大模型
arxiv.org ·
LLM 概率集中：对齐如何缩小生成视界
该 arXiv 论文（v4）研究对齐大模型输出多样性下降的机制，提出用 Branching Factor（BF，输出分布熵的指数化长度平均，可理解为生成时有效候选步数）量化概率集中。实证发现：BF 随生成推进而下降，受控干预显示这主要是自回归自条件的任务无关属性，与对齐无关，且可被意外上下文局部逆转；对齐微调从一开始就锐化分布，整体使 BF 降低 2–5 倍，早期位置最多降低一个数量级（如 12 降至 1.2），这解释了对齐模型对解码策略不敏感，也说明对齐 CoT 模型（如 DeepSeek 蒸馏模型）生成更长推理链、进入更确定阶段而输出更稳定。作者假设对齐并未根本改变模型行为，而是将模型导向 'Sure' 等风格 token，从而解锁基座模型已有的低熵轨迹，nudging 实验支持该观点。
AI推荐理由 arXiv 论文（v4，2026-09-12 修订），作者 Chenghao Yang、Sida Li、Ari Holtzman，摘要给出可验证的量化指标与实验结论：用 Branching Factor（BF，输出分布熵的指数化长度平均）衡量生成多样性；发现 BF 随生成推进下降，且该下降主要是自回归自条件的任务无关属性，可通过意外上下文局部逆转；对齐微调从一开始就锐化分布，整体使 BF 降低 2–5 倍，早期位置最多降低一个数量级（如 12 降至 1.2），并解释了对齐模型对解码策略不敏感、CoT 模型生成更长推理链进入低 BF 阶段更稳定；提出对齐可能只是把模型导向 'Sure' 等风格 token 以解锁基座模型已有的低熵轨迹，nudging 实验支持该假设。新信息具体、有诊断价值，但结论为作者假设与实验观察，尚未经独立复现，故未给更高分。
82
正文
Computer Science > Computation and Language
arXiv:2506.17871 (cs)
[Submitted on 22 Jun 2025 (v1), last revised 12 Sep 2026 (this version, v4)]
Title:LLM Probability Concentration: How Alignment Shrinks the Generative Horizon
Authors:Chenghao Yang, Sida Li, Ari Holtzman
View a PDF of the paper titled LLM Probability Concentration: How Alignment Shrinks the Generative Horizon, by Chenghao Yang and 2 other authors
View PDF
HTML (experimental)
Abstract:Despite their impressive capabilities, aligned large language models (LLMs) often generate outputs that lack diversity. What drives this consistency in the generation? We investigate this phenomenon through the lens of probability concentration in the model's output distribution. To quantify it, we use the Branching Factor (BF)--the exponentiated length-averaged entropy of the output distribution, interpreted as the effective number of plausible next steps during generation. Our empirical analysis reveals two key findings: (1) BF often decreases as generation progresses, suggesting that LLMs become more predictable; a controlled intervention indicates that this decline is largely a task-independent property of autoregressive self-conditioning, distinct from alignment, and can be locally reversed by unexpected context. (2) Alignment tuning sharpens the output distribution from the outset, reducing BF by a factor of 2--5 overall and up to an order of magnitude (e.g., from 12 to 1.2) at early positions. This reduction helps explain why aligned models are less sensitive to decoding strategies. It also has implications for complex reasoning: aligned Chain-of-Thought (CoT) models (e.g., DeepSeek-distilled models) generate longer reasoning chains that reach later, more deterministic (lower-BF) stages, yielding more stable outputs. We hypothesize that alignment does not fundamentally change model behavior, but instead steers the model toward stylistic tokens (e.g., ``Sure'') that unlock low-entropy trajectories already present in the base model. Nudging experiments support this view: prompting base models with such tokens similarly reduces BF. Together, our findings establish BF as a diagnostic for understanding and controlling LLM outputs, clarifying how alignment reduces variability, CoT stabilizes generation, and base models can be steered away from diversity.
Comments:
Subjects:
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Cite as:
arXiv:2506.17871 [cs.CL]
(or
arXiv:2506.17871v4 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2506.17871
Focus to learn more
arXiv-issued DOI via DataCite
Submission history
From: Chenghao Yang [view email]
[v1]
Sun, 22 Jun 2025 02:00:37 UTC (762 KB)
[v2]
Tue, 14 Oct 2025 22:47:33 UTC (1,089 KB)
[v3]
Mon, 2 Mar 2026 19:30:01 UTC (1,256 KB)
[v4]
Sat, 12 Sep 2026 23:41:05 UTC (1,479 KB)
Full-text links:
Access Paper:
View a PDF of the paper titled LLM Probability Concentration: How Alignment Shrinks the Generative Horizon, by Chenghao Yang and 2 other authors
View PDF
HTML (experimental)
TeX Source
view license
Current browse context:
cs.CL
new
|
recent
| 2025-06
Change to browse by:
cs
cs.AI
cs.LG
References & Citations
NASA ADS
Google Scholar
Semantic Scholar
export BibTeX citation
Loading...
BibTeX formatted citation
×
loading...
Data provided by:
Bookmark
Bibliographic Tools
Bibliographic and Citation Tools
Bibliographic Explorer Toggle
Bibliographic Explorer (What is the Explorer?)
Connected Papers Toggle
Connected Papers (What is Connected Papers?)
Litmaps Toggle
Litmaps (What is Litmaps?)
scite.ai Toggle
scite Smart Citations (What are Smart Citations?)
Code, Data, Media
Code, Data and Media Associated with this Article
alphaXiv Toggle
alphaXiv (What is alphaXiv?)
Links to Code Toggle
CatalyzeX Code Finder for Papers (What is CatalyzeX?)
DagsHub Toggle
DagsHub (What is DagsHub?)
GotitPub Toggle
Gotit.pub (What is GotitPub?)
Huggingface Toggle
Hugging Face (What is Huggingface?)
ScienceCast Toggle
ScienceCast (What is ScienceCast?)
Demos
Demos
Replicate Toggle
Replicate (What is Replicate?)
Spaces Toggle
Hugging Face Spaces (What is Spaces?)
Spaces Toggle
TXYZ.AI (What is TXYZ.AI?)
Related Papers
Recommenders and Search Tools
Link to Influence Flower
Influence Flower (What are Influence Flowers?)
Core recommender toggle
CORE Recommender (What is CORE?)
Author
Venue
Institution
Topic
About arXivLabs
arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Which authors of this paper are endorsers? |
Disable MathJax (What is MathJax?)