Hybrid Reasoning Agents

Jiaru Zou; Yunzhe Qi; Sirui Chen; Federico Bianchi; Shang Zhu; Pan Lu; Jingrui He; James Zou

cs.CLOct 13, 2026

Hybrid Reasoning Agents

Authors: Jiaru Zou, Yunzhe Qi, Sirui Chen, Federico Bianchi, Shang Zhu, Pan Lu, Jingrui He, James Zou

Organizations: Stanford University · Together AI · UIUC · Meta

Abstract

How can we build an agent that thinks beyond human language, yet still acts through it in the external environment? Modern LLM agents still reason and interact through text, while latent reasoning offers an efficient internal medium but lacks native tool access and reasoning transparency. We bridge these two directions with Hybrid Reasoning Agents (HybridAgents), which dynamically navigate between latent reasoning and text-based interaction throughout a unified generation process. Architecturally, a lightweight HybridLink enables seamless transitions between text and latent modes, while the agent's continuous working memory preserves lossless hybrid context across multiple turns. These designs preserve tool-integrated agentic capabilities while making latent thoughts directly inspectable. Building on this architecture, we further develop dedicated post-training algorithms for hybrid reasoning: HybridSFT stabilizes tool-integrated hybrid generation by teaching the backbone to coordinate with HybridLink, then HybridRL scales reasoning by learning when to think in latent space and when to act through text. We evaluate HybridAgent at 3 model scales across 11 benchmarks covering mathematics, science, medicine, search, and agentic reasoning. Compared with frontier LLM Agents and agentic frameworks spanning both text-based and latent methods, HybridAgent consistently improves performance by 10.4%, while delivering 35.5%-54.7% token reduction, 1.8x-2.6x faster inference, and 2.7x-3.7x cost savings. Code and data are available at https://github.com/togethercomputer/HybridAgent.

Figures & tables

Figure 1: HybridAgent Architecture Overview. Left: A HybridLink maps last-layer hidden states back into the input embedding space, letting the agent enter latent mode for recurrent, fast, inspectable inner thinking and return to text mode for tool calls and external monitoring . Right: HybridAgent performs hybrid reasoning that thinks in latent space and acts through text within a single generation process. Building on this architecture, HybridAgent (i) achieves multi-turn hybrid generation by carrying the entire hybrid context through its own working memory (Section 3 ), and (ii) can be converted from any standard pre-trained LLM through a dedicated post-training pipeline (Section 4 ).

Figure 2: Interpretable Latent Mode in HybridAgent. All latent steps can be expressed as readable pseudo-text carrying their semantics for external inspection. Left: Quantitatively, latent thoughts are highly readable and context-specific (see Appendix F ). Right: Qualitatively, latent pseudo-text stays semantically aligned with surrounding text reasoning and tool calls. (“Steps” counts forward passes).

Figure 3: Multi-Turn Hybrid Generation over HybridAgent’s Own Working Memory. HybridAgent accumulates key-value (KV) entries from text, latent thoughts, and tool observations in continuous working memory, directly reusing and updating it across turns to preserve the full hybrid context.

Figure 4: Self-Conditioned HybridSFT. A teacher-forced first pass produces next-step latent inputs via HybridLink. A second pass replaces ground-truth latent embeddings Emb

(\cdot)

with the predictions

\mathcal{R}(h)

and jointly trains the backbone and HybridLink using CE at text and regression at latent positions.

Method	GAIA				WebWalker				HLE				SealQA		DSQA
Lv.1	Lv.2	Lv.3	Avg.	Easy	Med.	Hard	Avg.	NS	CE	SF	Avg.	S-0	Hard	Acc.	F1

Frontier LLM Agents
QwQ-32B	30.9	6.5	5.2	18.9	7.5	2.1	4.6	4.3	11.5	7.3	5.2	9.6	1.8	10.6	11.3	23.9
GLM-5.3-Flash	38.5	28.8	8.3	30.1	8.9	8.3	9.9	9.0	13.3	14.7	16.8	14.2	9.0	22.0	16.0	38.5
DeepSeekV4.1-Flash	43.6	42.3	16.7	39.8	22.2	16.7	12.7	16.5	28.5	25.3	30.5	28.4	14.4	33.5	19.3	38.6

