--- title: 'LongSeeker: Elastic Context Orchestration' url: https://www.emergentmind.com/topics/longseeker type: topic --- # LongSeeker: Elastic Context Orchestration LongSeeker is a long-horizon search agent introduced as an instance of **elastic context orchestration**, a design in which the agent treats its working context as a controllable state rather than as an append-only transcript. It instantiates the **Context-ReAct** paradigm, under which each step jointly produces reasoning, a set of context-management meta-operations, and a tool call. LongSeeker is fine-tuned from **Qwen3-30B-A3B** on **10k synthesized Context-ReAct trajectories** and is reported to achieve **61.5% on BrowseComp** and **62.5% on BrowseComp-ZH**, while keeping average context size stable, plateauing around **15k tokens** even on runs of up to **300 steps** [2605.05191]. ## 1. Problem formulation and motivation LongSeeker is motivated by a specific failure mode of long-horizon search agents: as they plan, browse, and accumulate observations, their working context grows monotonically and begins to degrade both efficiency and reliability. In the formulation used for Context-ReAct, a standard ReAct step is written as $$ S_i^{std} = (r_i, c_i, o_i) $$ where $r_i$ is the chain-of-thought, $c_i$ the tool call, and $o_i$ the returned observation. The append-only history at time $t$ is $$ H_t = [S_1^{std}, S_2^{std}, \ldots, S_t^{std}], $$ and the working context state is denoted $C_t$, typically with $C_t = H_t$ or a managed variant. Under naive accumulation, token cost follows $$ cost(C_{t+1}) = cost(C_t) + |S_{t+1}|, $$ with $cost(C) \approx |C|$ as a token-cost proxy [2605.05191]. This setup yields three linked difficulties. First, inference cost and latency scale roughly with context length. Second, the signal-to-noise ratio deteriorates because stale or redundant traces remain visible to later reasoning steps. Third, hallucination risk increases because factual details are repeatedly paraphrased or re-generated instead of being preserved in a controlled form. A common misconception is that sufficiently large context windows eliminate this problem. LongSeeker explicitly argues otherwise: even with a large **256k** model context window and a **maximum tool-call budget of 300**, the agent still benefits from active memory shaping because the issue is not only capacity, but also the quality and relevance structure of the context presented to the model [2605.05191]. ## 2. Context-ReAct as the core agentic paradigm Context-ReAct generalizes ReAct by inserting context management directly into the decision loop. A Context-ReAct step is defined as $$ S_i^{meta} = (r_i, M_i, c_i, o_i), $$ where $$ M_i = [op_1^{(i)}, \ldots, op_k^{(i)}] $$ is a list of meta-operations applied to the current context before the next observation is appended. The effective transformation is $$ C_t' = T(C_t, M_t), $$ with composition $$ T(C, [op_1, \ldots, op_m]) = op_m(\ldots op_2(op_1(C)) \ldots). $$ The managed context at the end of step $t$ is then $$ C_t = Append(T(C_{t-1}, M_t), S_t). $$ In procedural terms, the model first generates reasoning, meta-operations, and a tool call in a single pass; the meta-operations are then executed; the environment returns the observation; and only after that is the current step appended [2605.05191]. The central design choice is that the agent decides **when, where, and how** to reshape memory before new evidence arrives. This differs from threshold-triggered summarization policies such as “summary-on-overflow,” because the management policy is state-dependent and co-generated with reasoning itself. The paper presents this as a unified loop in which context is neither a passive buffer nor an external post-processing artifact, but a first-class control variable. In practice, this means LongSeeker can preserve newly verified evidence verbatim, compress older resolved subthreads, abandon dead-end branches, or delete clearly irrelevant steps before they contaminate later inference [2605.05191]. ## 3. Atomic operations and theoretical properties Context-ReAct provides five atomic operations: **Skip**, **Compress**, **Rollback**, **Snippet**, and **Delete**. They act on a context $$ C = [S_1, \ldots, S_n] $$ where each step has the form $S_k = (r_k, c_k, o_k)$. | Operation | Semantics | Fidelity / complexity | |---|---|---| | **Skip** | $Skip(C)=C$ | Identity, $O(1)$ | | **Compress** | Replace a contiguous range $[a,b]$ with summary $\Sigma$ | Lossy, $O(|\Sigma|)$ | | **Rollback** | Truncate to step $k$ and summarize with $\Sigma$ | Localized loss, $O(|\Sigma|)$ | | **Snippet** | Replace an observation with an exact substring slice | Lossless, $O(1)$ | | **Delete** | Remove step $k$ entirely | Exact removal, $O(1)$ | The paper gives the operations explicitly. For example, $$ Compress(C, a, b, \Sigma) = [S_1, \ldots, S_{a-1}, S_{a:b}=\Sigma, S_{b+1}, \ldots, S_n], $$ $$ Rollback(C, k, \Sigma) = [S_1, \ldots, S_k=\Sigma], $$ $$ Snippet(C, k, pre, suf) = [S_1, \ldots, (r_k, c_k, o_k[pre:suf]), \ldots, S_n], $$ and $$ Delete(C, k) = [S_1, \ldots, S_{k-1}, S_{k+1}, \ldots, S_n]. $$ The distinction between **Snippet** and **Compress** is especially important. Snippet is pointer-based and lossless for the retained segment, which the paper highlights as eliminating hallucination on exact literals such as numbers, entities, URLs, code, or quotes. Compress is generative and lossy, but offers the largest token savings on long spans [2605.05191]. A common misunderstanding is that the five operations are merely an ad hoc toolkit. The paper instead proves a formal expressivity result: **Compress alone is expressively complete**. The theorem states that for any $H_{in}, H_{target} \in H$, there exists a single configuration $\kappa=(a=1, b=|H_{in}|, \Sigma=H_{target})$ such that $$ Compress_\kappa(H_{in}) = H_{target}. $$ Equivalently, for any history-to-history transformation $f:H \rightarrow H$, one can choose $\kappa_f(H)=(1, |H|, \Sigma=f(H))$ so that $$ Compress_{\kappa_f(H)}(H)=f(H). $$ This establishes universality, but not practical optimality. The specialized operators remain important because they provide efficiency and fidelity guarantees that Compress does not. Skip introduces zero cost and zero risk; Snippet preserves exact strings without regeneration; Delete removes noise without abstraction; and Rollback structurally discards failed suffixes while retaining a localized summary of lessons learned [2605.05191]. ## 4. Model, supervision, and training pipeline LongSeeker is built by supervised fine-tuning rather than reinforcement learning. The base model is **Qwen3-30B-A3B**, and each step is emitted in a four-field structured format: **``**, **``**, **``**, and **``**. The supervised objective is standard next-token prediction over the full structured output, conditioned on the managed context: $$ L_{SFT} = -\sum_{t=1}^{T}\sum_j \log p_\theta(x_j^{(t)} \mid x_{