Explicit Structured Reasoning

• Explicit Structured Reasoning (ESR) is a framework that makes intermediate reasoning steps explicit using structured formats such as graphs, schemas, and tables.
• It leverages staged optimization, curriculum learning, and explicit verification to boost accuracy, interpretability, and overall reasoning quality.
• ESR is applied across diverse domains including audio-language, visual reasoning, legal analysis, and text-to-SQL, demonstrating its versatile impact.

to=arxiv_search 福利彩票天天彩json {"query":"\"Explicit Structured Reasoning\" OR structured reasoning chain-of-thought scene graph reinforcement learning audio language 2025 2026", "max_results": 10, "sort_by": "relevance"} สามสิบเอ็ดassistant to=arxiv_search commentary 开号链接ាយ to=arxiv_search 一本道高清无码ություն code 亚历山大发ion 彩票平台招商? Across recent work, Explicit Structured Reasoning (ESR) denotes a family of approaches in which intermediate reasoning states are made explicit and organized by a structured substrate rather than left latent in hidden activations or an unconstrained chain-of-thought. In the literature surveyed here, that substrate can be a fixed output schema, an entailment tree, a scene graph, a box-embedded logical query, a syllogistic major-premise/minor-premise/conclusion decomposition, a structured retrieval artifact such as a table or knowledge graph, or a structured in-context environment built from graph data (Wen et al., 22 Apr 2025, Shi et al., 2018, Wang et al., 2023, Zhang et al., 5 Apr 2025, Wu et al., 16 Oct 2025, Yu et al., 27 Sep 2025). The common design principle is that reasoning is treated as a first-class external object: it can be prompted, trained, rewarded, inspected, and in many cases partially verified. ESR therefore occupies a distinct position between ordinary end-to-end latent reasoning and fully hand-coded symbolic systems.

1. Conceptual axes and boundaries

A recurring distinction in the ESR literature is between implicit and explicit reasoning, and between structured and unstructured explicit reasoning. In SARI, implicit reasoning means the model outputs only the final answer, whereas explicit reasoning requires an overt reasoning trace in <THINK> ... </THINK> before the answer; within explicit reasoning, the structured variant further constrains the trace into four mandated sections—<[PLANNING](https://www.emergentmind.com/topics/dynamic-heuristic-biasing-planning)>, <CAPTION>, <REASONING>, and <SUMMARY>—while the unstructured variant leaves the rationale as a single free-form block (Wen et al., 22 Apr 2025). SCR makes an analogous move in text reasoning by separating <answer>, <critic>, and <revised> components, with critique ending in T or F, so that verification and revision are no longer blended into a single stream (Han et al., 12 Jan 2026).

A second axis concerns what counts as “structure.” In some papers, structure is a typed natural-language schema; in others it is a graph, a logical query, or a world-state representation. XNMs make reasoning explicit through node attentions and edge attentions over scene graphs, so intermediate computation is represented directly as graph attentions rather than post hoc explanations (Shi et al., 2018). The text-pretraining framework of "Unifying Structure Reasoning and LLM Pre-training for Complex Reasoning" treats implicit text structure as triplets, paths, and intersection motifs, then executes explicit projection and intersection operators over box embeddings (Wang et al., 2023). SEER pushes the idea further by treating reasoning itself as a structured object—an entailment tree or reasoning graph—and defining returns over structural parent dependencies rather than over the temporal next step (Chen et al., 2024).

A third axis concerns the strength of explicitness. Some methods expose semantically typed states; others expose only controllable proxies. Ctrl-R is explicit at the level of lexical trajectory classes, DFA-encoded constraints, and a tractable behavior policy, but its “structured reasoning” remains a lexical proxy for latent cognitive behavior rather than a formal proof state or symbolic decomposition (Kung et al., 2 Mar 2026). UpstreamQA likewise makes reasoning explicit through upstream object-identification and scene-context traces, but these are prompt-structured natural-language modules rather than machine-verifiable symbolic representations (Nguyen et al., 25 Apr 2026). This suggests that ESR is best understood as a graded design space rather than a single formalism.

