---
title: 'Intention Tree: Hierarchical Reasoning'
url: https://www.emergentmind.com/topics/intention-tree
type: topic
---

# Intention Tree: Hierarchical Reasoning

to=arxiv_search.search  qq的天天中彩票  全民彩票天天
  "query": "\"intention tree\" OR \"Definitions of intent suitable for algorithms\" OR \"Hierarchical Intention Tracking\"",
  "max_results": 10,
  "sort_by": "relevance"
}
to=arxiv_search.search  微信的天天中彩票  天天中彩票开奖
  "query": "\"Intention Tree\" arXiv",
  "max_results": 10,
  "sort_by": "relevance"
}
to=arxiv_search.search  亚洲男人天堂  天天中彩票派奖
  "query": "Definitions of intent suitable for algorithms 2106.04235",
  "max_results": 5,
  "sort_by": "relevance"
}
An intention tree is a tree-structured representation used to organize, diagnose, infer, or operationalize intention. In the recent arXiv literature, the term does not denote a single canonical formalism. Instead, it appears in several technically distinct forms: a hierarchical legal-diagnostic procedure for classifying algorithmic intent; a probabilistic hierarchy of latent intentions in human–robot collaboration; a behavior-tree realization of goals extracted from natural-language instructions; a branching session model for evolving customer intent; and an intention-aware planning structure for tool-using language agents [2106.04235, 2506.07004, 2405.07474, 2507.20185, 2602.11541].

## 1. Principal meanings of the term

The surveyed papers use the term in several recurring ways. Some works treat the tree as a diagnostic object, some as a latent-state model, and some as an executable or explanatory structure. This variation is substantive rather than terminological: the node semantics, edge semantics, and inference procedures differ by domain.

| Domain | Tree meaning | Representative paper |
|---|---|---|
| Algorithmic culpability | Hierarchical decision procedure for direct, means-end, oblique, and ulterior intent | [2106.04235] |
| Collaborative robotics | Rooted tree of discrete intention variables across semantic levels | [2506.07004] |
| Industrial assembly | High-level interaction mode gating low-level task goals | [2203.09063] |
| Natural-language task execution | Behavior tree derived from logical goals; treated as an intention tree in practice | [2405.07474] |
| Dialogue systems | Intent graph that becomes tree-like under acyclicity and single-parent conditions | [2310.11818] |
| E-commerce session modeling | Time-indexed branching structure with intention, attributes, and comparisons | [2507.20185] |

A broader human-centered framing treats intention itself as multi-faceted rather than identical to a task goal. The taxonomy in [2602.15963] organizes human intention along five axes—goal-oriented versus implementation, implicit versus explicit, conscious versus unconscious, individual versus collective, and short-term versus long-term—thereby supplying a conceptual “tree” whose leaves are intention subcategories rather than task states.

## 2. Formal structures and semantics

A strict probabilistic intention tree is given in hierarchical intention tracking for collaborative robotics. There, an intention tree is a rooted, directed tree $T=(V,E)$ of arbitrary depth $D$, whose nodes are discrete intention variables arranged in levels $l=1,2,\dots,D$, with $l=1$ the most concrete level and $l=D$ the root. Each node has at most one parent; leaves encode the most concrete intentions. Parent nodes constrain children via $p(I_{c,t}\mid I_{p,t})$, transitions among siblings are permitted, and transitions across different parents are suppressed. Inference proceeds by Bayesian filtering at each level, upward measurement propagation from leaves to parents, and downward posterior propagation from parents to children [2506.07004].

A simpler but influential two-level instantiation appears in industrial assembly. The latent intention is $G_t=\{G^2_t,G^1_t\}$, where the high-level state $G^2_t$ encodes interaction pattern—Coexistence or Cooperation—and the low-level state $G^1_t$ encodes task goals or Failure Recovery. The hierarchy is explicitly asymmetric: under Cooperation, $P(FR\mid x_{1:t}^{h,r},CO)=1$ and $P(i\mid x_{1:t}^{h,r},CO)=0$, whereas under Coexistence, task goals $i\in\{1,2,3,4\}$ remain admissible and Failure Recovery is gated off. Higher levels evolve more slowly than lower levels through a persistence parameter with $\alpha^2>\alpha^1$ [2203.09063].

