LongCoT: Extended Reasoning in LLMs

Updated 3 July 2026

LongCoT is a method for producing long, compositional reasoning paths in LLMs, enabling detailed, multi-step decision processes for complex tasks.
It benchmarks long-horizon reasoning via structured DAGs and rigorous evaluation, highlighting challenges like context management and reward sparsity.
Research advances include novel RL training methods, bootstrapping techniques like BOLT, and inference strategies such as SmartSwitch that enhance reasoning depth and efficiency.

Long Chain-of-Thought (LongCoT) constitutes a paradigm within LLM research for producing long, compositional, and interpretable multi-step reasoning traces, capable of spanning tens to hundreds of thousands of tokens. LongCoT has emerged as a central capability for solving complex, stateful tasks across scientific, mathematical, code, logic, vision–language, and autonomous workflow domains, providing a quantitative and qualitative basis for frontier reasoning evaluation, algorithmic improvements, and scalable RL training (Motwani et al., 15 Apr 2026).

1. Definition and Conceptual Foundations

LongCoT generalizes standard Chain-of-Thought (CoT) prompting by requiring models to generate especially long, structured “reasoning paths,” supporting planning, backtracking, reflection, error detection, and credit assignment over interdependent reasoning steps. Unlike classical “short-CoT”—single or few-step, human-mimetic rationales—LongCoT encompasses entire solution DAGs or procedural search traces, with chain lengths from several thousand up to hundreds of thousands of tokens (Motwani et al., 15 Apr 2026, Pang et al., 6 Feb 2025).

Formally, a LongCoT response to prompt $x$ is a sequence $\mathbf{z}_{1:L}$ , with $L$ variable and typically large, optionally terminating with final answer $y$ . For compositional benchmarks, tasks instantiate a dependency graph $T=(V,E)$ or, in procedural settings, a state-space/search-tree $T=(S,s_0,R,Q)$ . Correctness requires accurate, coherent reasoning along all relevant subpaths, not just final answer accuracy (Motwani et al., 15 Apr 2026). In vision-language settings, LongCoT trajectories are sequence $\mathbf{z}_{1:L}$ detailing both spatial grounding and logic over multimodal context (Shen et al., 26 Jun 2025).

2. LongCoT Benchmarking: The LongCoT Benchmark

The "LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning" benchmark presented the first large-scale, multi-domain testbed for isolating long-horizon reasoning (Motwani et al., 15 Apr 2026). Comprising 2,500 expert-designed problems (500 each from chemistry, mathematics, computer science, chess, logic), it strictly measures the ability to sustain long, interdependent reasoning chains.

Key features:

Each instance contains a short input (median ~2K tokens, max 6.7K), but the solution graph embedded in the task requires navigating tens to hundreds of thousands of reasoning tokens.
Problems enforce strict compositionality via opaque DAGs or procedural search, with correctness assessed via programmable/automated verifiers; every step’s correctness is required.
Models are evaluated using accuracy:

$\mathrm{Accuracy} = \frac{N_\mathrm{correct}}{N_\mathrm{total}} \times 100\%$

with all-or-nothing scoring.

At launch, even frontier closed-source models—GPT 5.2, Gemini 3 Pro, Claude 4.5 Sonnet—scored $<10\%$ on the main set (e.g., GPT 5.2: $9.8\%$ , Gemini 3 Pro: $\mathbf{z}_{1:L}$ 0), with performance relatively uniform across domains.

LongCoT reveals that failures stem from long-horizon planning, branching, self-verification, context management, and credit assignment, rather than domain gaps. Steep degradation past 15 graph nodes exposes new failures beyond pure context-length limitations. Tool scaffolding (e.g., recursive code execution) helps on logic/chess but does not substitute for fundamental LongCoT proficiency (Motwani et al., 15 Apr 2026).

3. LongCoT Training Methodologies

3.1 Reinforcement Learning with Verifiable Rewards and Policy Optimization

LongCoT RL is characterized by policy-gradient optimization targeting the production of long, verifiably correct reasoning traces. Rather than relying exclusively on supervised distillation of short or medium-step CoT traces, models are optimized on the following core RL loops:

Seed-Coder employs Gradient-Regularized Policy Optimization (GRPO, a PPO variant), where each output token is treated as an action, and final and format rewards inform the overall trajectory return. Explicitly, given trajectory $\mathbf{z}_{1:L}$ 1:

$\mathbf{z}_{1:L}$ 2

with no KL regularization and a format reward added to enforce structural output constraints (Seed et al., 4 Jun 2025).

AMPO (Adaptive Multi-Guidance Policy Optimization) interleaves on-policy RL with on-demand, comprehension-filtered off-policy guidance drawn from multiple teacher models. Unlike single-teacher distillation or pure on-policy exploration, AMPO adaptively injects diverse correct traces only when the policy fails, balancing self-discovery with coverage. Policy gradient is computed over augmented batches with careful importance correction and advantage normalization (Yuan et al., 2 Oct 2025).
RL Environments and State Management: The Markovian Thinker (Aghajohari et al., 8 Oct 2025) introduces a Markovian "Delethink" environment that conditions the next chunk of reasoning only on a small carry-over state, breaking the quadratic scaling of standard LongCoT RL and enabling efficient reasoning over arbitrarily long horizons.

