arXiv subjects

You Shan

Publications and source records attributed to You Shan.

6 recordsLinked to original sources

arXiv2025

Towards Analyzing and Understanding the Limitations of VAPO: A Theoretical Perspective

The VAPO framework has demonstrated significant empirical success in enhancing the efficiency and reliability of reinforcement learning for long chain-of-thought (CoT) reasoning tasks with large language models (LLMs). By systematically addressing challenges such as value model bias, heterogeneous sequence lengths, and sparse reward signals, VAPO achieves state-of-the-art performance. While its practical benefits are evident, a deeper theoretical understanding of its underlying mechanisms and potential limitations is crucial for guiding future advancements. This paper aims to initiate such a discussion by exploring VAPO from a theoretical perspective, highlighting areas where its assumptions might be challenged and where further investigation could yield more robust and generalizable reasoning agents. We delve into the intricacies of value function approximation in complex reasoning spaces, the optimality of adaptive advantage estimation, the impact of token-level optimization, and the enduring challenges of exploration and generalization.

cs.LG↗

arXiv2025

VQ-Logits: Compressing the Output Bottleneck of Large Language Models via Vector Quantized Logits

Large Language Models (LLMs) have achieved remarkable success but face significant computational and memory challenges, particularly due to their extensive output vocabularies. The final linear projection layer, mapping hidden states to vocabulary-sized logits, often constitutes a substantial portion of the model's parameters and computational cost during inference. Existing methods like adaptive softmax or hierarchical softmax introduce structural complexities. In this paper, we propose VQ-Logits, a novel approach that leverages Vector Quantization (VQ) to drastically reduce the parameter count and computational load of the LLM output layer. VQ-Logits replaces the large V * dmodel output embedding matrix with a small, shared codebook of K embedding vectors (K << V ). Each token in the vocabulary is mapped to one of these K codebook vectors. The LLM predicts logits over this compact codebook, which are then efficiently "scattered" to the full vocabulary space using the learned or preassigned mapping. We demonstrate through extensive experiments on standard language modeling benchmarks (e.g., WikiText-103, C4) that VQ-Logits can achieve up to 99% parameter reduction in the output layer and 6x speedup in logit computation, with only a marginal 4% increase in perplexity compared to full softmax baselines. We further provide detailed ablation studies on codebook size, initialization, and learning strategies, showcasing the robustness and effectiveness of our approach.

cs.CL↗

arXiv2025

ComplexFormer: Disruptively Advancing Transformer Inference Ability via Head-Specific Complex Vector Attention

Transformer models rely on self-attention to capture token dependencies but face challenges in effectively integrating positional information while allowing multi-head attention (MHA) flexibility. Prior methods often model semantic and positional differences disparately or apply uniform positional adjustments across heads, potentially limiting representational capacity. This paper introduces ComplexFormer, featuring Complex Multi-Head Attention-CMHA. CMHA empowers each head to independently model semantic and positional differences unified within the complex plane, representing interactions as rotations and scaling. ComplexFormer incorporates two key improvements: (1) a per-head Euler transformation, converting real-valued query/key projections into polar-form complex vectors for head-specific complex subspace operation; and (2) a per-head adaptive differential rotation mechanism, exp[i(Adapt(ASmn,i) + Delta(Pmn),i)], allowing each head to learn distinct strategies for integrating semantic angle differences (ASmn,i) with relative positional encodings (Delta(Pmn),i). Extensive experiments on language modeling, text generation, code generation, and mathematical reasoning show ComplexFormer achieves superior performance, significantly lower generation perplexity , and improved long-context coherence compared to strong baselines like RoPE-Transformers. ComplexFormer demonstrates strong parameter efficiency, offering a more expressive, adaptable attention mechanism.

cs.LG↗

arXiv2025

How General Are Measures of Choice Consistency? Evidence from Experimental and Scanner Data

Choice consistency with utility maximization, as a key assumption in economics, has been extensively used to evaluate decision quality of individuals and to predict real-world outcomes across different contexts. Here we investigate the generalizability of consistency measures derived from budgetary decisions in the lab-in-the-field experiment and purchasing decisions using supermarket scanner data. In the first study, we observe a lack of correlation between consistency scores derived from risky decisions in the experiment and those from supermarket food purchasing decisions. In the second study, we observe moderate correlations between experimental tasks and low to moderate correlations across purchasing categories and over time periods within the supermarket. Moreover, consistency in the two settings exhibits distinct predictive validity in predicting consumer behavior. These results suggest that choice consistency, as a measure of decision quality, may be better characterized as a multidimensional skill set rather than a single-dimensional ability.

econ.GN↗

arXiv2023

The Emergence of Economic Rationality of GPT

As large language models (LLMs) like GPT become increasingly prevalent, it is essential that we assess their capabilities beyond language processing. This paper examines the economic rationality of GPT by instructing it to make budgetary decisions in four domains: risk, time, social, and food preferences. We measure economic rationality by assessing the consistency of GPT's decisions with utility maximization in classic revealed preference theory. We find that GPT's decisions are largely rational in each domain and demonstrate higher rationality score than those of human subjects in a parallel experiment and in the literature. Moreover, the estimated preference parameters of GPT are slightly different from human subjects and exhibit a lower degree of heterogeneity. We also find that the rationality scores are robust to the degree of randomness and demographic settings such as age and gender, but are sensitive to contexts based on the language frames of the choice situations. These results suggest the potential of LLMs to make good decisions and the need to further understand their capabilities, limitations, and underlying mechanisms.

econ.GN↗

arXiv2021

SURPRISE! and When to Schedule It

Information flow measures, over the duration of a game, the audience's belief of who will win, and thus can reflect the amount of surprise in a game. To quantify the relationship between information flow and audiences' perceived quality, we conduct a case study where subjects watch one of the world's biggest esports events, LOL S10. In addition to eliciting information flow, we also ask subjects to report their rating for each game. We find that the amount of surprise in the end of the game plays a dominant role in predicting the rating. This suggests the importance of incorporating when the surprise occurs, in addition to the amount of surprise, in perceived quality models. For content providers, it implies that everything else being equal, it is better for twists to be more likely to happen toward the end of a show rather than uniformly throughout.

cs.MA↗