Table 1: Main Results of HybridAgent over Agentic Reasoning Tasks. We report GAIA and WebWalker by difficulty level, HLE by category (NS/CE/SF), SealQA on the Seal-0 (“S-0") and Seal-Hard (“Hard") splits, and DSQA by full-correct rate (“Acc.") and F1 score. Across 5 independent runs, the average standard deviations of HybridAgent are

\pm

0.014 in accuracy, with a corresponding 95% CI of

\pm

0.017. At 8B and 14B scales, HybridAgent consistently outperforms advanced LLM agents, agentic frameworks, and matched text and latent baselines. Best performance is highlighted .

Method	AIME25	AIME26	MATH500	GPQA-D	MedQA	MuSiQue

Qwen3-8B-Base
Coconut	13.1	9.6	51.4	26.8	61.7	14.0
SoftCoT	16.5	17.9	63.6	28.9	62.0	19.5
Standard Text SFT	30.0	28.3	78.2	32.3	69.7	23.5
HybridAgent (SFT)	30.8	29.2	78.4	31.8	70.0	26.5

Table 2: Main Results of HybridAgent on Math and Knowledge-Intensive Search. We evaluate performance across both SFT and RL stages. Small-size benchmarks like AIME25/26 are reported on avg@16 accuracy. Across all tasks, HybridAgent yields an average std. of

\pm

0.031 (math) and

\pm

0.018 (search), with corresponding 95% CI of

\pm

0.038 and

\pm

0.022. Under the same training budgets and model setups, HybridSFT and HybridRL improve both text-based and latent reasoning methods.

Figure 5: Scalability of HybridAgent at Test-time. We evaluate pass@k ( $4,8,16,32,64$ ) performance on 100 hard math and science questions and compare with text agents trained with GRPO.

Figure 6: Efficient Hybrid Generation of HybridAgent across Model Scales. HybridAgent reduces both text-token usage and total generation steps compared with matched text-based LLM agents.

Figure 7: Inference time speedup of HybridAgent across all model scales.

Table 3 and Figure 8 : Effectiveness and Interpretability of Self-Conditioned HybridSFT. Left: We compare successive HybridSFT stages with latent and text training baselines using the same Qwen3-14B-Base backbone. Right: We evaluate the readability and context-specificity across 200 downstream questions, showing how self-conditioning improves latent transparency.

Figure 9: Training Dynamics of HybridRL. As training progresses, HybridAgent shifts more generation from text to latent mode while increasing reward and tool use. These trends reflect how HybridAgent adaptively learns to allocate its reasoning and tool calling efforts across the two modes.

Reward Design	MATH500	GAIA	WebWalker
HybridRL (8B)	84.6	43.7	42.5
w/o $r_{\text{correct}}$	79.2	35.9	37.5
w/o $r_{\text{format}}$	80.6	40.7	40.5
w/o $r_{\text{tool}}$	84.2	38.8	41.0

Table 4: Ablation on HybridRL rewards.

Latent-Mode Intervention	Setup Description	Strength $d$	8B-Scale	14B-Scale
Noise Perturbation
Gaussian Noise ( $\sigma=0.5$ )	Add perturbation with random noise of magnitude $0.5\\|v\\|$ before latent feedback.	0.5	72.6	75.0
Gaussian Noise ( $\sigma=2.0$ )	Add perturbation with random noise of magnitude $2.0\\|v\\|$ before latent feedback.	2.0	66.4	70.2
Latent Replacement
Random-token Replacement	Replace $v$ with a uniformly sampled token embedding, rescaled to $\\|v\\|$ .	1.6	41.2	44.6
Cross-problem Replacement	Replace each latent span with the corresponding latents from a different partner problem.	1.4	44.8	48.6

Table 5: Causal Interventions on HybridAgent Latent Mode. We report MATH500 accuracy (%) for 8B and 14B scales under different interventions on latent-mode thoughts generated by HybridAgent. Performance drops substantially when latent semantics or structure are disrupted, indicating that HybridAgent causally relies on semantically meaningful latent-mode thoughts. Here,

v

denotes the original latent vector and

v^{\prime}

denotes the intervened counterpart. The relative perturbation strength

d

is computed as

\|v^{\prime}-v\|/\|v\|

, averaged over latent steps, where

d=0

means no intervention.

