Skip to content

SciLENS: RL-Driven Autonomous Agents for Scientific Localized Evidence Navigation and Synthesis

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/LQgdwind/SciLENS
Area: LLM Agent
Keywords: Scientific Literature Synthesis, Offline Autonomous Agent, Structural Data Visualization, Reverse-Decomposition Rubrics, Reinforcement Learning Alignment

TL;DR

SciLENS introduces a fully localized autonomous research agent operating on a dual-tier 12-million-record scholarly store, pioneering the use of structured data-driven visualization as an actionable tool to overcome context bottlenecks and aligning the policy via citation subgraph consensus synthesis and reverse-decomposition rubric reinforcement learning.

Background & Motivation

As the volume of published scientific literature expands at an exponential rate, researchers face steep cognitive challenges when attempting to trace methodological lineages across disciplines, navigate intricate citation topologies, and synthesize macro-level development trends. In response, autonomous agents powered by large language models have emerged as a promising paradigm for automating scholarly literature reviews and complex academic inquiry. Recent developments fall into three main architectures: standard retrieval-augmented generation (RAG) pipelines for localized question answering, web-search agents that dynamically query public internet search engines, and deep-research agents designed for prolonged multi-document synthesis. However, deploying these paradigms to rigorous scientific synthesis exposes critical systemic limitations.

The primary bottleneck lies in the linear, text-centric representation model common to all prior systems: when tasked with synthesizing broad research evolutions across dozens or hundreds of papers, generating pure textual narratives inevitably causes context window saturation and induces severe hallucination due to diluted attention. Prior agents entirely lack mechanisms to autonomously compress high-dimensional citation topologies into structured visual summaries. Second, high-quality multi-hop scientific reasoning trajectories are notoriously scarce and prohibitively costly to annotate manually, leaving open-source models inadequately trained. Furthermore, standard reinforcement learning schemes rely on sparse scalar outcome rewards, failing to provide granular process-level supervision over initial problem decomposition and strict evidence grounding. Finally, web-search and deep-research agents depend heavily on commercial search APIs, which suffer from high latency, restrictive rate limits, network fragility, and ephemeral web content that impede deterministic, reproducible research at scale.

To resolve these interconnected bottlenecks, this paper proposes an autonomous scientific research agent operating entirely on a localized, deterministic infrastructure. Core idea: build a local dual-tier indexing infrastructure over 12 million academic papers, integrate data-driven chart visualization as an actionable tool directly into the agent's reasoning loop to compress topological evidence, and align the policy via multi-model consensus subgraph synthesis and reverse-decomposition rubric reinforcement learning.

Method

Overall Architecture

The SciLENS framework is organized around four tightly coupled sub-systems: at the foundational layer, a persistent, fully local dual-tier infrastructure indexes approximately 12 million academic records (coupling a MongoDB metadata store with a distributed FAISS dense vector index) to provide sub-second, deterministic query execution. Above this layer, a unified RESEARCHTOOLBOX provides twelve specialized tools grouped into semantic retrieval, topological graph traversal, validated chart visualization, and text summarization. The agent operates within a disciplined ReAct-style thought-action-observation loop governed by a universal system prompt that enforces strict formatting and dual-tier error self-healing. Finally, the policy model is aligned through a two-stage training curriculum: supervised fine-tuning (SFT) distilled from DeepSeek-V3.2 with observation masking, followed by rubric-based reinforcement learning (RL) using the IcePop algorithm with dynamic multi-dimensional process rewards.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Complex Scientific Query Input<br/>Macro Trend Evolution / Multi-Hop Citation Tracing"] --> B["Dual-Tier Localized Infrastructure<br/>MongoDB Metadata + Distributed FAISS Dense Retrieval (12M Papers)"]
    B --> C["Visualization-Augmented Research Toolbox<br/>12 Structured Tools (Retrieval / Graph Traversal / Chart Schemas / Summary)"]
    C --> D["Topological Subgraph Synthesis & Consensus Verification<br/>Random Walk Extraction of 30K Subgraphs + 20-Model Stochastic Consensus (k=4)"]
    D --> E["Two-Stage Learning & Reverse-Decomposition Rubric RL<br/>Observation-Masked SFT Distillation + IcePop with Process Planning & Grounding Rewards"]
    E --> F["Structured Multimodal Synthesis Report<br/>Integrated Text-and-Chart Output ยท Mitigating Context Exhaustion & Hallucination"]