3.2 Bootstrapping and Supervised Pretraining

BOLT: Bootstrapping LongCoT from an off-the-shelf instruct model is possible using as few as 10 in-context prompt–trace examples. Synthetic LongCoT data is generated with in-context learning and reward filtering, followed by SFT and online preference optimization. This permits scaling LongCoT capability to new model architectures and domains without reliance on privileged teacher models or expensive human supervision (Pang et al., 6 Feb 2025).

3.3 Vision-LLMs and Tree Search

SpatialReasoner-R1 uses Multi-Model Monte Carlo Tree Search (M3CTS) to grow and curate diverse, high-quality LongCoT trajectories, followed by fine-grained Direct Preference Optimization (fDPO) with segment-level rewards to separate descriptive grounding from logical reasoning (Shen et al., 26 Jun 2025).

4. Inference and Reasoning Strategies

LongCoT research has yielded a class of inference-time frameworks to mitigate underthinking (shallow, premature switching of reasoning chains) and token inefficiency:

SmartSwitch: A plug-and-play, training-free framework for LLMs to detect premature thought abandonment using a reward model and intervene by backtracking and prompting for deeper exploration. It monitors for thought switch linguistic cues, uses a process reward model to evaluate abandoned thoughts, and triggers up to three interventions inserting deepening prompts (Zhang et al., 22 Oct 2025).

SmartSwitch achieves consistent gains in pass@1 accuracy (5–25 points depending on model/benchmark), reduces average response length by up to 15%, and inference time by up to 35%, e.g.:

Model	Vanilla Accuracy (AIME24)	SmartSwitch Accuracy	Δ points
Qwen-1.5B	28.9	40.0	+11.1
Qwen-32B	79.5	86.7	+7.2

SmartSwitch reduces "Underthinking Frequency" $\mathbf{z}_{1:L}$ 3 by up to 40%, yielding more coherent, deeply explored reasoning paths (Zhang et al., 22 Oct 2025).

5. Empirical Results and Current Limitations

Across multiple domains, LongCoT-trained models and inference strategies produce substantial improvements on intricate reasoning tasks, especially with RL leveraging verifiable intermediate and final rewards. For example, Seed-Coder LongCoT RL boosts pass@1 on LiveCodeBench from 39% (distilled base) to 53.6%, with more concise, robust code (Seed et al., 4 Jun 2025). BOLT transforms an 8B instruct LLM's MT-Bench score from 32.1 to 52.9, with similar dramatic improvements on WildBench, Arena-Hard, ZebraLogic, and MATH500 (Pang et al., 6 Feb 2025).

Nonetheless, current limitations include:

Long-Horizon Generalization: Even the best models achieve only single-digit percent accuracy (≤10%) on the full LongCoT benchmark (Motwani et al., 15 Apr 2026).
Token/Compute Efficiency: Standard RL environments incur quadratic compute/memory in thinking length, only mitigated by explicit Markovian paradigms (Aghajohari et al., 8 Oct 2025).
Reward Sparsity and Proxy Quality: Relying on final answer or format rewards limits credit assignment and may introduce learning plateaus or instability (Seed et al., 4 Jun 2025, Pang et al., 6 Feb 2025).
Intervention Fragility: In SmartSwitch, intervention efficacy depends on linguistic cue detection and reward model calibration; harder domains may require further PRM tuning (Zhang et al., 22 Oct 2025).

6. Future Directions and Open Challenges

Promising research avenues include:

RL and Horizon Bootstrapping: Architectures that decouple context size from reasoning length—e.g., Markovian Thinker—allow models to reason over hundreds of thousands or millions of tokens linearly, enabling new scaling regimes (Aghajohari et al., 8 Oct 2025).
Curriculum and Reward Engineering: Introduction of denser intermediate rewards, value networks, and curriculum learning to accelerate LongCoT RL and improve chain quality (Seed et al., 4 Jun 2025).
Cross-Domain and Multimodal Generalization: Expansion of LongCoT methodology to reinforce long-horizon reasoning in vision-language, spatial, and cross-modal tasks, leveraging segment-specific rewards and fine-grained preference optimization (Shen et al., 26 Jun 2025).
Systematic Bootstrapping: BOLT demonstrates scalable LongCoT bootstrapping with minimal in-context exemplars; future work could target multilingual, multimodal, or low-resource language settings (Pang et al., 6 Feb 2025).
Comprehensive Benchmarks and Tool Use Integration: Extending LongCoT benchmarks to integrate tool-use, retrieval, and real-world workflows, without conflating tool-scaffolding with fundamental reasoning (Motwani et al., 15 Apr 2026).

LongCoT reasoning thus defines both a major capability frontier for modern LLMs and a fertile discipline for advancing the state of compositional, interpretable, and robust machine reasoning.