IntentionReasoner: Computational Intent Processing

• IntentionReasoner is a class of computational systems that represent, infer, and operationalize intentions to guide planning, execution, and adaptive control in robotics and language models.
• These systems integrate coarse symbolic reasoning with fine-grained action planning using methods like CR-Prolog, POMDP planning, and intention-conditioned perception to handle complex tasks.
• Empirical studies highlight significant improvements, with some architectures achieving 100% task success in simulated environments and robust adaptations to uncertainty in practical scenarios.

IntentionReasoner denotes a class of computational systems that represent, infer, and operationalize intentions in order to guide planning, execution, perception, explanation, or safeguarding. In the literature, the term names both a specific architecture for human–robot collaboration that encodes a formal theory of intentions and a broader family of mechanisms for active goal recognition, intention-conditioned grounding, embodied control, and intent-aware language processing. Across these usages, the common thread is that intention is treated not as a mere label but as a structured variable that constrains action selection, relevance filtering, and adaptation under uncertainty (Gomez et al., 2019, Ognibene et al., 2019, Shen et al., 27 Aug 2025).

1. Scope and major variants

The literature uses “IntentionReasoner” in several technically distinct but related senses. Some systems reason about the agent’s own intentions as named activities linked to goals; others infer another agent’s latent goal from motion, language, or scene context; others use intent as a control variable for LLMs or guard models. What unifies them is an explicit intermediate representation between raw observation and downstream action.

Family	Representative work	Technical core
Symbolic human–robot collaboration	"Towards a Theory of Intentions for Human-Robot Collaboration" (Gomez et al., 2019)	CR-Prolog ATI, coarse/fine transition diagrams, refinement and zooming
Active recognition under uncertainty	"Proactive Intention Recognition for Joint Human-Robot Search and Rescue Missions through Monte-Carlo Planning in POMDP Environments" (Ognibene et al., 2019)	POMDP belief updates, entropy-shaped Monte Carlo planning
Intention-conditioned perception	"RIO: A Benchmark for Reasoning Intention-Oriented Objects in Open Environments" (Qu et al., 2023)	Natural-sentence intentions, IOOD/IOIS, text–object compatibility
Embodied intention reasoning	"IntentionVLA: Generalizable and Efficient Embodied Intention Reasoning for Human-Robot Interaction" (Chen et al., 9 Oct 2025)	Compact reasoning, spatial grounding, VLA-conditioned action generation
Intent-aware language safeguarding	"IntentionReasoner: Facilitating Adaptive LLM Safeguards through Intent Reasoning and Selective Query Refinement" (Shen et al., 27 Aug 2025)	Multi-level safety labels, intent reasoning, selective query rewriting

A recurring distinction is between intention as an internal commitment of the acting system and intention as a latent variable attributed to another agent or to a user query. The first sense dominates symbolic BDI-like and activity-theoretic architectures; the second dominates POMDP active goal recognition, open-world grounding, and LLM safeguard pipelines.

2. Symbolic and hierarchical intention reasoning

The most explicit operationalization appears in the human–robot collaboration architecture of Gelfond- and CR-Prolog-based reasoning. That architecture encodes a theory of intentions around three principles: non-procrastination, persistence, and relevance. The coarse-resolution signature $\Sigma_c$ is expanded with mental fluents such as active_activity(activity), in_progress_activity(activity), next_action(activity, action), active_goal(goal), and current_action_index(activity, index), together with mental actions start(name), stop(name), select(goal), and abandon(goal). Coarse reasoning computes an activity of abstract actions, while each abstract action is implemented at fine resolution by refinement and zooming. Fine-level reasoning is bounded to the constants relevant to the current coarse transition or goal; bridge axioms such as $loc(O,P)\ \mathbf{if}\ component(C,P),loc^*(O,C)$ and $$in\_hand(R,O)\ \mathbf{if}\ component(Opart,O),in\_hand^*(R,Opart)$$ connect the levels. The architecture keeps ATI constructs only at the coarse level and applies refinement to physical fluents and actions (Gomez et al., 2019).