Key Designs

1. Dual-Tier Localized Infrastructure: Deterministic Sub-Second Dense Retrieval and Graph Topology

Relying on external web search APIs introduces network latency, strict rate quotas, and non-reproducible data drift. To guarantee deterministic and high-throughput scholarly exploration, SciLENS establishes a fully local storage backend indexing approximately 12 million cleaned records from the Open Academic Graph (OAG). The metadata layer is hosted in MongoDB, which manages raw metadata records, BM25 text indices (weighting title at 5 and keywords at 10), and citation relationships loaded into physical memory to bypass repeated disk seeks. The dense retrieval layer utilizes Qwen3-Embedding-8B to construct distributed IndexIVFFlat indices across GPU shards with 4,096 centroids per shard. To link the two systems with minimal overhead, each paper is assigned a globally unique integer identifier calculated as the shard index multiplied by \(10^{10}\) plus the within-shard document offset. This enables nearest-neighbor results from FAISS to resolve into full paper metadata via logarithmic-time index lookups, completely decoupling the agent from external internet dependencies.

2. Visualization-Augmented Research Toolbox: High-Density Topological Compression Beyond Linear Text

Existing research agents generate exclusively linear text, which quickly exhausts the context window and triggers hallucination when synthesizing macro-level academic trends across dozens of publications. SciLENS introduces an algorithmic paradigm shift by empowering the agent to autonomously call visualization tools during its reasoning loop, offloading high-dimensional statistics and citation topologies into compact, structured chart schemas. The RESEARCHTOOLBOX provides 12 specialized tools across four functional categories: - Semantic & Metadata Retrieval: EmbeddingSearch (returns Top-5 results with title, venue, year, and truncated abstract), KeywordSearch with Porter stemming, PaperInfo for exact string metadata lookup, and ReferenceSearch for batched outgoing reference resolution (capped at 3,000 items); - Graph Traversal: GetKhop for local breadth-first citation expansion (node cap of 500), GetInDegree for retrieving citing literature (capped at 50 documents), and ShortestPath for discovering the shortest directed causal connection between two papers using bidirectional breadth-first search. When forward frontier \(\pi_f\) and backward frontier \(\pi_b\) meet at node \(v^*\), the tool autonomously reconstructs the topological path: $\(\text{path} = \text{reverse}(\pi_f(v^*)) \parallel \pi_b(v^*)\)$ Bounded to a maximum depth of 5 hops with a 100,000-node layer expansion limit to prevent combinatorial explosion; - Structural Visualization: LineChart, BarChart, PieChart, and ScatterPlot. The agent emits structured JSON payloads specifying labels and numerical arrays; tools execute strict dimensional and numeric type validation before returning structured diagnostic feedback on syntax failures; - Text Utility: A dedicated Summarize tool powered by a local Qwen3-30B-A3B model that condenses verbose text segments before integrating them into the reasoning trajectory, preventing long raw documents from overwhelming the working context.

3. Topological Subgraph Synthesis & Consensus Verification: Annotation-Free High-Fidelity Data Engine

Complex multi-hop academic interaction trajectories are practically non-existent in public corpora, leaving open-source agents severely undertrained. SciLENS solves this via an automated synthesis pipeline built directly on citation subgraphs. First, the network is filtered by retaining nodes with total degree \(\ge 40\) and incoming citation degree \(\ge 10\). Random walks start from a uniformly sampled seed paper, stepping through the union of forward and backward citation neighbors for 30 steps with 10 independent walks per graph, generating 30,000 connected subgraphs. Next, a root paper is selected to formulate a factoid base question, and candidate papers along the graph are incrementally introduced to compose challenging multi-hop questions. A self-consistency filter mandates that the model must fail to resolve the compound question 4 consecutive times using only atomic statements. Finally, to eliminate factual ambiguity, an ensemble of 20 frontier commercial and open models (including Claude-3.5/3.7/Opus-4/Sonnet-4.5 and GPT-4o/GPT-5/5.1/5.2/O3 series) serves as a stochastic verification pool. For each candidate QA pair, 4 verifiers (\(k=4\)) are sampled uniformly at random to evaluate factual accuracy and answer uniqueness: $\(\text{keep}(q, a) = \prod_{m \in S} \mathbf{1}[\text{corr}_m(q, a) \land \text{uniq}_m(q, a)]\)$ Only instances achieving unanimous consensus across all 4 sampled verifiers are retained. Strict train-test disjointness guarantees 0% paper-level and edge-level overlap between splits.

