Hugging Face Daily PapersJaejun Shim, HyunJin Kim, Young Jin KimSep 17, 20261 min readpaperadvanced

When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models

Summary

When2Think introduces a post‑training framework that lets a large reasoning model decide per‑instance how much reasoning depth to allocate, using difficulty‑aware reward shaping (IDAC) and verifier rewards. It cuts token usage by ~28% while boosting Pass@3 by 10% on AIME24 and reaches 40% Pass@3 on AIME25, outperforming compression and routing baselines.

Formulates efficient reasoning as instance‑adaptive computation allocation.
IDAC uses pre‑computed accuracy/token statistics to shape rewards without a learned reward model.
Combines verifier‑based rewards and batch‑wise standardized advantages for stable, critic‑free optimization.
Hybrid “System 1 vs System 2” behavior: direct answer on easy inputs, extended reasoning on hard ones.

Large reasoning models waste compute on trivial queries and still fail on hard ones. A lightweight, post‑training control that adapts compute per instance can lower inference cost in production while preserving or improving accuracy, a practical win for any service deploying LLM‑based reasoning.

7/10

llm machine learning performance reasoning transformers

Related reading

Register Tokens for Bounded-State Reasoning in Diffusion Language Models arxiv.org
Register tokens are fixed‑position embeddings that store a compact hidden state across diffusion‑based language model generation chunks, enabling bounded‑state reasoning without retaining all prior text. Post‑training on LLaDA and Dream shows up to +8.5 math and +19.5 code benchmark points versus plain text carry, and RL fine‑tuning further improves long‑horizon tasks.
Hugging Face Daily Papersarxiv.orgAlbert Ge, Chandan Singh, Yufan Zhuang8d1 minpaper
When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation arxiv.org
On-policy distillation (OPD) can lead to excessively long student responses, a phenomenon called length inflation. This paper identifies "termination-token mismatch" between base students and post-trained teachers as a key source, where models place stopping probability on different EOS tokens. Treating functionally equivalent EOS tokens as a shared semantic stopping action substantially mitigate…
Hugging Face Daily Papersarxiv.orgYuxiao Yang, Tianrun Yu, Shangzhe Li5d1 minpaper
RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents arxiv.org
The paper presents RecreationWorld, a five‑platform framework that lets hybrid computer‑use agents learn by recreating the behavior of a running reference, and introduces RecreationBench, a 250‑task benchmark with programmatic and visual assertions. Experiments show GPT‑6 Astra reaches 58.1% overall but struggles with deeper programmatic tests, highlighting gaps in current agents.
Hugging Face Daily Papersarxiv.orgShuai Bai, Jiayong Deng, Yikun Fu4d2 minpaper
E2A-Bench: Benchmarking Evidence-to-Action Reliability in Financial Chart Reasoning arxiv.org
E2A‑Bench is a 969‑query benchmark for financial chart reasoning that evaluates vision‑language models across a full evidence‑to‑action chain using four metrics (UCR, RCI, ECI, NDR). Experiments on 20 VLMs expose hidden failures: low‑UCR models have only 6.4 % directional coverage, oracle‑aided verification cuts unsupported claims but can kill coverage, and fine‑tuning inflates BUY:SELL ratios by…
Hugging Face Daily Papersarxiv.orgXiaoya Wang, Yutong Xu, Junjie Wang9d1 minpaper
Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States arxiv.org
Mind2Dialogue introduces a psychology‑guided simulator that generates user mental states and uses privileged distillation to train LLM assistants on Oracle responses that know those states. The resulting models improve personalization and theory‑of‑mind metrics by up to 41 percentage points versus standard instruction‑tuned baselines.
Hugging Face Daily Papersarxiv.orgZixuan Wang, Yufan Zhou, Jinzhou Tang8d2 minpaper
SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness arxiv.org
SoL-Pi is a new approach for scaling auto-research loops in coding agents, focusing on token efficiency for long reasoning trajectories. It achieves comparable performance to existing methods while significantly reducing token traffic by 44.7-49.0% and API costs by about one third on the 51-task EdgeBench evaluation.
Hugging Face Daily Papersarxiv.orgHaozhe Liu, Tian Ye, Sensen Gao5d1 minpaper

When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models

Summary

Related reading

Register Tokens for Bounded-State Reasoning in Diffusion Language Modelsarxiv.org

When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillationarxiv.org

RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agentsarxiv.org

E2A-Bench: Benchmarking Evidence-to-Action Reliability in Financial Chart Reasoningarxiv.org

Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental Statesarxiv.org

SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harnessarxiv.org

Register Tokens for Bounded-State Reasoning in Diffusion Language Models arxiv.org

When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation arxiv.org

RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents arxiv.org

E2A-Bench: Benchmarking Evidence-to-Action Reliability in Financial Chart Reasoning arxiv.org

Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States arxiv.org

SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness arxiv.org