Magellan: Guided MCTS for Latent Space Exploration and Novelty Generation

AI summary

Overview

Research area: Inference-time search and guidance for large language models, applied to automated scientific idea generation (AI for Science). Technical level: Intermediate (assumes familiarity with LLM embeddings, Monte Carlo Tree Search, and LLM-as-a-Judge evaluation). Scope: The paper proposes Magellan, a Monte Carlo Tree Search framework with a hierarchical guidance system that steers an LLM's generation through its latent conceptual space to produce more novel and plausible scientific research proposals.

What This Paper Is About

Large language models tend to generate "safe," high-probability text and settle into what the author calls the "gravity wells" of their training distribution, which limits truly innovative idea generation. Existing search-based methods such as Tree of Thoughts (ToT) try to explore alternatives, but they rely on the LLM's own unprincipled and inconsistent self-evaluation to decide which branches to pursue. Magellan reframes creative generation as a guided exploration of the model's latent space, using MCTS driven by an explicit, multi-objective reward structure and a long-range directional vector.

Key Contributions

1. Proposes Magellan, a framework that applies Monte Carlo Tree Search at inference time, reframing LLM generation from passive sequence prediction into an active, guided exploration within the model's latent space.
2. Designs a hierarchical guidance mechanism combining a long-range "semantic compass" vector (v_target) for global goal-setting with a principled, multi-objective value function that makes the search explicitly aware of the latent landscape's features (probabilistic gradients, knowledge density, local topology).
3. Demonstrates through experiments that Magellan significantly outperforms strong baselines, including Chain-of-Thought (CoT) and Tree of Thoughts (ToT), in generating outputs with superior novelty and overall quality.
4. Provides ablations isolating the strategic guidance component, the tactical value function and pruning, and the effect of backbone model scale, plus an efficiency and cost analysis.

Main Findings

• Magellan leads the main comparison: Reported overall score of 8.94 with a 92.0% win rate, against CoT (overall 8.63, 8.0% win rate), Zero-shot (overall 7.87, 0.0%), ToT (overall 4.93, 0.0%), and ReAct (overall 4.56, 0.0%). Magellan scored 8.98 ± 0.32 on Plausibility, 9.30 ± 0.54 on Clarity, and 8.54 ± 0.71 on Innovation.
• CoT beats Magellan on one sub-metric: CoT scored higher on Clarity (9.48 ± 0.54 versus Magellan's 9.30 ± 0.54), which the paper attributes to CoT's inherently linear, step-by-step nature producing highly readable but less sophisticated outputs.
• Advanced agentic frameworks failed: ReAct and ToT scored far below both Magellan and CoT. The paper reports ReAct suffered thematic drift ("implausible due to irrelevant datasets deviating from the theme") and ToT produced shallow, repetitive explorations ("minimal content… no innovation, merely restates the theme"). The paper's stated insight is that for open-ended creative tasks, unconstrained agency is less effective than principled, guided search.
• Magellan also beats specialized AI-for-Science systems: On idea expansion, Magellan reached a 90.0% win rate versus AI Scientist (0.0%) and SciPip (10.0%). Magellan scored 8.84 ± 0.42 Plausibility, 9.34 ± 0.77 Clarity, 8.50 ± 0.58 Innovation; AI Scientist scored 6.68 ± 1.27, 6.14 ± 1.64, 6.96 ± 1.35; SciPip scored 6.16 ± 2.58, 5.66 ± 3.26, 5.40 ± 3.16.
• Higher cost but a different token profile: On a 5-theme subset, Magellan took 5548.5 seconds and used 107,045 input and 102,400 output tokens (209,445 total); ToT took 3563.5 seconds with 118,731 input and 68,443 output tokens (187,174 total); ReAct took 1024.3 seconds with 6,940 input and 7,538 output tokens (14,478 total). ToT's input-to-output ratio was 1.73:1, Magellan's 1.05:1.
• The guidance term is critical: Disabling the guidance term in the UCT formula (w_g = 0) dropped the win rate from 90.0% to 10.0%, with Plausibility 8.02 ± 0.65, Clarity 7.90 ± 0.79, and Innovation 7.60 ± 0.81 versus the full model's 8.86 ± 0.40, 9.10 ± 0.46, and 8.32 ± 0.59.
• Removing novelty is devastating: The w/o Novelty variant scored 6.98 ± 0.96 Plausibility, 6.76 ± 0.96 Clarity, 6.40 ± 1.23 Innovation, with a 2.0% win rate (the full model in that table is listed at 98.0%). The judge's feedback was "relies heavily on existing techniques with lower novelty."
• Removing progress breaks efficiency and coherence: With w_prog = 0, the search failed to converge, always running the maximum 30 iterations, and outputs were observed to be highly repetitive and logically disjointed, making them unsuitable for quantitative scoring.
• Scale matters: Qwen-4B achieved 88.0% win rate (9.10 ± 0.42 Plausibility, 9.24 ± 0.59 Clarity, 8.98 ± 0.51 Innovation), Qwen-1.7B achieved 12.0% (8.46 ± 0.61, 8.74 ± 0.56, 7.98 ± 0.65), and Qwen-0.6B achieved 0.0% (5.92 ± 0.92, 3.92 ± 0.94, 4.70 ± 1.05). The paper describes Magellan as a "performance amplifier."
• Convergence was fast: Although MCTS was configured for a maximum of 30 iterations, the efficiency mechanisms led to an average of only approximately 3 iterations before convergence.

