Alignment Games: A Framework for Conceptual Repair in Human-AI Collaboration

Hari Subramonyam; Maneesh Agrawala; Sean Follmer

cs.HCSep 28, 2026

Alignment Games: A Framework for Conceptual Repair in Human-AI Collaboration

Authors: Hari Subramonyam, Maneesh Agrawala, Sean Follmer

Organizations: Stanford University USA

Abstract

The meaning of a concept in use is shaped by the situation, task, goals, and prior knowledge. For example, a request to make a poster "visually appealing for a five-year-old" might evoke bright colors and cartoon imagery for one collaborator, but less text, bold shapes, and visual simplicity for another. We call such task-relevant differences conceptual misalignment. We introduce Alignment Games, a framework for making these differences visible and repairable during human-AI interaction. Drawing on theories of situated conceptualization, we characterize task-specific conceptual frames in terms of relevant attributes, values, relations, constraints, and priorities. We then define alignment moves that intervene on the situation, the reasoning used to interpret it, or the resulting frame. Through examples from educational content generation, creative coding, and argumentative writing, we show how these moves can be composed into repair sequences and derive design principles for supporting task-sufficient conceptual alignment at runtime.

Figures & tables

Figure 1. Conceptual Misalignment in Human-AI Collaboration: Given the same request to make a poster “more visually appealing for a five-year-old,” the requester (a) and AI agent (c) construct different conceptual frames from the compressed natural language instruction (b). The requester envisions bold shapes and minimal text (top), while the agent generates bright colors and cartoon imagery (right, d). Both interpretations are valid yet misaligned, demonstrating the challenge of conceptual coordination without explicit repair mechanisms. Diagram of a requester and an AI agent each forming a different mental picture from the same instruction. A plain blue science-camp poster on the left branches into two revisions: the requester's imagined version (bold shapes, minimal text) at top, and the agent's generated version (bright yellow, cartoon atom, colorful buttons) at right.

Figure 2. Examples from the iterative development of the Alignment Games framework. Across collaborative working sessions, we explored alternative frame primitives and representations, connected candidate representations to theories of situated cognition, and worked through concrete human–AI interaction examples to refine the framework and alignment-move vocabulary. Collage of four photographs and screenshots from working sessions, each with a rounded label. Clockwise from top left: a digital whiteboard of boxed frame primitives and notes; handwritten whiteboard sketches of node-and-arrow concept maps; a flowchart mockup ending in bar-chart dashboards; and a meeting room with walls covered in sticky notes and paper diagrams.

Figure 3. Process model of situated conceptualization. External situational elements, internal task context, and prior knowledge shape an iterative cycle of perception and attention, retrieval, reasoning, and simulation in working memory. Through this cycle, a collaborator constructs a situated conceptual frame that organizes the attributes, values, relations, constraints, priorities, and broader framings relevant to the task. Actions guided by the frame alter the artifact or situation, producing new evidence that can trigger subsequent cycles of construction and revision. Process diagram reading left to right along a "Time" arrow. A person views a science-camp poster and instruction; arrows labeled Perception, Retrieval, Reasoning, and Simulation move through a large cloud labeled Working Memory, where perceptual chunks and conceptual fields combine into a frame with attributes, values, and constraints, producing a Situated Conceptual Frame at the far right.

