Source-linked AI summary
Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time Scaling
Falcon LLM Team, Iheb Chaabane, Puneesh Khanna, Suhail Mohmad, Slim Frikha, Shi Hu, Abdalgader Abubaker, Reda Alami, Mikhail Lubinets, Mohamed El Amine Seddik, Hakim Hacid
TL;DR
Reasoning models often require costly test-time scaling, motivating compact systems that retain strong accuracy while reducing inference overhead. Falcon-H1R combines a hybrid Transformer–Mamba architecture with curated SFT, reinforcement learning, and DeepConf evaluation, achieving competitive benchmark performance and improved accuracy–cost efficiency. The reported results support the feasibility of small reasoning models, with evaluation contamination remaining very low except for near-zero MMLU-Pro levels.
Problem
Test-time scaling can improve reasoning but requires substantial computation, creating a need for efficient models that handle parallel inference and long sequences.
Method
Falcon-H1R uses a 7B hybrid Transformer–Mamba architecture trained with curated supervised fine-tuning and reinforcement learning, then evaluated with DeepConf test-time scaling.
Results
Falcon-H1R matches or exceeds larger reasoning models across challenging benchmarks and, with DeepConf, reaches 96.7% AIME25 accuracy using 38% fewer tokens than DeepSeek-R1-0528-Qwen3-8B.
Takeaways & Limitations
The results indicate that a compact model can combine competitive reasoning accuracy with faster inference, lower token use, and reduced computational cost.
Takeaways & Limitations
Exact-match analysis found near-zero contamination for MMLU-Pro rather than zero contamination: 0.0005% in SFT data and 0.035% in RL data.
Abstract
from arXiv · showhide
This work introduces Falcon-H1R, a 7B-parameter reasoning-optimized model that establishes the feasibility of achieving competitive reasoning performance with small language models (SLMs). Falcon-H1R stands out for its parameter efficiency, consistently matching or outperforming SOTA reasoning models that are $2\times$ to $7\times$ larger across a variety of reasoning-intensive benchmarks. These results underscore the importance of careful data curation and targeted training strategies (via both efficient SFT and RL scaling) in delivering significant performance gains without increasing model size. Furthermore, Falcon-H1R advances the 3D limits of reasoning efficiency by combining faster inference (through its hybrid-parallel architecture design), token efficiency, and higher accuracy. This unique blend makes Falcon-H1R-7B a practical backbone for scaling advanced reasoning systems, particularly in scenarios requiring extensive chain-of-thoughts generation and parallel test-time scaling. Leveraging the recently introduced DeepConf approach, Falcon-H1R achieves state-of-the-art test-time scaling efficiency, offering substantial improvements in both accuracy and computational cost. As a result, Falcon-H1R demonstrates that compact models, through targeted model training and architectural choices, can deliver robust and scalable reasoning performance.
1. Introduction
Falcon-H1R addresses the high cost of test-time scaling with a 7B hybrid Transformer–Mamba model trained for reasoning. It combines efficient inference, targeted SFT and RL, and strong benchmark performance across accuracy, token use, and parallel reasoning.
- Motivation: Test-time scaling improves reasoning by allocating additional inference-time compute, but generating and evaluating many candidate solutions creates substantial computational costs.The paper motivates efficiency improvements for models handling large parallel batches and long sequences.
- Approach: Falcon-H1R is a 7B reasoning-optimized model built on a hybrid Transformer–Mamba architecture for efficient high-throughput inference.The architecture is designed for large batch sizes and parallel test-time scaling.
- Training: The model combines cold-start SFT on curated long reasoning traces with GRPO reinforcement learning across mathematics, code, and science.The training strategy emphasizes difficult data and supports response lengths up to 48K tokens.
- Results: Falcon-H1R achieves 88.1% on AIME24, 83.1% on AIME25, 64.9% on HMMT25, 36.3% on AMO-Bench, and 68.6% on LiveCodeBenchv6.These results compete with larger reasoning models including GPT-OSS-20B, Qwen3-32B, Phi-4-Reasoning-Plus-14B, and DeepSeek-R1-0528-Qwen3-8B.
- Results: Under DeepConf test-time scaling, Falcon-H1R-7B reaches 96.7% accuracy on AIME25 while using 38% fewer tokens than DeepSeek-R1-0528-Qwen3-8B.The reported gains are attributed to strong base reasoning performance and confidence estimates that support early stopping.
2. Cold-start SFT Stage
The cold-start SFT stage builds Falcon-H1R’s reasoning capabilities through curated, verified, difficulty-weighted data and targeted ablations. The final recipe uses balanced token normalization and a math-dominant mixture with code and science data.
- Data and rationale: The SFT stage is motivated by evidence that RLVR can improve average accuracy while discovering no broader solvable-problem coverage than the base model.Cold-start SFT is therefore used to expand the reasoning traces available before RL.
- Data and rationale: The training corpus spans mathematics, coding, STEM, instruction-following, tool calling, and general chat, organized into mathematics, coding, science, and other domains.Coding data emphasizes algorithmic reasoning and functional correctness, while math and science data prioritize verified answers.