2. Representational substrates

The most important difference across ESR systems is the explicit substrate on which reasoning is carried out.

Substrate	Representative systems	Explicit structure
Fixed output schema	SARI, SyLeR, SCR	Planning/Caption/Reasoning/Summary; major premise/minor premise/conclusion; <answer>/<critic>/<revised>
Graph or state representation	XNMs, GSR, SaGe	scene graphs, node/edge attentions, world-state transitions, node-grounded traces
Query or retrieval structure	(Wang et al., 2023), Structure-R1, SIE, IESR	box queries, tables/knowledge graphs/catalogues, support/distractor subgraphs, semantic state

Schema-based ESR is most explicit when the schema is semantically typed. SARI’s four-part trace inserts a dedicated Caption stage between planning and inference so that audio evidence is externalized before reasoning (Wen et al., 22 Apr 2025). SyLeR defines legal deduction as a major premise synthesized from statutes and precedent cases, a minor premise distilled from case facts, and a conclusion aligned with the answer; the output is a syllogistic reasoning path rather than a bare legal response (Zhang et al., 5 Apr 2025). SCR decouples reasoning into generation, verification, and revision, then teaches dynamic stopping through explicit EOS supervision after a correct critique (Han et al., 12 Jan 2026).

Graph-based ESR uses explicit relational structure as the reasoning workspace. XNMs operate over scene graphs where objects are nodes and directed relations are edges, with AttendNode, AttendEdge, Transfer, and Logic as four meta-types; reasoning is therefore visible as attention flow over entities and relations (Shi et al., 2018). GSR explicitly models embodied world-state evolution as transitions over grounded scene graphs whose nodes contain objects, articulated parts, keypoints, and state labels such as open/close or empty/full (Hu et al., 2 Feb 2026). SaGe converts flat image-text corpora into scene graphs with hierarchical entities, bounding boxes, and depth ranges, then textualizes graph-grounded reasoning traces with <entity>, <bbox>, and <depth> tags (Yang et al., 7 Jul 2026).

Query-structured ESR instead exposes latent structure in text or retrieved evidence. The pretraining framework in (Wang et al., 2023) constructs simple triplets, two-step paths, outward intersections, and inward intersections, then executes relation projection and intersection over box embeddings. Structure-R1 lets the model emit <format: ...> blocks such as Chunk, Knowledge Graph, Table, Catalogue, and Algorithm, and can invent new formats when a fixed schema is insufficient (Wu et al., 16 Oct 2025). SIE builds reasoning environments from knowledge-graph subgraphs, partitioning them into supporting and distractor triples, so the environment itself becomes a structured in-context object (Yu et al., 27 Sep 2025). IESR for Text-to-SQL exposes a semantic state $S_q = (i, r, E, R, N, U, P)$, schema-compressed context, typed reasoning actions, and candidate SQL trajectories, making reasoning modular rather than monolithic (Liu et al., 5 Feb 2026).

3. Learning and optimization strategies

A striking regularity in ESR systems is that explicit structure is rarely left to emerge from reward-only learning. Instead, most methods combine structured supervision with RL, staged optimization, or structure-aware verifiers.

Several papers show that supervised warm-up is critical. SARI first performs supervised fine-tuning on structured and unstructured chains-of-thought, then applies curriculum-guided GRPO; RL-only structured or unstructured prompting from scratch produced traces that were often “meaningless,” while the two-stage regimen yielded coherent reasoning and higher accuracy (Wen et al., 22 Apr 2025). SyLeR uses GPT-4o to synthesize ten syllogistic reasoning paths per legal example, fine-tunes on these traces, and only then applies PPO with a structure-aware reward (Zhang et al., 5 Apr 2025). SaGe similarly performs graph-aligned SFT on 120K scene-graph-derived samples before RL (Yang et al., 7 Jul 2026). SCR reports that imitation of structured traces alone is not enough, but the explicit decomposition into generation, critique, and revision creates components that can then be optimized with stage-specific rewards (Han et al., 12 Jan 2026).