4. Two-Stage Learning & Reverse-Decomposition Rubric RL: Process-Level Planning and Evidence Grounding

Standard agent reinforcement learning relies on coarse, outcome-based scalar rewards, which fail to guide early problem decomposition and encourage agents to hallucinate answers using parametric memory to take shortcuts. SciLENS adopts a two-stage training pipeline: Stage 1 performs supervised fine-tuning (SFT) by distilling approximately 20,000 complete interaction trajectories generated by DeepSeek-V3.2 on the RESEARCHTOOLBOX, applying an observation mask to exclude tool return text from the loss function and forcing the model to master alternating reasoning and execution. Stage 2 optimizes the policy using the IcePop RL algorithm, which stabilizes exploration via token-level importance weight clipping and decoupled behavior policies. Instead of a scalar score, Claude-Opus-4.1 generates instance-specific rubrics based on the synthetic question's recorded editing history: $\(\mathcal{R} = \{(d_i, w_i)\}_{i=1}^N, \quad \sum_{i=1}^N w_i = 100\)$ The rubric evaluates three distinct dimensions: Answer RLVR (verifying numeric exactness and entity equivalence), Planning & Reverse-Decomposition Rubrics (checking if initial <think> steps reverse-engineer composite questions into fundamental sub-goals), and Strict Evidence Grounding Rubrics (heavily penalizing ungrounded parametric hallucinations). The final scalar reward combines format compliance and semantic rubric scores: $\(R = \lambda \cdot S_{\text{format}} + (1 - \lambda) \cdot S_{\text{rubric}}\)$ This granular process-level signal guides the policy toward deliberate planning and strict evidence anchoring.

A Worked Example

The practical reasoning process is illustrated through an epidemiological query regarding the immunogenetic susceptibility odds ratio for visceral leishmaniasis: 1. Initial Planning & Retrieval: In its opening <think> phase, the agent breaks down the query into finding the foundational study, autonomously invoking EmbeddingSearch to locate immunogenetic candidate papers; 2. Metadata Inspection & Extraction: Reviewing the retrieved candidates, the agent identifies the seminal 2007 paper by Alonso et al., calling PaperInfo to inspect the full abstract and accurately extracting the specific odds ratio figure; 3. Graph Traversal & Consensus Verification: Rather than stopping prematurely at a historical 2007 figure, the agent reasons that scientific findings may be updated by subsequent literature. It calls GetInDegree to traverse the paper's forward citation network, verifying across subsequent citations whether this odds ratio remains the accepted consensus; 4. Answer Synthesis & Diagnostic Rubric Scoring: Finding no contradictory evidence in the forward citation graph, the agent synthesizes the final validated answer. The local Qwen3-30B-A3B judge evaluates the trajectory against the dynamic rubric, awarding full marks across Answer Semantic Fidelity, Query Intent Satisfaction, and Evidence Grounding Quality.

Key Experimental Results

Main Results

All evaluations follow a unified toolbox protocol: all models (including frontier systems like GPT-5.2 and Claude-4.5-Sonnet) interact with the identical local RESEARCHTOOLBOX and 12-million-paper database without web search, ensuring that observed score differences reflect pure model policy and alignment quality. Benchmarks include the Structural Synthesis Benchmark (SSB), Scientific Fact and Reasoning (SciFR), and four established QA and citation datasets (QASA, SciFact, PubMedQA, ScholarQA-CS).