- Data and rationale: Solutions were verified and filtered for correctness, valid final answers, executable code, and reliable tool outputs, while selected difficult unverifiable traces were retained.The filtering pipeline used mathematical verification and factual cross-validation with official web sources when applicable.
- Ablations and weighting: Twelve rollouts per problem were most effective, especially for difficult queries, whereas incorrect rollouts produced only marginal gains on the hardest problems.The benefit of rollout diversity increased with problem difficulty.
- Ablations and weighting: Cross-domain teacher mixing reduced evaluation scores and increased output entropy relative to the best single-teacher baseline.The authors hypothesize that conflicting reasoning styles introduced distribution shifts and inconsistencies.
- Ablations and weighting: Hard problems received 1.25×–1.75× weighting, medium problems retained weight 1, and easy problems were down-weighted to 0.5× or removed.The scheme focused training on difficult examples while reducing emphasis on trivial samples.
- Ablations and weighting: A math-dominant mixture with moderate code and science data outperformed more balanced or code-centric mixtures, supporting stronger transfer from math reasoning to other domains.The final mixture broadened cross-domain coverage while remaining math-dominant.
- Distributed training: Balanced Data-Parallel Token Normalization produced 4–10% higher AIME-25 accuracy throughout the first 3,750 SFT steps.The method corrects unequal valid-token contributions across data-parallel ranks and improves training consistency.
3. Reinforcement Learning Stage
The RL stage uses verified, difficulty-filtered math and code data with GRPO-based training, then evaluates domain and curriculum choices. Math-focused RL provides the strongest general reasoning gains, while code-only training specializes but generalizes less effectively.
- Training framework: RLVR further improves reasoning performance and output quality after supervised fine-tuning.The authors use a custom GRPO-based framework built on verl.
- Data preparation: Training data is verified, difficulty-filtered, and deduplicated against SFT data to support reliable rewards and prevent memorization.Difficulty is estimated from the percentage of correct solutions across eight sampled rollouts per problem.
- Difficulty filtering: 100% success-rate problems were removed, while 0% success-rate problems were undersampled and heavily filtered when solutions exceeded the token limit.The resulting math dataset had a mirrored J-shaped difficulty distribution.
- Training framework: GRPO modifications target convergence and stability through positive-sample cross-entropy, online sampling, adjusted clipping, and truncated importance sampling.The framework also removes KL and entropy regularization by setting β = 0 and γ = 0.
- Ablation findings: Math-only RL produced strong reasoning and generalization, whereas code-only RL led code benchmarks but weakened math and science transfer.Science-only RL produced no meaningful GPQA-Diamond improvement, and mixed-domain training did not outperform alternatives.
- Ablation findings: Sequential math-then-code training achieved the best average main-benchmark performance, but its gains were modest relative to the additional compute.Training monitoring showed stable entropy and response length near 20k during Stage 1.
4. Standard Reasoning Tasks Evaluation
Falcon-H1R-7B is evaluated across math, code, and general reasoning benchmarks using specified sampling and response-length settings. It matches or exceeds larger state-of-the-art reasoning models, with especially strong mathematical performance and competitive code and general results.
- Evaluation setup: The evaluation covers Math, Code, and General benchmark categories, using pass@1 with task-specific response counts and maximum lengths.The benchmark suite includes AIME24/25, HMMT25, AMO-Bench, MATH500, LiveCodeBench v6, SciCode, τ2-Telecom, TB Hard, GPQA-Diamond, MMLU-Pro, HLE, and IFBench.
- Evaluation setup: Evaluation integrity was supported by near-zero contamination: all benchmarks showed 0% exact-match contamination except MMLU-Pro, at 0.0005% for SFT data and 0.035% for RL data.HLE used 2,158 text-only problems, while TB Hard, τ2-Telecom, and SciCode were each repeated three times.
- Overall results: Falcon-H1R matches or exceeds state-of-the-art reasoning models across tasks despite having significantly fewer parameters.Baselines span approximately 7B to 32B parameters, including Qwen3-8B, GPT-OSS-20B, and Qwen3-32B.
- Mathematical reasoning: 88.1% on AIME24, 64.9% on HMMT25, 36.3% on AMO-Bench, and 97.4% on MATH500 were the highest reported scores, while AIME25 reached second place at 83.1%.On AMO-Bench, Falcon-H1R exceeded GPT-OSS-20B by over 10 percentage points.
- Code generation: 68.6% on LiveCodeBench v6 was the second-highest code score, while SciCode, TB Hard, and τ2-Telecom remained competitive with similarly sized models.GPT-OSS-20B was the only model reported to surpass Falcon-H1R on LiveCodeBench v6.
- General reasoning: Falcon-H1R achieved second-best results on HLE at 11.1 and IFBench at 53.4, while GPQA-Diamond and MMLU-Pro remained competitive but left room for improvement.The authors characterize the weaker knowledge-intensive results as a trade-off associated with the model’s reasoning emphasis.
5. Test-time Scaling
Falcon-H1R is evaluated with DeepConf@512 to test-time scale parallel reasoning while controlling inference cost. Across selected mathematical and scientific benchmarks, it combines high accuracy with fewer generated tokens and efficient confidence-based filtering.
