Agentic Symbolic Search (ASYS): Structured Exploration

• Agentic Symbolic Search (ASYS) is defined by explicit search state, typed actions, and compact memory to structure long-horizon reasoning.
• The framework decomposes search-time cognition into modules with shared state and turn-level optimization to enhance evidence integration and retrieval.
• ASYS is applied in PDE characterization, RTL repair, legal reasoning, and biomedical intervention through hybrid neural-symbolic methods with trajectory-aware optimization.

Agentic Symbolic Search (ASYS) denotes a family of agentic search systems in which long-horizon reasoning is organized around explicit search state, typed actions, compact external memory, and controllable state transitions rather than unrestricted natural-language trajectories. In current work, ASYS is usually hybrid rather than fully symbolic: the control layer is symbolic and structured, while the semantic layer remains neural and natural-language based. The most explicit named instance treats PDE characterization as a prior-guided search over differentiable symbolic programs, while adjacent agentic-search frameworks supply the broader architectural pattern of shared state, graph-structured plans, structured protocols, and trajectory-aware optimization (Yu et al., 18 Jun 2026, Wang et al., 23 Dec 2025, Chen et al., 8 Jan 2026).

1. Conceptual emergence and problem setting

ASYS emerged against a critique of monolithic agentic search. In monolithic systems, one LLM is expected to perform planning, query formulation, retrieval usage, evidence integration, and answer generation inside one long evolving response. This produces coupled bottlenecks: unconstrained reasoning outputs, inflated trajectories, sparse outcome-level rewards, difficult credit assignment, and stochastic search noise from external tools (Chen et al., 8 Jan 2026). Closely related work describes the same failure regime as reasoning instability from unconstrained natural-language traces and context overflow from raw trace accumulation, with prior systems prone to “unstructured loops or premature conclusions” and “context drift” (Wang et al., 23 Dec 2025).

A further systems-level pressure comes from retrieval itself. Information retrieval has historically assumed an “organic query stream,” but agentic systems produce a “synthetic query stream,” changing workload volumes, query reformulation patterns, and cacheability. The resulting object of study is no longer a single query but a trajectory of reasoning-induced queries, retrieved documents, thoughts, and answers (Pezzuti et al., 19 Feb 2026). This shift is central to ASYS: the search object is a stateful process, not a one-shot retrieval event.

A common misconception is that ASYS simply means “more agents.” The literature is more specific. M-ASK is relevant precisely because it does not merely propose “more agents” for search; it proposes a structured decomposition of search-time cognition into explicit modules with a shared intermediate state and turn-level optimization (Chen et al., 8 Jan 2026). The emphasis is on explicit control flow, bounded memory growth, and stepwise scoring.

2. State, protocol, and symbolic control

The architectural center of ASYS is an explicit intermediate state. In M-ASK, the shared state is

$$\mathcal{K}_t = \left\{ \begin{aligned} & \text{``question"}: q, \ & \text{``thinking\_trajectory"}: \mathcal{T}_t, \ & \text{``predicted\_answer"}: a_t \end{aligned} \right\},$$

with the trajectory represented as a list of reasoning steps,

$\mathcal{T}_t = [\tau_1, \tau_2, \dots, \tau_m],$

where each step is a sub-query / sub-answer pair (Chen et al., 8 Jan 2026). This turns an implicit chain of thought into an explicit, editable reasoning record. The Update Agent then chooses between in-place refinement, \<Update> ... \</Update>, and expansion, \<Add> ... \</Add>, giving the system an explicit state-transition language rather than append-only transcript growth (Chen et al., 8 Jan 2026).

Laser pushes this further by defining a symbolic action protocol

$$\mathcal{A} = \mathcal{A}_\textrm{plan} \cup \mathcal{A}_\textrm{sol} \cup \mathcal{A}_\textrm{ret},$$

partitioned into planning, task-solving, and retrospection spaces (Wang et al., 23 Dec 2025). Its planning space contains <intent_refinement> and <problem_framing>, with the latter producing a DAG-structured task decomposition. Its task-solving space contains <tool_call>, <doc_extraction>, <task_answer>, and <final_answer>. Its retrospection space contains <revisit_task> and <replanning>, making backtracking explicit at both task and plan level (Wang et al., 23 Dec 2025). The corresponding context register is a typed memory,

$$\mathcal{R}^t=\{\mathcal{I},\mathcal{P}_{\textrm{his}},\mathcal{P}_{\textrm{cur}}\},$$

with refined intent, planning history, and current plan state stored separately rather than intermingled in a raw transcript (Wang et al., 23 Dec 2025).

Other ASYS-like systems adopt analogous typed interfaces. LRAS organizes legal reasoning into >, <search>, <information>, and <answer> actions, implementing a non-branching but explicit think-search-observe-rethink loop (Zhou et al., 12 Jan 2026). Clover defines a search node as a pair $(c,h)$, where $c$ is code status and $h$ is dialogue history, and treats verified RTL repair as search over program states and hypotheses rather than over unconstrained text (Luo et al., 19 Apr 2026). Across these systems, the recurrent pattern is the same: search state is externalized, action semantics are typed, and intermediate representations are designed to be inspectable and updateable.