In teleoperation manipulation, the hierarchy is task $\rightarrow$ action. The intention at time $t$ is $H_t=(T_t,A_t)$, with $m=6$ high-level tasks and $n=21$ low-level actions. The admissibility constraint is $A_t\in A^{T_t}$, where each task has a legal action set. The neural architecture shares a root encoder, adds task and action heads, conditions the action head on the task embedding, and trains with cross-entropy plus a hierarchical dependency loss. The same work uses a multi-window scheme in which the task head sees $35$ frames, approximately $3.5\,\text{s}$, while the action head sees only the most recent $10$ frames, approximately $1.0\,\text{s}$, through masking [2403.19770].

Other papers weaken or transform the tree assumption. In multi-turn dialogue, the primary object is a directed multigraph $G=\{(e_s,r,e_o)\mid e_s,e_o\in E,r\in R\}$ consisting of feature nodes, query nodes, and a root node; it behaves as an intention tree only if there are no cycles, every non-root node has exactly one parent, and query nodes are leaves [2310.11818]. In deliberation theory, the intention tree is the decision tree pruned by the deliberative policy: once a best plan is selected, unchosen decision branches are removed, and the resulting structure corresponds to intention-accessible subworlds in a branching-time possible-worlds model [1303.5742].

## 3. Diagnostic intention trees for algorithmic actors

A fully explicit diagnostic intention tree is developed for algorithmic agents. Its purpose is to classify whether an algorithm satisfies direct intent, means-end intent, oblique intent, or ulterior intent, using a sequence of stopping conditions. The root is a capacity gate: the agent must have state access, action choice, a subjective causal or probability model, plans, and aims. If that capacity is absent, the output is “No assessable intent.” If capacity is present, the procedure evaluates direct intent at commission through four conditions: Free Agency, Knowledge, Foreseeable Causality, and Aim. Aim can be explicit, as in $G(X=x)=1$ or an objective function $U$ that prefers $X=x$, or implicit, as in $\exists a' : P^D(X=x\mid A=a) > P^D(X=x\mid A=a')$ [2106.04235].

If direct intent fails, the tree proceeds to means-end intent, which requires that some other result $Y=y$ is directly intended, that $X=x$ is a necessary intermediate for $Y=y$, and that the action causing $X=x$ is a subsequence of the action sequence pursuing $Y=y$. If that also fails, the procedure checks oblique intent. Oblique intent requires a directly intended result $Y=y$ plus a side effect that is “virtually certain” according to the agent’s own subjective model at the point of commission. The operational threshold proposed in the details is a high subjective probability, for example $P^D(H=1\mid A=a)\ge 0.95$ or $P^D(H=1\mid Y=1,A=a)\ge 0.95$. Ulterior intent is then defined for future results: at time $t_1$, there must exist a foreseeable state of the world $S=s$ at some later time $t_2>t_1$ such that the agent would directly or obliquely intend $X=x$ via $A=a$, and the policy must already commit to $A=a$ should $S=s$ occur.

This diagnostic tree is explicitly distinguished from neighboring mental-state categories. Knowledge or observability is necessary but not sufficient for intent, because intent also requires aim. Recklessness involves conscious disregard of a substantial and unjustifiable risk but lacks either aim or the virtual-certainty condition required for oblique intent. Negligence applies where subjective awareness is absent but the agent objectively “should have known.” The paper’s worked examples—the unreliable plane bomb, the dud bomb judged under subjective belief, autonomous trading “spoofing,” and an autonomous vehicle that causes property damage while avoiding certain death—show how the same tree can classify distinct intentional modes without collapsing them into mere foreseeability.

## 4. Hierarchical intention tracking in robotics

In collaborative robotics, intention trees are primarily latent-state models for real-time adaptation. The 2025 Hierarchical Intention Tracking system represents human intentions at task, interaction, and verification levels and updates them every timestep by Bayesian filtering, upward measurement propagation, and downward posterior propagation. The implementation uses sibling-constrained transitions with persistence $\kappa^l=0.99$, leaf-level wrist-position likelihoods obtained from a Gaussian intention-aware motion model, and rule-based switching among three tracking modes: Interaction–Task, Verification–Task, and PAUSE. Switching thresholds are explicit: if $p(I_{2,t}=\text{Cooperation})>0.9$ continuously for at least $0.5\,\text{s}$, the system moves from Interaction–Task to Verification–Task; if Verification–Task infers Abnormal with probability $>0.8$ for at least $1.0\,\text{s}$, it returns immediately to Interaction–Task; if end-effector force exceeds $20\,\text{N}$, it enters PAUSE and activates admittance control [2506.07004].