Table 1: Performance on Structural Synthesis (SSB) and Scientific Fact and Reasoning (SciFR) (Mean ยฑ SD across 5 runs, p < 0.05)

Model Category / System SSB: ACS SSB: QAS SSB: FGS SSB: Overall (\(\uparrow\)) SciFR: ASF SciFR: QIS SciFR: EGQ SciFR: Overall (\(\uparrow\))
Closed-Source Models
Claude-4-Sonnet \(0.3007 \pm .0308\) \(0.3003 \pm .0167\) \(0.3020 \pm .0139\) \(0.3009 \pm .0455\) \(0.2922 \pm .0380\) \(0.3112 \pm .0496\) \(0.5686 \pm .0378\) \(0.3640 \pm .0267\)
Claude-4.5-Sonnet \(0.4552 \pm .0270\) \(0.5651 \pm .0467\) \(0.3845 \pm .0481\) \(0.4959 \pm .0388\) \(0.2328 \pm .0128\) \(0.3399 \pm .0476\) \(0.4190 \pm .0206\) \(0.3060 \pm .0443\)
Gemini-3.0-pro \(0.5991 \pm .0580\) \(0.7734 \pm .0493\) \(0.6868 \pm .0306\) \(0.6707 \pm .0419\) \(0.4954 \pm .0366\) \(0.7340 \pm .0396\) \(0.5956 \pm .0255\) \(0.5839 \pm .0354\)
GPT-5 \(0.5857 \pm .0428\) \(0.7340 \pm .0456\) \(0.6443 \pm .0391\) \(0.6437 \pm .0499\) \(0.5640 \pm .0465\) \(0.6585 \pm .0271\) \(0.5601 \pm .0357\) \(0.5885 \pm .0439\)
GPT-5.2 \(0.5680 \pm .0452\) \(0.7124 \pm .0529\) \(0.6321 \pm .0304\) \(0.6263 \pm .0460\) \(0.6349 \pm .0319\) \(0.7717 \pm .0313\) \(0.5772 \pm .0479\) \(0.6578 \pm .0578\)
OpenAI-DeepResearch \(0.5966 \pm .0553\) \(0.7439 \pm .0299\) \(0.6617 \pm .0392\) \(0.6564 \pm .0402\) \(0.5800 \pm .0469\) \(0.7223 \pm .0478\) \(0.5696 \pm .0528\) \(0.6158 \pm .0339\)
Open-Source Agents
A-Searcher \(0.4343 \pm .0571\) \(0.6716 \pm .0394\) \(0.5369 \pm .0386\) \(0.5291 \pm .0221\) \(0.4942 \pm .0375\) \(0.7501 \pm .0358\) \(0.5097 \pm .0410\) \(0.5672 \pm .0547\)
DR-Tulu \(0.3450 \pm .0153\) \(0.5778 \pm .0400\) \(0.4743 \pm .0320\) \(0.4433 \pm .0194\) \(0.4046 \pm .0483\) \(0.7906 \pm .0453\) \(0.4100 \pm .0323\) \(0.5101 \pm .0205\)
Mirothinker-v1.5-30B \(0.3930 \pm .0316\) \(0.5672 \pm .0283\) \(0.4922 \pm .0203\) \(0.4671 \pm .0377\) \(0.4896 \pm .0448\) \(0.7745 \pm .0519\) \(0.5316 \pm .0230\) \(0.5764 \pm .0249\)
Web-Dancer \(0.4902 \pm .0236\) \(0.7726 \pm .0551\) \(0.5788 \pm .0574\) \(0.5966 \pm .0300\) \(0.5315 \pm .0342\) \(0.8094 \pm .0390\) \(0.5232 \pm .0281\) \(0.6042 \pm .0557\)
O-Researcher-72B-rl \(0.4742 \pm .0485\) \(0.7599 \pm .0355\) \(0.5501 \pm .0525\) \(0.5796 \pm .0384\) \(0.4726 \pm .0274\) \(0.8676 \pm .0466\) \(0.5055 \pm .0225\) \(0.5860 \pm .0363\)
SciLENS Pipeline
Qwen3-30B (Base+Tools) \(0.1232 \pm .0309\) \(0.1419 \pm .0130\) \(0.1401 \pm .0432\) \(0.1324 \pm .0277\) \(0.1523 \pm .0154\) \(0.1834 \pm .0237\) \(0.1605 \pm .0119\) \(0.1654 \pm .0317\)
SciLENS-30B (SFT) \(0.6336 \pm .0269\) \(0.9054 \pm .0382\) \(0.6820 \pm .0501\) \(0.7290 \pm .0398\) \(0.6301 \pm .0472\) \(0.7863 \pm .0537\) \(0.5834 \pm .0567\) \(0.6611 \pm .0434\)
SciLENS-30B (RL) [Ours] \(\mathbf{0.6811 \pm .0579}\) \(\mathbf{0.9127 \pm .0534}\) \(\mathbf{0.7138 \pm .0370}\) \(\mathbf{0.7607 \pm .0569}\) \(\mathbf{0.7395 \pm .0326}\) \(\mathbf{0.8364 \pm .0512}\) \(\mathbf{0.7144 \pm .0416}\) \(\mathbf{0.7594 \pm .0575}\)