HybridLink Design	MATH500	MedQA	MuSiQue
Identity (No Alignment)	71.4	73.7	24.5
1-Layer	83.6	79.7	28.5
1-Layer+Res.	86.4	80.3	31.0
2-Layer	85.8	81.0	31.5
2-Layer+Res. (ours)	88.2	81.7	34.5

Table 6: Effectiveness of HybridLink. We compare accuracy across alternative architectural designs.

Figure 10: Agentic Tool-Use Analysis on HybridAgent. Left : Valid tool-call rate across 8B and 14B scales; "Execute Success" measures error-free execution among valid tool calls. Middle : Final-answer accuracy on questions where the agent performs at least one tool call. Right : Useful tool-call rate, measuring the fraction of tool calls whose execution results contain the final answer.

Metric	Measurement	8B Scale		14B Scale
Metric	Measurement	HybridAgent	Text Agent	HybridAgent	Text Agent
	Text-Mode tokens per instance	302	596	284	497
	Latent-Mode steps per instance	170	–	183	–
Visible Trace	Key reasoning steps per 100 visible tokens $\uparrow$	0.72	0.57	0.86	0.67
	Weak-monitor AUROC (MATH500) $\uparrow$	0.840	0.772	0.860	0.818
Monitoring	Strong-monitor AUROC (all tasks) $\uparrow$	0.961	0.935	0.989	0.990

Table 7: Monitorability and Debuggability through HybridAgent Text-Mode. We compare HybridAgent with matched same-size text agents to examine whether its exposed text remains sufficient for external monitoring, debugging, correction, and rationale-based inspection under hybrid reasoning. All comparisons are evaluated using paired bootstrap 95% confidence intervals.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 11: Train-Inference Mismatch in Latent Reasoning. We illustrate the exposure bias that arises when teacher-forced training meets free-running inference. Left: During activation, the backbone consumes ground-truth embeddings $\mathrm{Emb}(y)$ at latent positions, while HybridLink outputs $\mathcal{R}(h)$ are only trained to match their targets and never fed back as inputs. Right: At inference, every latent input is the model’s own $\mathcal{R}(h)$ , so deviations from the target embeddings accumulate across the span.

Figure 12: Hybrid Data Curation Pipeline. We convert text-based demonstrations into multi-turn hybrid trajectories using complementary signals from LLM judgments and model entropy to assign reasoning modes, followed by a visibility audit that preserves decision-critical steps in text mode.

Method	AIME25	AIME26	MATH500	GPQA-D

Qwen3-4B-Base
Coconut	10.6	13.3	40.6	17.2
SoftCoT	14.8	18.5	55.2	26.3
Standard Text SFT	19.0	22.7	77.4	31.8
HybridAgent (SFT)	22.5	23.3	76.8	30.8

Appendix

Table 8: 4B-Scale HybridAgent on Math and Knowledge-Intensive Search. We evaluate performance across both SFT and RL stages. AIME25/26 are reported on avg@16 accuracy.

Model Backbone	Text Agent	HybridAgent	Cost / 1M questions	Cheaper by
Model Backbone	(TFLOP)	(TFLOP)	(Text Agent $\rightarrow$ HybridAgent)	Cheaper by
Qwen3-4B-Base	51.2	13.9	$359$ \rightarrow98	3.7 $\times$
Qwen3-8B-Base	62.5	17.2	$439$ \rightarrow121	3.6 $\times$
Qwen3-14B-Base	66.9	25.1	$470$ \rightarrow176	2.7 $\times$

Appendix

Table 9: Cost Analyses on HybridAgent. We compare HybridAgent with text agents under the same backbones, reporting compute per question and estimated monetary cost per 1M questions. Each text or latent generation step is charged as one full backbone forward pass. Monetary costs assume

2.50 per H100-hour and 10% utilization of peak BF16 throughput (approximately

0.007 per PFLOP).