Source	Dimensions that can vary	Possible frame differences
External situation	Grain size, tangibility, familiarity, and temporal, spatial, social, or hypothetical distance can alter what aspects of a situation are available or treated as relevant ( Yeh and Barsalou, 2006 ; Barsalou, 2008 ; Schank and Abelson, 2013 ; Logan, 1988 ; Dane, 2010 ; Trope and Liberman, 2010 ) .	Different attributes or priorities become salient ( $\Delta_{A},\Delta_{W}$ ), while different values, constraints, or levels of abstraction may be inferred ( $\Delta_{V},\Delta_{C},\Delta_{CF}$ ).
Internal task context	Goals, goal abstraction, motivations, values, roles, resource constraints, and affect shape perceptual selection and what knowledge is recruited ( Barsalou et al., 2018 ; Hommel et al., 2001 ; Barsalou, 1983 ; Hass et al., 2019 ; Vallacher and Wegner, 1987 ; Trope and Liberman, 2003 ; Isen and Daubman, 1984 ; Gasper and Clore, 2002 ; Han et al., 2014 ) .	Different attributes become salient ( $\Delta_{A},\Delta_{W}$ ), and collaborators may establish different optimization criteria, preferred values, or constraints ( $\Delta_{O},\Delta_{V},\Delta_{C}$ ).
Prior knowledge, acquisition, and expertise	Different exemplars, prototypes, causal theories, episodic experiences, learning modalities, and levels of expertise provide different material for conceptual construction ( Tenenbaum et al., 2011 ; Posner and Keele, 1968 ; Rosch and Mervis, 1975 ; Medin and Schaffer, 1978 ; Murphy and Medin, 1985 ; Gopnik, 1993 ; Barsalou, 1985 ; Medin et al., 1997 ; Barsalou, 1999 ; Wauters et al., 2003 ; Villani et al., 2019 ; Juhasz, 2005 ; Barsalou, 2020 ) .	Different candidate attributes, defaults, relations, or broader conceptual structures may be recruited ( $\Delta_{A},\Delta_{V},\Delta_{R},\Delta_{CF}$ ).
Social and cultural experience	Instruction, social learning, community norms, and cultural experience shape which distinctions are learned, which values are conceivable, and which ideals are desirable ( Mervis, 1987 ; Waxman et al., 1997 ; Schwartz et al., 2011 ; Norenzayan et al., 2002 ; Naous et al., 2023 ; Borghi and Cimatti, 2009 ; Malt, 1995 ; Kövecses, 2005 ; Atran and Medin, 2008 ; Waxman et al., 2007 ; Ojalehto and Medin, 2015 ; Barsalou, 2023 ) .	Different defaults, acceptable values, constraints, optimization criteria, and broader framings can emerge ( $\Delta_{V},\Delta_{C},\Delta_{O},\Delta_{CF}$ ).

Table 1. Sources of divergence in conceptual inputs. Differences in the information available to collaborators can lead them to construct different situated conceptual frames.

Source	Dimensions that can vary	Possible frame differences
Perception and attention	Perceptual sensitivity, expertise, goals, attentional focus, joint attention, and cognitive load affect which information collaborators extract from the same situation ( Hommel et al., 2001 ; Chase and Simon, 1973 ; Brennan, 1995 ; Tomasello and Farrar, 1986 ; Swallow and Jiang, 2013 ; Xu et al., 2011 ) .	Different evidence enters conceptualization, affecting represented attributes and their salience ( $\Delta_{A},\Delta_{W}$ ), with downstream effects on other frame components.
Retrieval	Frequency and recency of use, contextual relevance, encoding strength, interference, and semantic-network structure affect which knowledge becomes accessible ( Murphy, 2004 ; Anderson and Reder, 1999 ; Buchanan et al., 2001 ; Pexman et al., 2007 ; Kenett and Faust, 2019 ; Popov et al., 2019 ) .	Different attributes, candidate values, exemplars, relations, or conceptual associations may be recruited ( $\Delta_{A},\Delta_{V},\Delta_{R},\Delta_{CF}$ ).
Reasoning	Collaborators can employ different analogies, metaphors, causal models, levels of abstraction, conceptual combinations, or inferential strategies ( Hummel and Holyoak, 1997 ; Thibodeau and Boroditsky, 2011 ; Kalogerakis et al., 2010 ; Ozkan and Dogan, 2013 ; Chi et al., 1981 ; Johnson and Keil, 2000 ) .	Different inferential paths can produce different values, relations, constraints, optimization criteria, or broader framings ( $\Delta_{V},\Delta_{R},\Delta_{C},\Delta_{O},\Delta_{CF}$ ).
Simulation	Collaborators may imagine different users, future situations, contexts of use, or consequences when evaluating an emerging interpretation ( Barsalou, 2009 ; Barsalou and Wiemer-Hastings, 2005 ) .	Different attributes, constraints, consequences, or preferred directions may become relevant ( $\Delta_{A},\Delta_{C},\Delta_{O},\Delta_{W}$ ).
Cognitive effort and metareasoning	Working-memory capacity, depth of deliberation, confidence calibration, perceived fluency, and metacognitive control affect how extensively an interpretation is elaborated and evaluated ( Hammer et al., 2019 ; Unsworth et al., 2014 ; Ackerman and Thompson, 2017 ; Büchel and Wildschut, 2013 ; Topolinski and Reber, 2010 ; Krämer et al., 2023 ; Reason, 1992 ; Kahneman, 2011 ) .	One collaborator may construct a more elaborated or internally constrained frame than another, producing differences across multiple components.

