoming a cornerstone of low-altitude networks, supporting smart cities, emergency response, and more. However, the open wireless environment, dynamic topology, and resource constraints of UAVs expose low-altitude networks to severe DoS threats. Traditional defense approaches, which rely on fixed configurations or centralized decision-making, cannot effectively respond to the rapidly changing conditions in UAV swarm environments. To address these challenges, we propose a novel federated multi-agent deep reinforcement learning (FMADRL)-driven moving target defense (MTD) framework for proactive DoS mitigation in low-altitude networks. Specifically, we design lightweight and coordinated MTD mechanisms, including leader switching, route mutation, and frequency hopping, to disrupt attacker efforts and enhance network resilience. The defense problem is formulated as a multi-agent partially observable Markov decision process, capturing the uncertain nature of UAV swarms under attack. Each UAV is equipped with a policy agent that autonomously selects MTD actions based on partial observations and local experiences. By employing a policy gradient-based algorithm, UAVs collaboratively optimize their policies via reward-weighted aggregation. Extensive simulations demonstrate that our approach significantly outperforms state-of-the-art baselines, achieving up to a 34.6% improvement in attack mitigation rate, a reduction in average recovery time of up to 94.6%, and decreases in energy consumption and defense cost by as much as 29.3% and 98.3%, respectively, under various DoS attack strategies. These results highlight the potential of intelligent, distributed defense mechanisms to protect low-altitude networks, paving the way for reliable and scalable low-altitude economy.

cs.CR↗
arXiv2025

BootOOD: Self-Supervised Out-of-Distribution Detection via Synthetic Sample Exposure under Neural Collapse

Out-of-distribution (OOD) detection is critical for deploying image classifiers in safety-sensitive environments, yet existing detectors often struggle when OOD samples are semantically similar to the in-distribution (ID) classes. We present BootOOD, a fully self-supervised OOD detection framework that bootstraps exclusively from ID data and is explicitly designed to handle semantically challenging OOD samples. BootOOD synthesizes pseudo-OOD features through simple transformations of ID representations and leverages Neural Collapse (NC), where ID features cluster tightly around class means with consistent feature norms. Unlike prior approaches that aim to constrain OOD features into subspaces orthogonal to the collapsed ID means, BootOOD introduces a lightweight auxiliary head that performs radius-based classification on feature norms. This design decouples OOD detection from the primary classifier and imposes a relaxed requirement: OOD samples are learned to have smaller feature norms than ID features, which is easier to satisfy when ID and OOD are semantically close. Experiments on CIFAR-10, CIFAR-100, and ImageNet-200 show that BootOOD outperforms prior post-hoc methods, surpasses training-based methods without outlier exposure, and is competitive with state-of-the-art outlier-exposure approaches while maintaining or improving ID accuracy.

cs.CV↗
arXiv2025

Controlling Performance and Budget of a Centralized Multi-agent LLM System with Reinforcement Learning

Large language models (LLMs) exhibit complementary strengths across domains and come with varying inference costs, motivating the design of multi-agent LLM systems where specialized models collaborate efficiently. Existing approaches predominantly rely on decentralized frameworks, which invoke multiple LLMs for every input and thus lead to substantial and uncontrolled inference costs. In this work, we introduce a centralized multi-LLM framework, where a controller LLM selectively coordinates a pool of expert models in a cost-efficient and cost-controllable manner. We formulate this coordination problem as reinforcement learning with dual objectives: maximizing task performance while minimizing the overall inference cost. In addition, we expect the multi-agent system to have adapted behavior with different budget conditions during inference. To this end, we propose CoRL, a reinforcement learning framework that optimizes the performance cost trade-off in a controllable multi-budget setting. Experiments on four diverse benchmarks demonstrate that CoRL enables a single system to surpass the best expert LLM under high-budget settings, while maintaining strong performance in more economical low-budget modes, highlighting the effectiveness of centralized coordination for scalable and cost-efficient multi-agent LLM systems.