3. Optimization, retrieval, and search regulation

ASYS research has converged on the view that long-horizon search requires trajectory-level optimization rather than answer-only supervision. M-ASK uses turn-level reward shaping under Independent PPO with parameter sharing. For iterative search it defines a shared incremental reward based on marginal answer improvement,

$$r_{\text{iter}}^{(t)} = \text{F}_1(a_t, y) - \text{F}_1(a_{t-1}, y),$$

assigned to the Search, Summary, and Update agents for the current turn (Chen et al., 8 Jan 2026). This reframes credit assignment from “which of the last 300 tokens caused success?” to “did this state transition improve the answer?”

Behavior-level supervision appears equally important. “Beneficial Reasoning Behaviors in Agentic Search and Effective Post-training to Obtain Them” identifies four beneficial reasoning behaviors—Information Verification, Authority Evaluation, Adaptive Search, and Error Recovery—and shows that trajectories exhibiting these behaviors are a stronger foundation for RL than correctness-only trajectories. On Qwen3-1.7B, Behavior [Prime](https://www.emergentmind.com/topics/partial-masking-scheme-prime) (Incorrect) + RL reaches 23.5, nearly identical to Behavior Prime (Correct) + RL at 23.6, and better than SFT (Correct) + RL at 20.0 (Jin et al., 8 Oct 2025). The paper’s mechanism claim is that these behaviors improve exploration, pass@k, entropy, and test-time scaling (Jin et al., 8 Oct 2025).

Retriever design has also become trajectory-aware. Agentic-R argues that passages in multi-turn search must be judged by both local query-passage relevance and global answer correctness, not semantic similarity alone. It constructs positives and negatives for intermediate queries using candidate passages, listwise local relevance scores, and exact-match correctness of rollouts conditioned on those passages, then iteratively co-trains the retriever and the agent (Liu et al., 17 Jan 2026). This is directly ASYS-relevant because it aligns evidence selection with trajectory success rather than one-step topicality.

Search regulation itself has become a dedicated problem. SAAS defines a policy-relative “search boundary” by contrasting search-disabled and search-enabled rollouts and then penalizes unnecessary and redundant search only after correctness has been achieved (Tang et al., 28 May 2026). Its boundary model is

$$\mathcal{S}(q)= \begin{cases} \text{NoSearch}, & n_{\mathrm{d}(q)} \ge \delta, \ \text{NeedSearch}, & n_{\mathrm{d}(q)}=0,\ n_{\mathrm{e}(q)}>0, \ \text{Undetermined}, & \text{otherwise}. \end{cases}$$

This is coupled with a stage-wise optimization strategy: first learn how to use search, then regularize when not to use it (Tang et al., 28 May 2026).

4. Domain-specific realizations

The named ASYS framework for PDEs makes the symbolic ambition explicit. It searches over differentiable symbolic programs to characterize PDE solutions from the governing equation, initial and boundary conditions, public mathematical guidance, admissibility constraints, and search history, without exposing hidden reference trajectories during search (Yu et al., 18 Jun 2026). Across five problems spanning bounded dynamics, finite-time blow-up, and free-boundary focusing, it recovers the exact Satsuma–Yajima breather core for NLS, constructs a geometric interface formula for Allen–Cahn 2D dynamics, and discovers a nine-parameter contraction law for Keller–Segel chemotactic blow-up (Yu et al., 18 Jun 2026).

In verified RTL repair, Clover instantiates ASYS as structured search over code manipulations under hard feedback from compilation, linting, simulation, and SMT solving. Its Stochastic Tree-of-Thoughts stores, scores, samples, and restores nodes $(c,h)$, and the full system fixes 31 of 32 bugs, i.e. 96.8%, with average pass@1 of 87.5% (Luo et al., 19 Apr 2026). Here ASYS is not equation discovery but a branch-and-verify search over executable program states.

In legal reasoning, LRAS recasts legal QA as sequential Active Inquiry. It trains a model to decide whether it can answer from internal knowledge, what to search for if not, whether retrieved evidence is sufficient, and when to stop. On 705 failed JEC-QA samples, search was triggered in only 28.7% of failure cases, while 71.3% of the failures involved no search trigger at all; LRAS then uses Introspective Imitation Learning and Difficulty-aware Reinforcement Learning to address precisely this introspection deficit, reporting gains of 8.2–32% over baselines (Zhou et al., 12 Jan 2026).