That architecture is tied to a collaborative assembly task with six task leaves—$\{0,1,2,3,\text{prep},\text{failure recovery}\}$—and two alternative parents depending on the active tree. In a user study with $10$ participants, the HIT-based systems eliminated assembly failures, reporting $0.0$ failures per trial compared with $1.84$ for the Coexistence baseline. The Cooperation baseline required substantial average force, approximately $9.42\,\text{N}$, and energy, approximately $31.83\,\text{J}$, whereas HIT-ITVT reduced these to approximately $1.03\,\text{N}$ and approximately $8.88\,\text{J}$. The same study reports that HIT-ITVT reduced interruptions by correcting false cooperation detections before physical contact.

The 2022 precursor already established the same core logic in a two-level form. It tracked low-level intentions at $30\,\text{Hz}$ and high-level intentions at $5\,\text{Hz}$, used two RGBD cameras, OpenPose, Kalman filtering, and force/torque sensing, and integrated the intention tree with coexistence and cooperation control modules. In the reported assembly study, the HIT system achieved $97.2\pm 15.5\,\text{s}$ completion time, guided path length $0.65\pm 0.37\,\text{m}$, human force $0.48\pm 0.24\,\text{N}$, human energy $5.71\pm 3.30\,\text{J}$, and $0.0$ failures, whereas the Coexistence baseline had $1.2$ failures and the Cooperation baseline required $5.51\pm 0.39\,\text{N}$ average human force [2203.09063].

A parallel deep-learning line addresses teleoperation manipulation. Hierarchical LSTM, GCN, and SlowFast models predict both tasks and actions online, and the hierarchical variants improve per-frame accuracy over independent baselines. On motion features, Hie-LSTM reports $95.41\%$ action accuracy and $98.25\%$ task accuracy versus $92.27\%$ and $96.79\%$ for independent LSTM. The multi-window variant also outperforms a non-specialized hierarchical model: Hie-NN-W reports $95.41\%$ action and $98.25\%$ task accuracy, compared with $93.82\%$ and $95.89\%$ for Hie-NN-O [2403.19770].

The taxonomy of human intention in robotics places these systems in a wider conceptual frame. It argues that intention is not exhausted by task goals and instead must be classified across five axes: type of goal, communication channel, conscious access, social scope, and temporality. A plausible implication is that many existing robotic intention trees model only a subset of the relevant facets—most commonly goal-oriented versus implementation and short-term versus long-term—while leaving consciousness, explicitness, or collective commitment to auxiliary modules or user interfaces [2602.15963].

## 5. Intention trees as executable, explanatory, and sequential structures

In task execution from language, the tree becomes an executable control object. A two-stage framework first maps natural-language instructions to first-order logical goals represented as well-formed formulas using only $\land$, $\lor$, and $\neg$, then normalizes those goals to disjunctive normal form, and finally constructs a Behavior Tree through the Optimal Behavior Tree Expansion Algorithm. The resulting BT is treated as an intention tree in practice because it encodes the interpreted goal, contingency structure, and reactive recovery semantics. The paper gives theoretical guarantees of finite-time success and optimal cost under reachability assumptions, with worst-case complexity $O(|\mathcal{A}||\mathcal{S}|\log|\mathcal{S}|)$. In the café scenario, OBTEA reduced both BT cost and condition-node ticks relative to BT Expansion; for hard goals, cost was $111.7$ versus $204.8$, and condition ticks were $809.7$ versus $1437.4$ [2405.07474].

In dialogue systems, the tree is often a traversal structure over intent elements rather than a latent-variable hierarchy. IntentDial organizes feature nodes, query nodes, and a root node into an intent graph and uses reinforcement learning to traverse from the root toward a standard query. States are of the form $s_t=(e_t,e_r,C_i)$, actions are outgoing edges from the current node, and the policy is optimized with REINFORCE. Query nodes have out-degree $0$, and when the path terminates at a key feature rather than a query node, the system issues a clarifying prompt and visualizes the current reasoning path for that turn [2310.11818].

Session modeling introduces a different tree semantics again. SessionIntentBench defines the intention tree over products $P_1,\dots,P_T$ as a time-indexed branching structure whose nodes at time $t$ contain an intention $I_t^\zeta$, decisive attributes $A_t^\zeta$, and a comparison $C_t^\zeta$ explaining the transition from the previous product. The branching policy is explicit: for $t\le 5$, the factor is $b_t=5$, and for $t>5$, it is $b_t=1$. The resulting benchmark contains $1{,}952{,}177$ intention entries, $1{,}132{,}145$ session intention trajectories, and $13{,}003{,}664$ available tasks mined using $10{,}905$ sessions. The benchmark is designed precisely to force models to reason over inter-session intention dynamics rather than over product titles alone [2507.20185].