Table 2. Sources of divergence in conceptualization processes. Even given similar inputs, collaborators can construct different interpretations because the processes that select, retrieve, organize, and evaluate information can vary.

Source	Dimensions that can vary	Possible frame differences
Concept properties	Concepts differ in concreteness, abstraction, context availability, situational systematicity, semantic diversity, associative structure, semantic richness, and hierarchical position ( Barsalou et al., 2018 ; Borghi et al., 2014 ; Borghi et al., 2017 ; Paivio, 1990 ; Trope and Liberman, 2010 ; Schwanenflugel and Shoben, 1983 ; Davis et al., 2020 ; Barsalou and Wiemer-Hastings, 2005 ; Hoffman, 2016 ; Lakhzoum et al., 2021 ; Recchia and Jones, 2012 ; Tanaka and Taylor, 1991 ; Chi et al., 1981 ) .	Some concepts constrain interpretation relatively tightly, while others permit broader differences in attributes, values, relations, and framing ( $\Delta_{A},\Delta_{V},\Delta_{R},\Delta_{CF}$ ).
Attribute structure	People vary in which attributes they associate with a concept and which properties they consider central, diagnostic, ideal, or causally explanatory ( Barsalou, 1981 ; Barsalou, 1993 ; Rosch and Mervis, 1975 ; Barsalou, 1985 ; Murphy and Medin, 1985 ; Johnson and Keil, 2000 ) .	Collaborators may include or omit different attributes ( $\Delta_{A}$ ) or assign them different relative importance ( $\Delta_{W}$ ).
Values and acceptable ranges	Defaults, prototypes, ideals, thresholds, and acceptable ranges vary with experience, context, community, and culture ( Rosch and Mervis, 1975 ; Barsalou, 1981 ; Barsalou, 1985 ; Markus and Kitayama, 1991 ; Medin and Atran, 2004 ; Norenzayan et al., 2002 ) .	Shared attributes may be instantiated differently ( $\Delta_{V}$ ), or collaborators may disagree about which values satisfy relevant constraints ( $\Delta_{C}$ ).
Relations and constraints	People differ in causal theories, assumptions about which properties co-vary, and beliefs about which combinations are possible, permissible, or necessary ( Murphy and Medin, 1985 ; Gopnik and Meltzoff, 1997 ; Gopnik and Wellman, 2012 ; Johnson and Keil, 2000 ) .	Frames may encode different dependencies ( $\Delta_{R}$ ) or constraints ( $\Delta_{C}$ ), causing changes to propagate differently through each frame.
Priorities and optimization	When desirable properties compete, collaborators can differ in which outcomes they optimize and which trade-offs they accept ( Keeney and Raiffa, 1993 ; Payne et al., 1993 ; Barsalou, 1983 ; Barsalou, 1985 ) .	Collaborators may pursue different preferred directions ( $\Delta_{O}$ ) or assign different importance to otherwise shared concerns ( $\Delta_{W}$ ).
Broader conceptual framing	The same task can be situated within different conceptual fields, abstraction levels, analogical frames, or culturally learned systems of meaning ( Barsalou, 2012 ; Yeh and Barsalou, 2006 ; Trope and Liberman, 2010 ; Kövecses, 2005 ; Barsalou, 2023 ) .	Different higher-level framings ( $\Delta_{CF}$ ) can reorganize which attributes, values, relations, constraints, and priorities appear relevant.