Methodology in Plain English

The pipeline has three stages.

1. Build a map of existing knowledge and pick a direction. The researchers encode 16,582 paper abstracts (7,937 CVPR papers from 2023–2025, 7,695 ICML papers from 2023–2025, and 950 Nature Medicine papers from 2022–2025) into dense embeddings using Qwen3-1.7B and store them in a FAISS-indexed vector database, called the novelty database. They split the embedding space into 20 clusters with K-Means, then sample two clusters that sit at a medium semantic distance from each other and prompt an LLM to synthesize a bridging research theme from representative concepts of both. The theme becomes the root node of the search tree. The theme is then decomposed into a core "problem" and "mechanism"; the mechanism vector is orthogonalized against the problem vector using Equation 1 to isolate its novel component, and the final guidance vector is v_target = v_p + α·v_m′ (Equation 2). All vectors are L2-normalized.

2. Search with guided MCTS. Each node is a text segment; a root-to-leaf path is a complete line of reasoning. Every new node is scored with a value function V = w_coh·V_coh + w_nov·V_nov + w_prog·V_prog. Coherence is the average log-probability the LLM assigns to the token sequence. Novelty is one minus the maximum cosine similarity between the node's vector and any entry in the novelty database. Progress is one minus the cosine similarity between the node's vector and its parent's vector. Node selection uses a UCT formula augmented with a third term: the cosine similarity between the node's narrative vector and the guidance vector, scaled by a guidance weight w_g. At a leaf, the LLM generates K distinct continuations; any continuation whose progress value falls below a threshold θ_prog is pruned. Scores are backpropagated up the path. Search stops at a maximum iteration count or when the best path (highest visit count) remains unchanged for a set number of iterations (Patience).

3. Extract the result. The final output is the concatenation of the text blocks along the best path, giving a layered narrative rather than a single generated passage.

Experimental setup. All methods used Qwen3-1.7B as the backbone for a fair comparison. The evaluation used a pool of 200 diverse cross-disciplinary themes, of which 50 were randomly sampled as the core test set. Baselines were Zero-shot, Chain-of-Thought, ReAct, and Tree of Thoughts, plus AI Scientist and SciPip for the idea-expansion comparison. Evaluation used an LLM-as-a-Judge protocol with DeepSeek-V3.1-Think scoring 1–10 on Plausibility, Clarity, and Innovation, plus a forced-choice "Overall Best" selection used to compute Win Rate. Hyperparameters: maximum 30 MCTS iterations, expansion width K = 3, Patience = 2, θ_prog = 0.05, UCT exploration constant C = 1.5, guidance weight w_g = 1.0, value weights w_coh = 0.5, w_nov = 0.3, w_prog = 0.2, guidance vector weight α = 1.0, sampling temperature 0.7, top-p 0.9. Baselines were given strong settings: ToT explored 5 candidates at each of up to 5 steps; ReAct was allowed up to 10 steps. Experiments ran on a four-node Tesla V100 GPU server with a total experiment time of approximately 36 hours. The paper notes that the AI agent used in the research was configured in the gemini-cli interactive environment with Gemini 2.5 Pro as its reasoning engine, and states in a footnote that the work was primarily contributed by Gemini AI, with details in the appendix.