The execution cycle alternates two-phase non-monotonic reasoning, zoomed ASP planning, probabilistic execution, and coarse-history revision. Diagnostics first select a consistent history interpretation using defaults, exceptions, and exogenous actions; planning then computes an activity with projected_success. During execution, fine-resolution outcomes are elevated to observations, mapped back to coarse observations, and used to re-evaluate intended_action(...) and futile_activity(...). Empirically, this design produced 100% aggregate accuracy against 74% for traditional planning in simulated office scenarios; with zooming, successful completion remained 100% across complexity levels $L1$–$L8$, whereas without zooming success fell to 65% at $L3$ and 0% at $L4$–$L8$. On Baxter and Turtlebot trials, ATI+refinement+zooming achieved 100% goal completion, while removing ATI reduced accuracy to approximately 60% and increased time by approximately $2$–$3\times$ in dynamic scenarios (Gomez et al., 2019).

3. POMDP-based active intention recognition

A second lineage treats IntentionReasoner as an online planner over latent intentions under partial observability. In joint search-and-rescue, the problem is cast as a POMDP $loc(O,P)\ \mathbf{if}\ component(C,P),loc^*(O,C)$0 in which a drone infers a human responder’s intention from sparse motion observations and augments reward with information-theoretic terms. The responder follows a stochastic policy that, with probability $loc(O,P)\ \mathbf{if}\ component(C,P),loc^*(O,C)$1, moves one step along an optimal path toward the survivor location, with the remaining mass assigned to stillness or random movement. Belief is updated with a complete discrete filter or a truncated approximation, and goal hypotheses are reweighted by an inverse-planning cost ratio $loc(O,P)\ \mathbf{if}\ component(C,P),loc^*(O,C)$2. Entropy-shaped variants such as chES, thES, ehES, and fhES improve performance when Monte Carlo budget is limited: in the large environment with 100 simulations, chES reached 50% success versus 18% for the default sparse-reward planner (Ognibene et al., 2019).

Industrial robotic assistance adopts the same POMDP framing but with a generative target simulator $loc(O,P)\ \mathbf{if}\ component(C,P),loc^*(O,C)$3 for worker behavior. In maintenance and assembly domains, the assistant interleaves probing actions and assistance tasks while reasoning over a state space that includes worker activity, object focus, results of recent actions, and robot-side variables. RAGE, a relevance-based planner with Partial Goal Satisfaction shaping, consistently outperformed POMCP. In the reported maintenance setting, best average returns were $loc(O,P)\ \mathbf{if}\ component(C,P),loc^*(O,C)$4, $loc(O,P)\ \mathbf{if}\ component(C,P),loc^*(O,C)$5, and $loc(O,P)\ \mathbf{if}\ component(C,P),loc^*(O,C)$6 for low-, medium-, and high-expertise workers, compared with $loc(O,P)\ \mathbf{if}\ component(C,P),loc^*(O,C)$7, $loc(O,P)\ \mathbf{if}\ component(C,P),loc^*(O,C)$8, and $loc(O,P)\ \mathbf{if}\ component(C,P),loc^*(O,C)$9 for POMCP; in the assembly domain, the state space exceeded $$in\_hand(R,O)\ \mathbf{if}\ component(Opart,O),in\_hand^*(R,Opart)$$0, and 100% completion required at least 512 simulations per step for RAGE and at least 1024 for POMCP (Saborio et al., 2024).