Table 3. Sources of divergence in concepts and frame structure. Properties of concepts and their learned organization influence how situated conceptual frames can be instantiated.

Source	Dimensions that can vary	Possible frame differences
Underspecification	Speakers routinely omit information they expect collaborators to recover from shared context and common ground; when this assumed overlap is weak, recipients must infer missing details ( Grice, 1975 ; Clark, 1996 ) .	Different attributes, values, constraints, or priorities may be supplied during reconstruction ( $\Delta_{A},\Delta_{V},\Delta_{C},\Delta_{W}$ ).
Ambiguity and polysemy	The same expression can support multiple interpretations, and contextual cues may be insufficient to determine which sense or conceptual frame was intended ( Klein and Murphy, 2001 ; Clark, 1996 ) .	Collaborators may activate different values, relations, or broader conceptual framings ( $\Delta_{V},\Delta_{R},\Delta_{CF}$ ).
Vagueness and granularity	Expressions can intentionally or unintentionally leave boundaries, thresholds, or levels of precision unspecified ( Austin, 1975 ; Jucker et al., 2003 ) .	Collaborators may adopt different acceptable ranges, levels of abstraction, constraints, or optimization thresholds ( $\Delta_{V},\Delta_{C},\Delta_{O},\Delta_{CF}$ ).
Compression	Rich conceptual representations must be externalized through comparatively sparse linguistic, visual, or symbolic expressions, so only a subset of the underlying frame is communicated ( Grice, 1975 ) .	Recipients may reconstruct omitted structure differently, producing divergence across multiple frame components.
Differences in common ground	Collaborators can differ in what they believe is mutually known, salient, or already established in the interaction ( Clark and Brennan, 1991 ; Clark, 1996 ; Lewis, 1979 ) .	Information one collaborator treats as implicit may be absent from the other’s frame, producing differences in attributes, constraints, values, or framing.
Channel and modality	Different communicative channels provide different affordances for grounding and repair: language can express abstractions efficiently, while visual artifacts, gesture, or direct manipulation can make referents and spatial relations more explicit.	What can be externalized or verified differs across channels, affecting which frame components remain implicit or become jointly inspectable.

Table 4. Sources of divergence introduced through communication. Even when collaborators begin with relatively compatible situated frames, communicating those frames through limited external channels can introduce additional differences. We treat these factors as cross-cutting rather than as a fourth locus in the situated conceptualization model.

Figure 4. The Alignment Game cycle. Evidence from an utterance, action, or artifact can reveal a possible conceptual difference. A collaborator responds with an alignment move intended to diagnose, repair, or validate that difference. The partner’s uptake—acceptance, rejection, counterproposal, or partial enactment—provides new evidence and may update one or both situated frames. The cycle continues until the remaining difference is sufficiently small for the joint task or the repair is abandoned. State diagram of six circular nodes. S1 Misalignment Detected leads via "Initiate Alignment" to S2 Move Proposal, then via "Evaluate" or "Counter" to S3 Respond. From S3, "Accept" leads to S4 Frame Update, which "Verify" leads to S5 Alignment Complete or loops back to S1. S3 can also fall to S6 Breakdown, which returns to S1 via "Counter."