Why This Matters

The paper argues that for creative discovery, a structured search with both a strategic compass and a tactical, principled value function beats unconstrained agentic behavior. This matters because it offers an alternative to relying on an LLM's self-evaluation, which the paper identifies as the weak point of prior search-based methods such as ToT.

Real-world applications (as implied by the framework and its evaluation):

• Assisting researchers in drafting cross-disciplinary research proposals and research themes.
• Idea expansion, turning an existing seed idea into a more detailed and structured research direction.
• Automated literature-grounded ideation, since the novelty signal is computed against a corpus of peer-reviewed abstracts.
• Serving as a component in scientific discovery pipelines that need explicit, inspectable scoring of candidate ideas rather than opaque self-assessment.

Industry relevance: The efficiency and token-cost analysis informs deployment decisions—Magellan uses a 1.05:1 input-to-output token ratio versus ToT's 1.73:1, meaning more of its computation goes into the output rather than internal search overhead. The scale ablation indicates that the approach is only useful above a certain backbone size in the Qwen family tested here, which matters for organizations choosing which model to serve.

Future Directions

• Formal normalization of the value components: The paper notes in a footnote that V_nov and V_prog have a theoretical range of [0, 2], which was empirically sufficient because MCTS is a rank-based search where monotonicity rather than scale matters; formal normalization is left for future work.
• Generalization beyond Qwen: The authors state that experiments are confined to the Qwen model family and that future work should verify generalizability across other architectures (e.g., Llama, Mistral).
• Human expert evaluation: The paper acknowledges that LLM-as-a-Judge, while consistent, is not a substitute for human expert assessment; the paper does not report any human evaluation results.
• Reducing computational cost: The authors list Magellan's higher computational cost relative to simpler methods like CoT as a limitation, since MCTS inherently involves more generation and evaluation steps.

Target Audience

Researchers and practitioners working on inference-time search, LLM reasoning frameworks, and AI-for-Science systems will benefit most, particularly those interested in replacing LLM self-evaluation with explicit reward structures. It is also relevant to engineers building automated ideation or research-assistant pipelines who need to weigh quality against token and latency costs. The paper requires intermediate background: readers should understand embeddings, cosine similarity, Monte Carlo Tree Search, and the UCT formula. Those unfamiliar with search algorithms can still follow the higher-level framing and the experimental comparisons.

Authors’ abstract

Large Language Models (LLMs) often struggle with generating truly innovative ideas, typically defaulting to high-probability, familiar concepts within their training data's "gravity wells." While advanced search-based methods like Tree of Thoughts (ToT) attempt to mitigate this, they are fundamentally limited by their reliance on unprincipled, inconsistent self-evaluation heuristics to guide exploration. To address this gap, we introduce \textbf{Magellan}, a novel framework that reframes creative generation as a principled, guided exploration of an LLM's latent conceptual space. At its core, Magellan employs Monte Carlo Tree Search (MCTS) governed by a hierarchical guidance system. For long-range direction, a "semantic compass" vector, formulated via orthogonal projection, steers the search towards relevant novelty. For local, step-by-step decisions, a landscape-aware value function replaces flawed self-evaluation with an explicit reward structure that balances intrinsic coherence, extrinsic novelty, and narrative progress. Extensive experiments demonstrate that Magellan significantly outperforms strong baselines, including ReAct and ToT, in generating scientific ideas with superior plausibility and innovation. Our work shows that for creative discovery, a principled, guided search is more effective than unconstrained agency, paving the way for LLMs to become more capable partners in innovation.