Hyperparameter	Value
Base model	Qwen3-4B/8B/14B-Base
Maximum sequence length	16,888
Effective batch size	16
Backbone learning rate	$7\times 10^{-6}$
HybridLink learning rate (Activation / Self-conditioning)	$2\times 10^{-4}$ / $5\times 10^{-5}$
Weight decay	0.0

Appendix

Table 10: Hyperparameter Setups of HybridSFT.

Hyperparameter	Value
Initialization	8B/14B HybridSFT checkpoint
Unique prompts per rollout batch	64
Rollouts per prompt (group size)	8
Minibatches per rollout batch	8
Optimization epochs per rollout batch	1
Total rollout batch size	512 trajectories ( $64$ $\times$ $8$ )

Appendix

Table 11: Hyperparameter Setups of HybridRL on Agentic Search.

Hyperparameter	Value
Initialization	4B/8B/14B HybridSFT checkpoint
Unique prompts per rollout batch	64
Rollouts per prompt (group size)	8
Minibatches per rollout batch	8
Optimization epochs per rollout batch	1
Total rollout batch size	512 trajectories ( $64\times 8$ )

Appendix

Table 12: Hyperparameter Setups of HybridRL on Deep Reasoning.

Figure 24

Figure 25

Figure 26

Figure 27

Figure 28

Figure 29

LLM Agents

Large language model agents use language models to plan, use tools, and act across multi-step tasks and environments.

See all 705 papers

Adaptive Latent Agentic Reasoning
Jun 1, 2026Dongwon Jung, Peng Shi, Yi Zhang +2Agentic Reasoning LLM Agents
University of California, Davis · University of Waterloo · Greenshoe, Inc.
Eigen-Agent: Adaptive Multi-Agent Scientific Reasoning with Monitor-Based RAG
Sep 25, 2025Xiangru Tang, Wanghan Xu, Yujie Wang +13Scientific Reasoning in Language Models Multi-Agent Reasoning
Yale University · Shanghai Jiao Tong University · Fudan University +4
A Survey of Agentic Reasoning for Large Language Models: Towards Recursively Self-Improving and Collective Agents
Jan 18, 2026Tianxin Wei, Ting-Wei Li, Zhining Liu +26Continual Learning for LLM Agents Multi-Agent Reasoning
University of Illinois Urbana-Champaign · Meta · UCSD +2

Jun 1, 2026J/K move · Enter open · S save

Adaptive Latent Agentic Reasoning

Dongwon Jung, Peng Shi, Yi Zhang, Junshan Zhang+1

University of California, Davis · University of Waterloo · Greenshoe, Inc.

Agentic Reasoning LLM Agents LLM Agent Training Efficient Language Model Inference Latent CoT Reasoning

Open paper AskPDF arXiv

Large reasoning models improve performance by generating extended chain-of-thought (CoT) reasoning, but this behavior becomes inefficient when applied to LLM agents. Current LLM agents often generate verbose textual reasoning at every decision step and allocate reasoning effort nearly uniformly across turns, leading to substantial inefficiency in multi-turn agentic trajectories. We propose Adaptive Latent Agentic Reasoning (ALAR), a dual-mode framework that uses compact latent reasoning for routine turns and selectively escalates to explicit chain-of-thought when deeper deliberation is needed. ALAR learns latent reasoning by using the agent's actions as supervision anchors and is further optimized to use latent reasoning when it is sufficient for task success and reserve explicit CoT for harder decisions. Experiments on agentic search and tool-use benchmarks show that ALAR maintains comparable or better task accuracy while substantially reducing generated tokens by up to 43.6% in search and 84.6% in tool use. These results demonstrate that ALAR improves the accuracy-efficiency trade-off of LLM agents by reducing unnecessary textual reasoning while preserving explicit deliberation for harder decision steps.

Abstract

LLM Agents

Explore similar work

Adaptive Latent Agentic Reasoning

Eigen-Agent: Adaptive Multi-Agent Scientific Reasoning with Monitor-Based RAG

A Survey of Agentic Reasoning for Large Language Models: Towards Recursively Self-Improving and Collective Agents