cs.CL↗
arXiv2025

COMPASS: Benchmarking Constrained Optimization in LLM Agents

Human decision-making often involves constrained optimization. As LLM agents are deployed to assist with real-world tasks like travel planning, shopping, and scheduling, they must mirror this capability. We introduce COMPASS, a benchmark that evaluates whether LLM agents can perform constrained optimization in realistic travel planning settings. To success in these tasks, agents must engage in multi-turn conversations with user to gather task information as well as use tools to gather information from the database. Then agents must propose a solution that not only satisfies hard constraints but also optimizes user's utility objective. Evaluating state-of-the-art models, we reveal a significant feasible-optimal gap: while models achieve 70-90% feasibility (constraint satisfaction), they reach only 20-60% optimality (utility optimization). Our analysis shows that tool use is not the bottleneck. Instead, the core limitation is insufficient exploration of the search space, with success strongly correlating with information gathered. Coding agents show a promising approach to mitigate this gap. Together, COMPASS provides a testbed for developing LLM agents that can truly mirror human decision-making by both satisfying constraints and optimizing objectives.

cs.LG↗
arXiv2025

To Backtrack or Not to Backtrack: When Sequential Search Limits Model Reasoning

Recent advancements in large language models (LLMs) have significantly improved their reasoning abilities, particularly through techniques involving search and backtracking. Backtracking naturally scales test-time compute by enabling sequential, linearized exploration via long chain-of-thought (CoT) generation. However, this is not the only strategy for scaling test time-compute: parallel sampling with best-of-N selection provides an alternative that generates diverse solutions simultaneously. Despite the growing adoption of sequential search, its advantages over parallel sampling-especially under a fixed compute budget-remain poorly understood. In this paper, we systematically compare these two approaches on two challenging reasoning tasks: CountDown and Sudoku. Surprisingly, we find that sequential search underperforms parallel sampling on CountDown but outperforms it on Sudoku, suggesting that backtracking is not universally beneficial. We identify two factors that can cause backtracking to degrade performance: (1) training on fixed search traces can lock models intro suboptimal strategies, and (2) explicit CoT supervision can discourage implicit (non verbalized) reasoning. Extending our analysis to reinforcement learning (RL), we show that models with backtracking capabilities benefit significantly from RL fine-tuning, while models without backtracking see limited, mixed gains. Together, these findings challenge the assumption that backtracking universally enhances LLM reasoning, instead revealing a complex interaction between task structure, training data, model scale, and learning paradigm.

cs.LG↗
arXiv2025

Sometimes I am a Tree: Data Drives Unstable Hierarchical Generalization

Early in training, LMs can behave like n-gram models, but eventually they often learn tree-based syntactic rules and generalize hierarchically out of distribution (OOD). We study this shift using controlled grammar-learning tasks: question formation and tense inflection. We find that a model learns to generalize hierarchically if its training data is _complex_-in particular, if it includes center-embedded clauses, a special syntactic structure. Under this definition, complex data drives hierarchical rules, while less complex data encourages shortcut learning in the form of n-gram-like linear rules. Furthermore, we find that a model uses rules to generalize, whether hierarchical or linear, if its training data is _diverse_-in particular, if it includes many distinct syntax trees in the training set. Under this definition, diverse data promotes stable rule learning, whereas less diverse data promotes memorization of individual syntactic sequences. Finally, intermediate diversity and intermediate complexity form an *unstable regime*, which is characterized by oscillatory learning dynamics and inconsistent behaviors across random seeds. These results highlight the central role of training data in shaping generalization and explain why competing strategies can lead to unstable outcomes.

cs.LG↗
arXiv2025

Limit Analysis for Symbolic Multi-step Reasoning Tasks with Information Propagation Rules Based on Transformers

Transformers are able to perform reasoning tasks, however the intrinsic mechanism remains widely open. In this paper we propose a set of information propagation rules based on Transformers and utilize symbolic reasoning tasks to theoretically analyze the limit reasoning steps. We show that the limit number of reasoning steps is between $O(3^{L-1})$ and $O(2^{L-1})$ for a model with $L$ attention layers in a single-pass.

cs.AI↗
arXiv2025

PRIME: Plasticity-Robust Incremental Model for Encrypted Traffic Classification in Dynamic Network Environments