- Evaluation setup: DeepConf dynamically filters parallel reasoning chains using model-derived confidence scores, terminating low-confidence paths early while continuing higher-potential chains.This design reduces computational overhead during test-time scaling.
- Evaluation setup: The evaluation uses a fixed trace budget of K = 512, with warm-up threshold estimation and early stopping based on confidence over 2,048-token windows.The aggressive filtering configuration sets η = 10%.
- Evaluation setup: DeepConf aggregation is compared using majority, mean confidence-weighted, tail confidence-weighted, bottom window-weighted, and minimum window-weighted voting.These strategies differ in how confidence scores determine the final answer.
- Evaluation results: The voting strategies converge to very equivalent accuracies, indicating that DeepConf produces a robust and flexible high-quality candidate trace set.The result is reported across the evaluated models and aggregation choices.
- Evaluation results: The TTS evaluation covers AIME-2024, AIME-2025, GPQA-D, and the parser-verifiable 39-problem subset of AMO-Bench.Final answers are extracted with the math_verify parsing framework rather than only the original boxed-expression rule.
- Evaluation results: Falcon-H1R consistently generates substantially fewer tokens while maintaining high accuracy across all evaluated benchmarks.The results are presented as advantages in solution quality and generated-token amount, alongside computational efficiency.
- Evaluation results: The combination of faster inference, token efficiency, and higher accuracy makes Falcon-H1R-7B practical for scaling systems that generate substantial parallel chain-of-thoughts.This positions the model for reasoning workloads using parallel test-time scaling.
6. Conclusion
Falcon-H1R is a 7B-parameter reasoning-optimized model showing that small language models can achieve state-of-the-art reasoning performance associated with much larger systems. Its hybrid architecture, SFT and RL training, and DeepConf integration support efficient and scalable reasoning with high accuracy and reduced resource use.
- Conclusion: Falcon-H1R is a 7B-parameter reasoning-optimized model demonstrating that small language models can achieve state-of-the-art reasoning performance associated with much larger systems.
- Conclusion: Its hybrid Transformer–Mamba architecture and training strategy combine supervised fine-tuning with reinforcement learning to deliver competitive accuracy and inference efficiency.
- Conclusion: Falcon-H1R consistently matches or surpasses larger state-of-the-art models across challenging reasoning benchmarks.
- Conclusion: The architecture enables faster inference, greater token efficiency, and effective parallelization for test-time scaling methods.
- Conclusion: DeepConf integration improves scalability and cost-effectiveness in test-time scaling while maintaining high accuracy and reducing resource consumption.
A. Training Optimizations
Training optimizations combine Liger Kernels with optimizer and gradient-accumulation improvements for Falcon-H1-7B. These changes reduce memory use and training overhead while improving long-context workload behavior.
- Kernel optimizations: Liger Kernels provide consistent speedups and a reduced memory footprint for Falcon-H1-7B training.
- Framework optimizations: Fused AdamW and efficient gradient accumulation are integrated into the verl framework.
- Framework optimizations: Together, the kernel and framework improvements reduce optimizer overhead and smooth gradient accumulation for long-context workloads.
B. Inference Analysis
The inference analysis compares Falcon-H1R-7B with transformer-based Qwen3-8B using vLLM across batch sizes from 2 to 128. It evaluates throughput under both shorter-input/longer-output and longer-input/shorter-output token configurations.
- Throughput comparison: vLLM inference throughput is compared between Falcon-H1R-7B and Qwen3-8B across batch sizes from 2 to 128.
- Benchmark settings: The benchmark averages configurations with 512 input tokens and 32K output tokens, or 8K input tokens and 16K output tokens.
- Benchmark settings: The comparison uses the optimal parallelism configuration for each model to provide a fair throughput evaluation.
C. RL Data Filtering Diagram
Figure 9 presents the RL data filtering procedure described in Section 3.1.
- Figure 9 depicts the RL data filtering procedure described in Section 3.1.
D. Benchmarks Descriptions
Table 9 provides an overview of the evaluation benchmarks considered in the paper.
- Table 9 summarizes the evaluation benchmarks considered in the study.
E. Safety Evaluation
Falcon-H1R-7B shows strong final-answer safety across a comprehensive evaluation, while chain-of-thought reasoning receives lower safety scores because it engages more directly with potentially harmful content. This distinction has deployment implications for exposing raw reasoning traces.
- 81,970 prompts were used to evaluate jailbreak attempts, adversarial prompts, and harmful content detection across three output configurations.
- 98.18% CoT+Answer and 98.19% Answer Only weighted averages indicate robust safety alignment in final outputs.
- 92.60% CoT safety versus 98.19% final-answer safety yields a 5.59 percentage point gap observed across nearly all benchmarks.
- CatHarmfulQA records 82.16% CoT safety versus 98.0% Answer safety, illustrating the difference between reasoning and final responses.
- Answer Only scores range from 90.0% to 100.0% across benchmarks, supporting effective refusal or safe responses to harmful queries.
- Raw reasoning traces may contain concerning exploratory content out of context, so deployment may favor showing final answers while retaining traces for auditing.