A related planning interpretation appears in budget-constrained tool-using language agents. INTENT reformulates planning as an intention-aware hierarchical world model in which the latent variable $z_t\in\{0,1\}$ indicates whether the tool outcome satisfies the semantic intention expressed in the reasoning trace $r_t$. The planning tree alternates intention nodes, action nodes, and observation nodes, while each tool action carries a price $c_t$, a predicted success probability $p_t$, and a geometric expected cost $\hat{C}_t=c_t/p_t$. Acceptance is governed by a hard feasibility rule: the immediate action must satisfy $\mathrm{COST}(a_t)\le b_t$, and the projected risk-adjusted cumulative cost $\gamma\sum_{k=t}^K c_k/p_k$ must also fit within the remaining budget. On cost-augmented StableToolBench, INTENT reports $100\%$ feasible rate for both GPT-4.1 mini and GPT-5 nano backbones, with pass rates of $63.8\%$ and $76.0\%$, respectively [2602.11541].

Other uses are looser and partly interpretive. In early malice detection for Bitcoin, the paper does not use the exact term but induces a tree-structured representation from decision-tree-based segmentation, clustered global status vectors, and survival-guided path selection; the discovered “intention motif” is a path through status nodes such as the reported Binance hack sequence $[0\!-\!0\!-\!9\!-\!5\!-\!0\!-\!9\!-\!0\!-\!0\!-\!0]$ [2209.12001]. In migration-intention prediction, the term denotes interpretable tree-based classifiers—Decision Trees, Random Forests, and XGBoost—used to explain who intends to migrate and why, with feature-importance and partial-dependence analyses standing in for an explicit hierarchical semantics [2012.02794].

## 6. Limits, misconceptions, and open questions

A persistent misconception is that an intention tree is necessarily a goal tree. The literature does not support that restriction. Some intention trees are about culpability classes rather than goals; some are about task–action hierarchies; some encode explanatory metadata such as attributes and comparisons; some are behavior trees compiled from logical goals; and some are only tree-like after additional assumptions or editorial reformulation [2106.04235, 2405.07474, 2507.20185, 2209.12001]. The taxonomy paper makes the deeper point explicit: a universally accepted definition of intention remains elusive, and existing works often equate human intention with specific task-related goals [2602.15963].

Another important distinction concerns whether the underlying structure is truly a tree. HIT and related robotics formalisms satisfy the strict rooted-tree definition directly. IntentDial is natively a graph and becomes a tree only under acyclicity and single-parent constraints. Behavior trees are executable trees but not necessarily probabilistic latent-intention models. Session intention trees branch early and collapse later by design, making tractability part of the representation rather than a by-product of inference [2506.07004, 2310.11818, 2405.07474, 2507.20185].

Methodological limits recur across domains. Algorithmic-intent diagnosis depends on access to the agent’s subjective model, alternative actions, and prediction logs; model-free RL, distributed systems, reward hacking, omissions, and intervening acts complicate that access [2106.04235]. Hierarchical tracking in robotics assumes discrete intention sets, short-horizon Markovian dynamics, sibling-restricted transitions, and hand-designed thresholds or Gaussian observation models; deep hierarchies and large branching factors raise computational cost, and continuous or hybrid intention spaces require extensions such as approximate inference or learned observation models [2506.07004, 2203.09063]. In budgeted LLM planning, the geometric retry approximation $\hat{C}=c/p$ and the usefulness of the intention-aware gate depend on calibration quality; overconfidence or underconfidence in $p_t$ changes pruning behavior even though hard budget feasibility remains guaranteed by per-step affordability checks [2602.11541].

The main research direction suggested by these papers is not convergence on one universal structure but sharper alignment between structure and use case. Where legal diagnosis is required, the tree must expose stopping conditions and subjective foreseeability. Where real-time collaboration is required, it must support multilevel filtering, switching, and control integration. Where language instructions or browsing sessions are primary, the tree must encode alternatives, explanations, and recovery. This suggests that “intention tree” is best understood as a structural design pattern for intention-centric reasoning rather than as a single formal object.

Source: https://www.emergentmind.com/topics/intention-tree