Move	Function	Target / Intended effect	Example
Ground $(e)$	Repair, validate	Situational elements: establishes a shared referent or directs collaborators to the same observable element.	“This icon is the mascot.”
Augment $(e)$	Repair	Situation / context: introduces missing information or resources that should enter conceptualization.	“Here is the brand palette.”
Filter $(e,\mathrm{scope})$	Repair	Situational elements: narrows the relevant evidence or excludes information that should not influence the interpretation.	“Ignore the footer; focus on the title area.”
SpecifyContext $(g,r,c)$	Diagnose, repair, validate	Goals, roles, constraints: makes relevant task context explicit so collaborators reason from compatible conditions.	“Brand compliance is required, and this is for five-year-olds.”
ProvideSchema $(f)$	Repair	Situation–frame coupling: provides an organizing schema through which available information can be interpreted.	“Read this as a Z-layout: title at the top, image in the center, caption below.”
Simulate $(scenario)$	Diagnose, repair	Prospective situation: introduces an imagined condition or context that can reveal additional constraints, goals, or consequences.	“Picture this being viewed on a billboard outdoors at noon.”

Table 5. Situation-level alignment moves. These moves intervene on situational and task information that serves as input to situated conceptualization.

Move	Function	Target / Intended effect	Example
Decompose $(a)$	Repair	Attributes ( $A$ ): splits an overloaded or coarse attribute into more specific sub-attributes that can be reasoned about separately.	“Decompose typography into family, weight, size, and spacing.”
Abstract $(level\,\uparrow/\downarrow)$	Repair	Attributes / values: changes the level of abstraction or granularity at which the concept is being reasoned about.	“Not navy versus teal—think warm versus cool palette.”
Restructure $(h)$	Repair	Frame / hierarchy: reorganizes the structure or ordering of frame elements.	“Shift to a type-first layout; text leads and the image supports.”
AnalogicalMap $(src)$	Probe, repair	Relations / values: transfers relational structure from a familiar source domain to reorganize the current interpretation.	“Lay it out like a metro map: line colors are categories and stops are sections.”
BlendConcepts $(c_{1},c_{2})$	Repair, create	Values / conceptual fields: synthesizes selected structures or values from two conceptual sources.	“Blend magazine minimalism with children’s-book playfulness.”
Elaborate $(detail)$	Repair	Attributes / values: expands an underspecified concept by generating additional detail or distinctions.	“Define what you mean by ‘playful’.”

Table 6. Process-level alignment moves. These moves intervene on reasoning processes used to construct, elaborate, reorganize, or evaluate situated conceptual frames.

Move	Function	Target / Intended effect	Example
ScopeAttribute $(a)$	Repair, expand	Attributes ( $A$ ): introduces an attribute that should be represented as relevant, reducing an attribute-level mismatch.	“Let’s add texture as an attribute.”
PruneAttribute $(a)$	Repair	Attributes ( $A$ ): removes an attribute that is irrelevant, redundant, or unnecessarily represented.	“We don’t need to track border width separately.”
Group $([a])$	Repair	Attributes ( $A$ ): coordinates several attributes as a meaningful conceptual bundle.	“Treat color, type, and icon style together as one playfulness bundle.”
CalibrateMetric $(a,m)$	Repair, validate	Values / measurement: aligns how variation along an attribute is represented or measured.	“Font size here means points, not pixels.”
ProposeValue $(a,v)$	Elicit, repair	Values ( $V$ ): proposes a candidate value for a shared attribute.	“The header color should be deep blue.”
Relate $(a_{i},a_{j})$	Repair	Relations ( $R$ ): establishes or revises a dependency between frame elements.	“Line height should increase when font size increases.”

Table 7. Frame-level alignment moves. These moves intervene directly on components of the situated conceptual frame.

Figure 5. Example Alignment Games. Three rows of comic-style panels, each showing a person and a robot exchanging alignment moves. Row A, a teacher and AI refine what "important" means in a solar-energy quiz question. Row B, an artist and AI blend a jellyfish sketch with a star field. Row C, a writer and AI structure evidence for an essay using a Toulmin argument frame.

Figure 6. Examples of existing systems that expose conceptual frame elements and support conceptual alignment interactions. Collage of six annotated screenshots from prior systems, each named in an orange tab: Spellburst, Stylette, DynEx, ChartEditor, Misty, Intent Tagger, and Brickify. Callouts highlight how each exposes frame elements, such as presenting attributes and values in prompts, decomposing a chart into an editable attribute hierarchy, setting constraints, blending concepts by selecting attributes, and supplying examples to induce concepts.