Table 2: Standard Scientific Benchmarks (Correctness and Citation F1)

Model Category / System QASA Corr QASA Cite SciFact Corr SciFact Cite PubMedQA Corr PubMedQA Cite ScholarQA Corr ScholarQA Cite
Claude-4.5-Sonnet 46.27 29.03 86.06 40.33 73.90 43.29 67.92 25.64
Gemini-3.0-pro 31.10 33.20 69.71 67.13 48.87 72.13 45.73 49.46
GPT-5.2 44.22 26.79 80.29 33.98 65.95 48.19 71.88 32.91
OpenAI-DeepResearch 38.92 17.02 73.56 42.94 60.02 55.20 66.20 24.32
DR-Tulu 25.26 8.10 70.48 26.02 59.79 30.50 51.10 72.34
Mirothinker-v1.5-30B 42.95 33.59 80.75 58.52 69.80 62.92 66.21 39.45
O-Researcher-72B-rl 35.51 28.78 74.23 25.04 55.52 33.51 59.34 71.22
Qwen3-30B (Base+Tools) 5.90 2.44 14.42 10.21 14.95 7.71 5.44 5.68
SciLENS-30B (SFT) 46.43 50.20 85.87 79.03 73.84 72.97 66.52 75.03
SciLENS-30B (RL) [Ours] \(\mathbf{47.62}\) \(\mathbf{52.16}\) \(\mathbf{88.94}\) \(\mathbf{83.72}\) \(\mathbf{77.53}\) \(\mathbf{76.54}\) \(\mathbf{73.98}\) \(\mathbf{76.82}\)

Ablation Study

The evaluation investigates the contribution of individual reinforcement learning rubrics alongside tool necessity under the unified offline setting.

Table 3: RL Component Ablation on SciLENS (Correctness for standard benchmarks, Overall for SciFR and SSB)

Configuration QASA SciFact PubMedQA ScholarQA-CS SciFR SSB Note
SciLENS (Full RL) \(\mathbf{47.62}\) \(\mathbf{88.94}\) \(\mathbf{77.53}\) \(\mathbf{73.98}\) \(\mathbf{0.7594}\) \(\mathbf{0.7607}\) Full rubric-based reinforcement learning
w/o Answer RLVR 41.20 72.45 64.33 62.10 0.7115 0.7350 Removing objective verification drops correctness across all tasks
w/o Planning Rubrics 47.10 86.80 74.90 70.40 0.7310 0.6855 Removing early decomposition causes large degradation on SSB
w/o Grounding Rubrics 46.50 86.17 75.80 72.40 0.7321 0.7020 Removing evidence rubrics induces parametric hallucination

Table 4: Tool Necessity Ablation (Holding SciLENS-RL Policy Fixed)

Configuration QASA SciFact PubMedQA ScholarQA SciFR SSB Key Observation
Full RESEARCHTOOLBOX \(\mathbf{47.62}\) \(\mathbf{88.94}\) \(\mathbf{77.53}\) \(\mathbf{73.98}\) \(\mathbf{0.7594}\) \(\mathbf{0.7607}\) Complete suite (retrieval, graph, visualization)
w/o Retrieval & Graph 7.15 15.12 11.05 6.21 0.1144 0.1902 Forces closed-book regime; triggers catastrophic collapse
w/o Visualization Tools 46.80 88.50 73.10 71.65 0.7312 0.5215 Restricts to linear text; SSB drops sharply from 0.7607 to 0.5215