The reward design in ESR systems is typically component-specific rather than answer-only. SEER defines a fine-grained step reward of $1$ for perfectly matched reasoning steps, $-0.5$ for redundant steps, and $-1$ for erroneous ones, then replaces chained TD targets with a structure-based return over parent nodes in the predicted entailment tree or reasoning graph (Chen et al., 2024). Structure-R1 rewards both direct answer correctness and re-inference correctness from extracted structures alone, with $R(t_q)=R_{\text{direct}}+\lambda R_{\text{reinf}}$, so a structure is rewarded when it is sufficiently self-contained to support answering without the original documents (Wu et al., 16 Oct 2025). SaGe adds node-grounded and node-relevance rewards on top of answer accuracy and formatting, explicitly rewarding visits to visually grounded and semantically relevant graph nodes (Yang et al., 7 Jul 2026). SyLeR’s reward decomposes over the major premise, the minor premise, and the conclusion, and sets reward to zero if the response is not syllogistic (Zhang et al., 5 Apr 2025).

A second common pattern is progressive or staged RL. SCR splits optimization into Stage I, which targets initial generation and self-verification, and Stage II, which optimizes revision; this avoids reward interference between abilities that otherwise cohabit a single chain (Han et al., 12 Jan 2026). Ctrl-R uses an explicit behavior policy $\mu_\alpha$ to force rollout trajectories toward predefined reasoning motifs, then applies exact importance weighting $w$ and power-scaling $w^\beta$ so exploration and learning can be balanced rather than collapsed into ordinary on-policy sampling (Kung et al., 2 Mar 2026). SARI sorts RL data from easy to hard using baseline pass rates, reporting that easy-to-hard curricula stabilize training, accelerate convergence, and improve final performance (Wen et al., 22 Apr 2025).

A third pattern is verification through explicit structure rather than pure outcome scoring. Structure-R1 checks whether extracted structures are self-contained by re-running inference on structure-only context (Wu et al., 16 Oct 2025). SEER aligns predicted trees or graphs with gold structures before assigning rewards (Chen et al., 2024). IESR performs trajectory consistency verification by masking trajectory suffixes and asking a second model to complete them, then combines execution correctness, discriminator confidence, and peer agreement in a final trajectory score (Liu et al., 5 Feb 2026). These mechanisms are not equivalent, but they all treat the reasoning object itself as an evaluable artifact.

4. Modalities and domain-specific instantiations

ESR has now been instantiated across audio-language reasoning, vision, retrieval-intensive QA, legal analysis, embodied manipulation, text-to-SQL, and video question answering.

In audio-language reasoning, SARI is the clearest explicit formulation: the model must emit Planning, Caption, Reasoning, and Summary before the answer, and the dedicated Caption stage grounds inference in salient sounds identified from the input (Wen et al., 22 Apr 2025). In visual reasoning, the design space spans symbolic graph execution and test-time strategy scaffolds. XNMs reason over scene graphs through attention transfer and logic operations (Shi et al., 2018). Ahab converts image content and DBpedia knowledge into RDF triples, parses questions into SPARQL templates, and answers by explicit graph traversal, category closure, and typed predicate lookup (Wang et al., 2015). The PSL-based VQA system in (Aditya et al., 2018) layers weighted logical rules over semantic parses of questions and dense-caption-derived image triplets, with background knowledge from word2vec and ConceptNet. AStar stores MCTS-derived “thought cards” such as VP -> [OST](https://www.emergentmind.com/topics/on-demand-steiner-tree-ost) -> OST or SA -> OST -> SR -> CoT and reuses them as external structured guidance at test time (Wu et al., 4 Feb 2025). SaGe moves from object-centric prompting to graph-aligned post-training and graph rewards (Yang et al., 7 Jul 2026).