A later robotic assistant framework makes uncertainty itself central to intention recognition. Its POMDP state includes latent human intention $$in\_hand(R,O)\ \mathbf{if}\ component(Opart,O),in\_hand^*(R,Opart)$$1, environment/task status $$in\_hand(R,O)\ \mathbf{if}\ component(Opart,O),in\_hand^*(R,Opart)$$2, human task progress $$in\_hand(R,O)\ \mathbf{if}\ component(Opart,O),in\_hand^*(R,Opart)$$3, and robot execution context $$in\_hand(R,O)\ \mathbf{if}\ component(Opart,O),in\_hand^*(R,Opart)$$4; perception actions cost $$in\_hand(R,O)\ \mathbf{if}\ component(Opart,O),in\_hand^*(R,Opart)$$5, premature restocking can incur $$in\_hand(R,O)\ \mathbf{if}\ component(Opart,O),in\_hand^*(R,Opart)$$6, and human assembly events yield $$in\_hand(R,O)\ \mathbf{if}\ component(Opart,O),in\_hand^*(R,Opart)$$7, $$in\_hand(R,O)\ \mathbf{if}\ component(Opart,O),in\_hand^*(R,Opart)$$8, or $$in\_hand(R,O)\ \mathbf{if}\ component(Opart,O),in\_hand^*(R,Opart)$$9. Observation accuracy is parameterized by $L1$0, with online planners POMCP and RAGE evaluated over budgets from 2 to 65,536 simulations per step. Starting at approximately 256 simulations per step, the planners consistently reached terminal states under heavy noise, and 20 Gazebo assistance runs all completed successfully, with mean scenario time $L1$1s (Saborío et al., 26 Aug 2025).

4. Perception-grounded and embodied variants

In vision-centric work, intention reasoning is formulated as grounding a natural-language intention in an open scene. RIO defines intention-oriented object detection and instance segmentation with inputs $L1$2, where $L1$3 is an image and $L1$4 is a natural-language intention sentence such as “Something you can use to create music.” The dataset contains 40,214 images, 130,585 intention–object pairs, and 69 COCO-derived categories, with intentions represented as scene-related natural sentences rather than verb words or short phrases. Evaluation is split into IOOD, measured by Top1-Accuracy and [email protected], and IOIS, measured by Top1-IoU and mIoU. On the common test set, end-to-end models outperform the detector–LLM hybrid: TOIST reaches [email protected] $L1$5 and Top1-Acc $L1$6, while D-R reaches $L1$7 and $L1$8; on uncommon intentions, all models degrade sharply, underscoring the difficulty of long-tail affordance reasoning (Qu et al., 2023).

Embodied VLA work internalizes intention reasoning inside the policy itself. IntentionVLA trains a Qwen2.5-7B-based backbone to emit intention tokens, spatial grounding, and discrete action cues, then compresses them into compact short reasoning of the form “move <direction> to <object>.” These compact outputs condition a connector and a diffusion-based action generator. The reported average success rates are 48.3% on direct instructions and 45% on intention instructions, versus 30% and 20% for $L1$9; on unseen instructions IntentionVLA averages 30% success, and in zero-shot real human–robot interaction it reaches 40% for “Phone on real hand” and 30% for “Phone on moving hand,” with average inference time 0.72s versus 3.41s for ECoT (Chen et al., 9 Oct 2025).

A related humanoid system, INTENTION, combines a VLM-based Intuitive Perceptor, a MemoGraph of prior scene graphs, and a motion-primitive library. A task graph is represented as $L8$0, and retrieval scores combine node, link, and instruction similarity through $L8$1. On interactive intuition tasks, INTENTION achieved Plan 84% and Succ 72%, versus Plan 20% and Succ 15% for LLM-BT; on standard manipulation, INTENTION reported Plan 96%, Succ 92%, and Avg. Time $L8$2 (Wang et al., 6 Aug 2025).

Egocentric anticipation systems also treat intention as an explicit intermediate. INSIGHT uses HOI regions detected by 100DOH and SAM2, verb–noun co-occurrence correction, and a GRPO-trained think $L8$3 reason $L8$4 answer module. On Ego4D v2 it reports Verb ED 0.6643, Noun ED 0.6092, and Action ED 0.8463; on EK-55 it reports mAP 45.2/62.4/36.0 for All/Freq/Rare; on EGTEA it reports 81.7/85.9/74.4. Ablations show the largest degradation when cognitive reasoning is removed, with Action ED rising from 0.8463 to 0.8612 (Chu et al., 3 Aug 2025).