Key Findings

  • Tool access alone does not yield agentic mastery: Providing the unaligned Qwen3-30B-A3B base model with the full tool suite achieves only 0.1324 on SSB and 0.1654 on SciFR due to repeated format parsing failures and infinite looping, demonstrating that multi-turn tool orchestration requires dedicated behavioral SFT and rubric RL.
  • Visualization as an essential structural compression mechanism: Removing visualization tools causes negligible change on atomic QA benchmarks (SciFact drops slightly from 88.94 to 88.50), but leads to a steep decline on SSB (0.7607 to 0.5215), directly confirming that data-driven charts are essential for overcoming text-centric bottlenecks in macro synthesis.
  • Context management through aggressive observation pruning: Comparing unmanaged context (which degrades rapidly due to attention dilution), preserving 20 observations, preserving 5 observations (ours), and discard-all rollout restart, keeping the 5 most recent tool observations maintains strict pass@1 execution while beating the 20-observation baseline by 2-3%, filtering stale distractors without losing the overarching reasoning plan.
  • Task-adaptive tool scheduling: On atomic factual verification (SciFR), retrieval operations dominate at 86% of all calls (31% KeywordSearch, 28% EmbeddingSearch). On macro literature review (SSB), graph traversal surges nearly fourfold to 19% (GetKhop accounting for 9%) and visualization doubles to 10%, showing that the policy model dynamically adapts its operational depth to the cognitive demands of the task.

Highlights & Insights

  • Autonomous visualization as an actionable reasoning tool: Overcomes the linear text generation bottleneck by enabling the agent to output validated chart schemas that compress multi-document quantitative distributions, preventing context exhaustion during long-horizon synthesis.
  • Self-contained data synthesis with 20-model consensus: Solves the severe lack of multi-hop scholarly interaction trajectories without human annotation by sampling multi-hop subgraphs from an empirical citation network and enforcing strict unanimity across 4 randomly sampled frontier LLMs.
  • Reverse-decomposition rubrics against parametric shortcuts: Leverages the question composition history to supervise early planning steps and penalize parametric hallucinations during reinforcement learning, fostering rigorous evidence attribution.

Limitations & Future Work

  • Reliance on a static citation corpus: The local dual-tier index operates on a snapshot of 12 million Open Academic Graph records and cannot absorb newly released preprints in real time; bridging offline high-throughput indexing with incremental streaming updates remains an open challenge.
  • Limited non-textual parsing of full papers: The system currently indexes titles, abstracts, and citation topologies, leaving deep multi-modal content such as complex mathematical derivations, detailed experimental schematics, and raw code repositories partially unmined.
  • Long-tail latency in RL rollout synchronization: During distributed reinforcement learning on 64 H800 GPUs, the high variance in reasoning trajectory lengths across complex instances caused synchronization bottlenecks across worker nodes, limiting training throughput.
  • vs Standard Scientific RAG (OpenScholar / SciRAG): Standard RAG pipelines rely on single-turn flat retrieval that isolates localized QA from macro-level trend analysis and cannot traverse citation networks; SciLENS integrates graph traversal and structural visualization to synthesize macro evolutions.
  • vs Online Web-Search Agents (Web-Dancer / Web-Shaper): Web agents query live external search engines and face severe latency, API rate quotas, and non-reproducible web drift; SciLENS operates on a fully localized 12M-paper infrastructure, guaranteeing deterministic and privacy-preserving execution.
  • vs Scalar-Reward Research Agents (DR-Tulu / O-Researcher): Prior deep research models use scalar outcome rewards that encourage shortcut-taking via parametric memory; SciLENS introduces reverse-decomposition rubrics to enforce explicit problem breakdown and sentence-level evidence grounding.

Rating

  • Novelty: โญโญโญโญโญ Pioneering data visualization as an actionable tool within the reasoning loop, accompanied by 20-model consensus subgraph synthesis and reverse-decomposition rubric RL.
  • Experimental Thoroughness: โญโญโญโญโญ Comprehensive evaluation across 6 scientific benchmarks under a unified offline protocol, featuring detailed tool necessity and RL component ablations alongside tool invocation profiling.
  • Writing Quality: โญโญโญโญโญ Rigorous technical descriptions of the dual-tier storage architecture, formal mathematical objectives for IcePop and rubric scoring, and thorough qualitative trace analyses.
  • Value: โญโญโญโญโญ Demonstrates that a 30B open-weights model aligned with process-level rubrics and localized tools can rival or exceed frontier proprietary models (e.g., GPT-5.2) on complex scientific synthesis.