With the continuous development of network environments and technologies, ensuring cyber security and governance is increasingly challenging. Network traffic classification(ETC) can analyzes attributes such as application categories and malicious intent, supporting network management services like QoS optimization, intrusion detection, and targeted billing. As the prevalence of traffic encryption increases, deep learning models are relied upon for content-agnostic analysis of packet sequences. However, the emergence of new services and attack variants often leads to incremental tasks for ETC models. To ensure model effectiveness, incremental learning techniques are essential; however, recent studies indicate that neural networks experience declining plasticity as tasks increase. We identified plasticity issues in existing incremental learning methods across diverse traffic samples and proposed the PRIME framework. By observing the effective rank of model parameters and the proportion of inactive neurons, the PRIME architecture can appropriately increase the parameter scale when the model's plasticity deteriorates. Experiments show that in multiple encrypted traffic datasets and different category increment scenarios, the PRIME architecture performs significantly better than other incremental learning algorithms with minimal increase in parameter scale.

cs.NI↗
arXiv2025

Can Interpretation Predict Behavior on Unseen Data?

Interpretability research often predicts model responses to targeted mechanistic interventions. But can we predict responses to unseen input data? We propose and demonstrate this alternate objective by using model internals to predict their out-of-distribution (OOD) behavior. We train hundreds of Transformers on simple synthetic tasks, where perfect in-distribution accuracy is compatible with multiple OOD generalization rules. We successfully use attention patterns -- observed only on in-distribution data -- to predict which rule each model follows on OOD data. Our experiments decouple the mechanistic faithfulness of our interpretation from its predictive value; ablations reveal such internal patterns can suppress rather than support the rule they predict, showing observational analysis can forecast behavior even when causal analysis fails to support a simple cause-effect link. Our findings are a proof-of-concept for a new interpretability objective: understanding model internals to predict behavior and assess reliability under distribution shift.

cs.LG↗
arXiv2025

Decomposing Elements of Problem Solving: What "Math" Does RL Teach?

Mathematical reasoning tasks have become prominent benchmarks for assessing the reasoning capabilities of LLMs, especially with reinforcement learning (RL) methods such as GRPO showing significant performance gains. However, accuracy metrics alone do not support fine-grained assessment of capabilities and fail to reveal which problem-solving skills have been internalized. To better understand these capabilities, we propose to decompose problem solving into fundamental capabilities: Plan (mapping questions to sequences of steps), Execute (correctly performing solution steps), and Verify (identifying the correctness of a solution). Empirically, we find that GRPO mainly enhances the execution skill-improving execution robustness on problems the model already knows how to solve-a phenomenon we call temperature distillation. More importantly, we show that RL-trained models struggle with fundamentally new problems, hitting a 'coverage wall' due to insufficient planning skills. To explore RL's impact more deeply, we construct a minimal, synthetic solution-tree navigation task as an analogy for mathematical problem-solving. This controlled setup replicates our empirical findings, confirming RL primarily boosts execution robustness. Importantly, in this setting, we identify conditions under which RL can potentially overcome the coverage wall through improved exploration and generalization to new solution paths. Our findings provide insights into the role of RL in enhancing LLM reasoning, expose key limitations, and suggest a path toward overcoming these barriers. Code is available at https://github.com/cfpark00/RL-Wall.

cs.AI↗
arXiv2025

On Kernel-based Variational Autoencoder

In this paper, we bridge Variational Autoencoders (VAEs) and kernel density estimations (KDEs) by approximating the posterior by KDEs and deriving an upper bound of the Kullback-Leibler (KL) divergence in the evidence lower bound (ELBO). The flexibility of KDEs makes the optimization of posteriors in VAEs possible, which not only addresses the limitations of Gaussian latent space in vanilla VAE but also provides a new perspective of estimating the KL-divergence in ELBO. Under appropriate conditions, we show that the Epanechnikov kernel is the optimal choice in minimizing the derived upper bound of KL-divergence asymptotically. Compared with Gaussian kernel, Epanechnikov kernel has compact support which should make the generated sample less noisy and blurry. The implementation of Epanechnikov kernel in ELBO is straightforward as it lies in the "location-scale" family of distributions where the reparametrization tricks can be directly employed. A series of experiments on benchmark datasets such as MNIST, Fashion-MNIST, CIFAR-10 and CelebA further demonstrate the superiority of Epanechnikov Variational Autoenocoder (EVAE) over vanilla VAE in the quality of reconstructed images, as measured by the FID score and Sharpness.

stat.ML↗