CIRF: Tokenizing Chain-of-Thoughts into Reusable Functional Units for Efficient Latent Reasoning in Large Language Models

arXiv — cs.CL•Thursday, May 28, 2026 at 4:00:00 AM

PositiveArtificial Intelligence

What Happened
A new framework named CIRF has been introduced, which tokenizes Chain-of-Thoughts (CoT) into reusable functional units, enhancing the efficiency of reasoning in large language models (LLMs). This approach allows for dynamic reasoning through discrete functional tokens, aligning latent reasoning with explicit rationales.
Why It Matters
The development of CIRF is significant as it addresses the limitations of existing CoT methods, particularly their lack of adaptability to complex examples and alignment with explicit rationales, potentially improving the performance of LLMs in various reasoning tasks.
The Bigger Picture
This advancement reflects a broader trend in AI research focusing on enhancing reasoning capabilities in LLMs, with various frameworks emerging that integrate CoT with other methodologies, such as graph representation learning and multi-teacher distillation, indicating a shift towards more sophisticated and adaptable reasoning processes in artificial intelligence.

— via World Pulse Now AI Editorial System

Read Original

Was this article worth reading? Share it

Continue Readings

arXiv — cs.CL18 hours ago

ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing

PositiveArtificial Intelligence

The introduction of ALIGNBEAM represents a significant advancement in the safety of large language models (LLMs) by enabling inference-time alignment transfer through cross-vocabulary logit mixing. This method allows for the translation of logits from a safe anchor model into the vocabulary of a target model, enhancing the refusal rate of harmful prompts without requiring model retraining.

Read full article

via arXiv — cs.CL

arXiv — cs.CL18 hours ago

DEFINED: A Data-Efficient Computational Framework for Fine-Grained Creativity Assessment in Debate Scenarios

PositiveArtificial Intelligence

A new computational framework named DEFINED has been proposed for assessing creativity in debate scenarios, addressing the limitations of current automated scoring methods that rely heavily on human evaluation. This framework utilizes a hierarchical eight-dimensional metric system to operationalize creativity, particularly in complex environments where traditional assessments fall short.

Read full article

via arXiv — cs.CL

arXiv — cs.CL18 hours ago

On the Optimal Reasoning Length for RL-Trained Language Models

NeutralArtificial Intelligence

A recent study published on arXiv explores the optimal reasoning length for reinforcement learning (RL)-trained language models, revealing that while increased output length can enhance reasoning, it also raises computational costs. The research indicates that accuracy peaks at an intermediate output length, with mode accuracy improving even as sample accuracy plateaus or declines.

Read full article

via arXiv — cs.CL

arXiv — cs.CL18 hours ago

Can News Predict the Market? Limits of Zero-Shot Financial NLP and the Role of Explainable AI

NeutralArtificial Intelligence

A recent study explores the ability of financial news to predict short-term stock movements using a zero-shot natural language processing framework. Despite advancements in large language models, the research indicates that these models struggle to extract actionable signals from financial news without domain-specific training, leading to a structured pipeline that incorporates temporal aggregation and an explainability framework.

Read full article

via arXiv — cs.CL

arXiv — cs.CL18 hours ago

Judging Against the Reference: Uncovering Knowledge-Driven Failures in LLM-Judges on QA Evaluation

NeutralArtificial Intelligence

Recent research highlights significant failures in large language models (LLMs) when used as judges for question answering (QA) tasks, particularly when the reference answers conflict with the models' inherent knowledge, leading to unreliable evaluation scores.

Read full article

via arXiv — cs.CL

arXiv — cs.CL2 days ago

Where You Inject Diversity Matters: A Unified Framework for Diverse Generation

PositiveArtificial Intelligence

A new framework for enhancing diversity in open-ended generation tasks has been introduced, addressing the limitations of large language models that often produce similar outputs. This framework characterizes diverse generation methods by the source of diversity introduced during generation and includes a transmission score to measure the effectiveness of this variation in the final output.

Read full article

via arXiv — cs.CL

arXiv — cs.LG2 days ago

Generative Archetype-Grounded Item Representations for Sequential Recommendation

PositiveArtificial Intelligence

A new framework named GenAIR has been proposed to enhance sequential recommendation systems by utilizing Generative Archetype-grounded Item Representations. This approach addresses the limitations of existing methods that rely on static item attributes, thereby improving the prediction of users' next interactions based on their historical behavior.

Read full article

via arXiv — cs.LG

arXiv — cs.CL2 days ago

Large Language Models as Modal Models in Linguistics

NeutralArtificial Intelligence

The rapid advancement of large language models (LLMs) has sparked significant debates within linguistic theory, categorized into three main positions: insulationism, eliminativism, and conciliationism. This discourse highlights the epistemic value of LLMs as minimal models, which can provide insights into language acquisition and linguistic competence despite lacking structural correspondence to human cognition.

Read full article

via arXiv — cs.CL

Ready to build your own newsroom?

Subscribe to unlock a personalised feed, podcasts, newsletters, and notifications tailored to the topics you actually care about