5. Language-model reasoning and safeguard systems

For LLMs, intention reasoning is often introduced as an explicit textual scaffold. SWI (“Speaking with Intent”) requires the model to generate free-form intent before and during analysis, then to conclude with “Final Answer:”. It is an inference-time method with no parameter updates, implemented on LLaMA3-8B-Instruct. On mathematical reasoning benchmarks, the reported average rises from 51.74 to 55.08 over Baseline; on reasoning-intensive QA, the average gain is +10.08 points; on summarization, average ROUGE rises from 17.11 to 19.03, and atomic-fact precision improves consistently even when recall drops because summaries become more concise (Yin et al., 27 Mar 2025).

Multi-agent LLM settings recast the problem as attributional inference. Attributional NLI formalizes a two-stage process in which Stage 1 selects intention hypotheses $L8$5 such that $L8$6, and Stage 2 derives conclusions $L8$7 such that $L8$8. In the Undercover-V game, neuro-symbolic Att-NLI agents combine abductive–deductive inference with Isabelle/HOL, Sledgehammer, and autoformalization via Neo-Davidsonian event semantics. Across experiments, neuro-symbolic agents consistently outperform standard NLI and standard Att-NLI, with an average spy win rate of 17.08% (Quan et al., 13 Jan 2026).

Safety-oriented work uses IntentionReasoner as a front-end guard model. The system is trained on approximately 163,000 queries annotated with <thinking>, a four-level label—Completely Unharmful, Borderline Unharmful, Borderline Harmful, Completely Harmful—and a <refined query> rewrite. Supervised fine-tuning is followed by GRPO-style multi-reward optimization over format, classification, rewrite safety, rewrite utility, and length. At 7B scale, the model reports F1 99.4 with ASR/ORR 1.2/0.0 across six harm-detection benchmarks; on jailbreak evaluation against Qwen2.5-7B-Instruct, average ASR falls to 0.4 compared with 74.4 for the unguarded model, and on GPT-4o it falls to 0.8 compared with 40.0 without the guard (Shen et al., 27 Aug 2025).

6. Formal foundations, interpretation, and recurrent limits

Theoretical work frames intention as a structured mental state rather than a synonym for goals or plans. One influential analysis decomposes intention through planning, action, and control lenses: intentions store results of deliberation for future action, stabilize behavior through commitment, and support guidance, monitoring, and triggering conditions across a distal $L8$9 proximal $L3$0 motor cascade. That analysis emphasizes skill, awareness, and “intentional under a description,” and argues that cognitive systems frustrate users when they lack the folk-psychological profile associated with beliefs, goals, and intentions (Bridewell, 2022).

A more formal decision-theoretic account defines a branching-time possible-worlds model $L3$1, together with a transformation from decision trees into goal-accessible worlds. Deliberation functions such as maximin and maxexpval compute best action sequences $L3$2, and Theorem 1 shows that if $L3$3, then $L3$4. This makes intention formation explicitly dependent on deliberation with probabilities and payoffs rather than on a purely modal postulate (Rao et al., 2013).

Hyperintensional logic pushes the point further by rejecting not only closure under entailment but also closure under equivalence. In the proposed system, $L3$5 iff $L3$6 and $L3$7, where $L3$8 maps formulas to the decision problems for which they are partial solutions. The logic validates agglomeration, consistency, and a restricted closure principle, but blocks $L3$9 even when $L4$0 is valid. This is designed to preserve the distinction between intended outcomes and unintended side-effects, and between propositions that are truth-conditionally equivalent yet differ in control profile (Khaitovich et al., 27 Nov 2025).

Taken together, these papers indicate recurrent limitations. Some architectures assume designer-provided domain descriptions and refinement schemas rather than online model learning; some elevate high-probability observations to certainty; some represent other agents only through exogenous actions or latent goal variables; some remain computationally expensive under online planning; and perception-grounded systems still degrade sharply on uncommon intentions, unseen instructions, or novel objects (Gomez et al., 2019, Saborío et al., 26 Aug 2025, Qu et al., 2023, Chen et al., 9 Oct 2025). A plausible implication is that “IntentionReasoner” is best understood not as a single architecture but as an interface problem: how to connect commitment, relevance, and latent-goal inference to scalable perception, planning, and explanation without collapsing intention into either raw classification or unrestricted logical closure.