In retrieval-intensive reasoning, Structure-R1 is a representation-learning variant of ESR: the model may emit tables, knowledge graphs, catalogues, algorithms, or novel formats between > blocks, and these structures become the object over which subsequent reasoning proceeds (Wu et al., 16 Oct 2025). SIE instead externalizes the environment itself as a structured in-context subgraph built from supporting and distractor KG triples (Yu et al., 27 Sep 2025). In text reasoning and pretraining, (Wang et al., 2023) uses text-derived graph motifs and box-query operators to align PLM representations with explicit structure reasoning during pretraining.

In law, SyLeR uses a tree-structured hierarchical retriever to ground a legal major premise in statutes and cases, then trains the model to produce a syllogistic path comprising major premise, minor premise, and conclusion (Zhang et al., 5 Apr 2025). In embodied manipulation, GSR turns the world state into an explicit scene graph and treats planning as state transition reasoning over object states and spatial relations (Hu et al., 2 Feb 2026). In Text-to-SQL, IESR decomposes reasoning into action types such as Equation Analysis, Schema Selection, Identify Columns, SQL Generation, and SQL Revision, then explores multi-path solutions with MCTS and execution-based verification (Liu et al., 5 Feb 2026). In VideoQA, UpstreamQA inserts explicit upstream modules for object identification and scene-context generation before the final answering model, thereby making some of the perceptual reasoning process visible and modular (Nguyen et al., 25 Apr 2026).

5. Empirical patterns and performance

Across the surveyed work, ESR is associated with gains in accuracy, reasoning efficiency, or interpretability, but the gains are neither uniform nor attributable to a single mechanism.

SARI reports a 16.35% average accuracy improvement over Qwen2-Audio-7B-Instruct on MMAU, and its Qwen2.5-Omni-based structured variant reaches 67.08% on MMAU test-mini (Wen et al., 22 Apr 2025). XNMs achieve 100% accuracy on CLEVR and CLEVR-CoGenT with perfect scene graphs, 97.9% on CLEVR with detected graphs and GT programs, and 67.5% on VQAv2.0 (Shi et al., 2018). SEER reports an absolute improvement of 6.9% over RL-based methods on EntailmentBank and a 4.4% average improvement on STREET, while also improving efficiency over prior structured baselines (Chen et al., 2024). SCR reports that, compared with existing reasoning paradigms, it reduces output token length by up to 50%, while improving final reasoning performance and self-verification (Han et al., 12 Jan 2026).

In multimodal reasoning, AStar achieves 53.9% accuracy on MathVerse and 32.7% on MathVision with a 7B backbone, compared with GPT-4o’s 50.2% and 30.4%, respectively (Wu et al., 4 Feb 2025). SaGe raises Qwen2.5-VL-3B on VStarBench from 75.4 to 89.0, and lifts Qwen2.5-VL-7B on CVBench-3D from 73.3 to 80.5; on VStarBench spatial, the 3B model improves from 67.1 to 93.4 (Yang et al., 7 Jul 2026). GSR-8B reports 82.40 on LIBERO-Long compared with base Qwen3-8B’s 54.00, and strong gains on RLBench categories such as KO and SA (Hu et al., 2 Feb 2026). UpstreamQA shows that explicit upstream reasoning can substantially help weaker downstream models—Gemini 2.5 Flash rises from 58.8 to 67.8 on OpenEQA with scene-context reasoning from Gemini 2.5 Pro—but can degrade a stronger baseline such as GPT-4o, which drops from 67.7 to 47.8 under one upstream configuration (Nguyen et al., 25 Apr 2026).