In biomedical intervention search, “Contextual Invertible World Models” implements a small-scale neuro-symbolic intervention-search framework. The state is a hybrid feature vector $\mathcal{T}_t = [\tau_1, \tau_2, \dots, \tau_m],$0 containing a binary mutation vector, 15 transcriptomic principal components, and MSI status; a frozen Random Forest world model predicts $\mathcal{T}_t = [\tau_1, \tau_2, \dots, \tau_m],$1, and an LLM-based agentic layer queries it through in silico perturbations such as APC or TP53 repair (Baker et al., 1 Mar 2026). The contextual model reaches $\mathcal{T}_t = [\tau_1, \tau_2, \dots, \tau_m],$2, and adding MSI yields an 18.8% relative gain in predictive fidelity (Baker et al., 1 Mar 2026). This is not formal symbolic planning, but it is counterfactual search over symbolic state edits.

5. Evaluation, datasets, and diagnostics

ASYS has also changed how agentic search is evaluated. Mind2Web 2 treats long-horizon web search as hierarchical constraint satisfaction rather than exact answer matching. Its Agent-as-a-Judge framework represents a rubric as a tree whose leaves are binary criteria and whose internal nodes aggregate critical and non-critical subcriteria via a gate-then-average rule. The benchmark contains 130 realistic tasks built with at least 1,000 hours of human labor. The best system, OpenAI Deep Research, reaches 0.54 Partial Completion, 0.28 Success Rate, and 0.40 Pass@3, while humans reach 0.79, 0.54, and 0.83 on the subset study (Gou et al., 26 Jun 2025).

RAVine extends this evaluation philosophy by targeting “multi-point queries and long-form answers that better reflect user intents” and by adding attributable nugget construction and process metrics such as cumulative relevant coverage, marginal search gain, search recall, and search precision (Xu et al., 22 Jul 2025). Its results show that even strong models achieve task completeness only in the mid-40s and have weak citation recall and precision, underscoring the gap between retrieval, synthesis, and provenance (Xu et al., 22 Jul 2025).

“A Picture of Agentic Search” contributes instrumentation rather than a scoring formalism. It defines a frame as

$\mathcal{T}_t = [\tau_1, \tau_2, \dots, \tau_m],$3

and a trace as

$\mathcal{T}_t = [\tau_1, \tau_2, \dots, \tau_m],$4

then releases ASQ, a dataset of reasoning-induced queries, retrieved documents, and thoughts collected from agentic retrieval-augmented systems (Pezzuti et al., 19 Feb 2026). It also introduces a reformulation-state taxonomy—IN, ADD, REM, REP, DUP, CH, OUT—that makes looping, repetition, and trace instability observable (Pezzuti et al., 19 Feb 2026).

Vertical-domain benchmarks reveal the same issues under more operational constraints. LocalSearchBench provides 150,031 anonymized merchant entries and 300 multi-hop QA tasks for local life services. Even with LocalPlayground, the best model, DeepSeek-V3.1, achieves only 34.34% correctness, while average completeness is 77.33% and average faithfulness is 61.99% (He et al., 8 Dec 2025). This gap between seeming completeness and actual correctness is precisely the kind of failure mode that motivates stronger symbolic control.

6. Limitations, misconceptions, and research directions

Current ASYS-like systems remain predominantly hybrid. Laser is explicit that the content of refined intent, extracted facts, subtask answers, and final answers remains natural-language or JSON-like neural generation, and that the policy over actions is still a neural LLM; it is therefore “not fully symbolic in the classical sense” (Wang et al., 23 Dec 2025). M-ASK likewise uses a natural-language structured state and answer-$\mathcal{T}_t = [\tau_1, \tau_2, \dots, \tau_m],$5-based rewards rather than formal symbolic validators (Chen et al., 8 Jan 2026). A common misconception is therefore that current ASYS already provides theorem-proving-style guarantees. In most cases it does not.

Another limitation is that trajectory-aware retrieval and optimization remain expensive. Agentic-R’s global answer correctness signal requires rollout-based passage evaluation, and its strongest lesson is conceptual rather than computationally cheap: evidence should be judged by whether it helps the search transition toward a successful terminal state (Liu et al., 17 Jan 2026). Evaluation is also still substrate-specific. RAVine is web-centric, ASQ lacks subquery relevance judgments, and LocalSearchBench is vertical-domain and tool-specific (Xu et al., 22 Jul 2025, Pezzuti et al., 19 Feb 2026, He et al., 8 Dec 2025).

Methodological concerns remain substantial. ASIA shows that closed-loop autonomous experimentation can discover strong hybrid structures, but it also raises explicit concerns about implicit test leakage, reduced methodological transparency, and reproducibility (Piga et al., 11 May 2026). This suggests that a mature ASYS stack will require not only explicit state, symbolic operators, and trajectory-aware rewards, but also trusted evaluators, provenance-preserving memory, and clearer separation between search, selection, and final reporting.

The immediate research direction is not a return to monolithic reasoning, but deeper integration of explicit state, typed action protocols, compact symbolic working memory, on-policy search regulation, trajectory-aware retrieval, and provenance-sensitive evaluation. The literature increasingly supports a single conclusion: controllable long-horizon search is achieved less by making one model “reason harder” than by structuring how search state is represented, updated, scored, and verified.