Optimizing Test-Time Scaling for Large Language Models

The field of large language models (LLMs) is witnessing significant advancements in test-time scaling techniques, which aim to improve model performance without retraining. A key direction in this area is the development of methods that optimize resource allocation during test-time scaling, ensuring efficient use of computational resources. Researchers are exploring innovative approaches, such as dynamic budget allocation, speculative decoding, and latent steering vectors, to improve the accuracy and speed of LLMs. These techniques have shown promising results in various benchmarks, including mathematical reasoning and multilingual tasks. Notably, some papers have made significant contributions to the field, including:

  • Every Rollout Counts, which proposes a provably optimal method for resource allocation during test-time search.
  • Bohdi, which enables heterogeneous LLM fusion with automatic data exploration.
  • SPECS, which introduces a latency-aware test-time scaling method using speculative drafts.
  • Fractional Reasoning, which allows for continuous control over reasoning intensity at inference time.
  • DynScaling, which proposes an efficient verifier-free inference scaling method via dynamic and integrated sampling.
  • Lookahead Reasoning, which raises the algorithmic ceiling for speculative decoding.
  • When Life Gives You Samples, which studies the benefits of scaling up inference compute for multilingual LLMs.
  • Utility-Driven Speculative Decoding for Mixture-of-Experts, which presents a framework for selectively enabling speculation to avoid slowdowns.
  • Test-time Scaling Techniques in Theoretical Physics, which evaluates the effectiveness of test-time scaling methods on the TPBench physics dataset.

Sources

Every Rollout Counts: Optimal Resource Allocation for Efficient Test-Time Scaling

Bohdi: Heterogeneous LLM Fusion with Automatic Data Exploration

$\texttt{SPECS}$: Faster Test-Time Scaling through Speculative Drafts

Fractional Reasoning via Latent Steering Vectors Improves Inference Time Compute

DynScaling: Efficient Verifier-free Inference Scaling via Dynamic and Integrated Sampling

Scaling Speculative Decoding with Lookahead Reasoning

When Life Gives You Samples: The Benefits of Scaling up Inference Compute for Multilingual LLMs

Utility-Driven Speculative Decoding for Mixture-of-Experts

Test-time Scaling Techniques in Theoretical Physics -- A Comparison of Methods on the TPBench Dataset

Built with on top of