In domains with explicit process rewards, the evidence often favors structured over unstructured explicitness, but only under suitable training. SARI finds that structured reasoning is more robust than unstructured reasoning, especially with curriculum and on stronger backbones (Wen et al., 22 Apr 2025). Structure-R1 shows that multi-structure reasoning and self-reward both help most on multi-hop tasks such as 2Wiki and Bamboogle, while simpler retrieval tasks such as PopQA show only small improvements (Wu et al., 16 Oct 2025). Ctrl-R shows that explicit control over structured lexical motifs improves both language and vision-language mathematical reasoning when paired with partial importance correction: Qwen3-8B-Base rises from 54.89 to 56.27, Qwen3-1.7B-Base from 32.58 to 33.85, and a Qwen2.5-VL-7B-Instruct baseline from 44.57 to 47.17 (Kung et al., 2 Mar 2026). SyLeR improves both automatic metrics and human judgments, with human evaluation averages of 4.66 for SyLeR versus 4.26 for CoT-SFT on Layperson legal questions, and especially large gains on Explainability and Trustworthiness (Zhang et al., 5 Apr 2025).

These results suggest two broad empirical regularities. First, explicit structure tends to help most when the task demands multi-hop composition, relational grounding, or state tracking rather than direct answer selection. Second, explicitness by itself is not enough: SARI’s RL-only traces can be “meaningless,” UpstreamQA can add noise when the base model is already strong, and SaGe’s node-grounded reward alone can degrade performance without a complementary relevance signal (Wen et al., 22 Apr 2025, Nguyen et al., 25 Apr 2026, Yang et al., 7 Jul 2026).

6. Limitations, controversies, and open questions

ESR methods make reasoning more inspectable, but they also relocate failure modes into the quality of the explicit structure, the faithfulness of the trace, and the adequacy of the reward.

A major limitation is that explicitness does not guarantee semantic faithfulness. Ctrl-R’s structures are lexical proxies such as “wait,” “double-check,” or “try another way,” and the paper explicitly notes that such patterns do not guarantee genuine verification or backtracking (Kung et al., 2 Mar 2026). UpstreamQA’s structured outputs are readable but remain prompt-structured natural language without a machine-verifiable formal schema (Nguyen et al., 25 Apr 2026). Structure-R1 verifies structures through answer recovery, not through a symbolic checker for completeness or consistency (Wu et al., 16 Oct 2025). This leaves open the question of how to score the quality of intermediate reasoning itself rather than only its downstream utility.

A second limitation is that explicit reasoning inherits the quality of the explicit substrate. XNMs are strongest with perfect scene graphs but degrade when detection quality is poor (Shi et al., 2018). GSR depends on scene-graph extraction that can miss fine-grained geometry or subtle object states (Hu et al., 2 Feb 2026). The text-structure pretraining framework in (Wang et al., 2023) relies on heuristic structure extraction from text, and its reported relation extraction F1 is 25.4%, which bounds the quality of the induced structures. SIE is highly scalable and verifiable at the answer level, but its main RL signal remains sparse because correctness is checked by exact answer match rather than step-level structure verification (Yu et al., 27 Sep 2025).

A third limitation is cost and redundancy. SARI notes that structured thoughts are longer than unstructured ones because explicit modules add redundancy (Wen et al., 22 Apr 2025). Structure-R1 incurs a second inference pass for self-reward verification (Wu et al., 16 Oct 2025). IESR is lightweight in model size but not in inference volume: with 32 rollouts, a full LogicCat run takes roughly 2 days per model on a single NVIDIA L20 GPU (Liu et al., 5 Feb 2026). The practical ESR question is therefore not only whether structure helps, but which structures are worth their token and search cost.

The remaining open questions are mostly about schema choice, reward design, and generalization. SARI explicitly asks how much of the gain comes from explicitness itself versus the specific four-stage schema, whether other schemas would work better for different audio tasks, and how to score intermediate reasoning quality rather than only final answer and format (Wen et al., 22 Apr 2025). Structure-R1 leaves open how diverse and stable invented formats can become beyond the examples shown (Wu et al., 16 Oct 2025). SyLeR shows that a domain-specific schema can organize retrieval, targets, and rewards, but its evaluation still lacks a dedicated structural metric and formal validity checker (Zhang et al., 5 Apr 2025). More broadly, the ESR literature suggests a stable design lesson: reasoning improves when the model is not merely encouraged to “think,” but is required to reason through an explicit, typed, and at least partially verifiable workspace. What remains unresolved is how to make that workspace both faithful and economical across domains.