Explore similar work

Jun 21, 2026cs.RO

A Taxonomy of Conceptual Alignment in Human-Robot Dialogue

Successful conversations require speakers to align on the meaning of concepts, a challenging but crucial task for human-robot interaction. Understanding the process of establishing such alignment is hindered by competing interpretations of the term and isolated, unidirectional investigations of its design space. This paper argues for a design-centric understanding of conceptual alignment as a bidirectional and co-constructive process. We introduce a taxonomy that characterizes conceptual alignment dialogues along what triggers its initiation and what level(s) of conceptual understanding it concerns. We further present a dialogue act schema as an operational tool that captures the interactional moves through which alignment is achieved. Together, these contributions provide a structured foundation for analyzing, comparing, and designing conceptual alignment in human-robot interaction.

Shengchen Zhang, Xiaohua Sun, Weiwei Guo

College of Design and Innovation, Tongji University, Shanghai, China. · School of Design, Southern University of Science and Technology, Shenzhen, Guangdong, China.

Human-Robot Interaction Artificial Intelligence Alignment Dialogue Taxonomy

Ask PDF

Jun 8, 2026cs.LG

A Unifying Framework for Concept-Based Representational Similarity

Learned representations across models and modalities often exhibit striking structural similarities, suggesting shared underlying concept decompositions. However, concept alignment remains poorly defined: existing approaches optimize different objectives under the same terminology, obscuring what is actually aligned. We propose a unifying framework that decomposes alignment along two axes: what is aligned (representations vs. concepts) and at what level (instance-wise vs. distributional). This induces four corresponding properties -- instance-wise and distributional variants of translation and concept consistency -- and reveals precisely which of these guarantees existing methods provide. We further introduce \InterVenchA, an intervention-based benchmark that separately measures extraction quality, translation quality, and concept consistency. Through theory and experiments, we show that commonly assumed equivalences between alignment objectives fail in practice: optimizing one property does not reliably recover the others, and purely unsupervised objectives fail to recover meaningful instance-level alignment. We then propose the Coupled Sparse Autoencoder (CoSAE), which jointly enforces complementary alignment objectives. Strong alignment emerges only in this regime. Surprisingly, as little as 0.1% paired data is sufficient to recover instance-level alignment when anchoring distributional objectives. Overall, our results show that concept alignment is fundamentally multi-objective: it must be defined, measured, and optimized as such.

Grégoire Dhimoïla, Victor Boutin, Agustin Martin Picard +2

Brown University · ENS Paris Saclay · CNRS +2

Representational Similarity Analysis Unified Framework Distributional Information

Ask PDF

Jul 15, 2026cs.AI

Align AI to Dynamic Human-AI Workflows

Current alignment approaches typically focus on emulating human behavior using static representations of human preferences, failing to capture the dynamic, context-dependent nature of real-world human-AI interactions. In this paper, we argue for a shift from static and emulative to interactive and complementary alignment, where preferences emerge through interaction and alignment is defined not by satisfying preferences alone. We first formalize this gap by contrasting existing alignment with a trajectory-level view in which human and model behavior co-evolve over time. Because these interaction dynamics have not been adequately captured within existing ML formulations, we ground this perspective in insights from an interdisciplinary workshop. We draw on lessons from social-science accounts of human-human collaboration and then argue that human-AI systems amplify these dynamics, introducing new asymmetries that make reasoning about uncertainty harder and introduce new coordination challenges. Based on these lessons and new challenges, we conclude by outlining a research agenda for developing AI systems that align with humans in interaction, requiring an interdisciplinary synthesis of machine learning and the social and decision sciences.

Valerie Chen, Cleotilde Gonzalez, Anita Williams Woolley +4

1Carnegie Mellon University · University of California, Irvine

Human-Ai Collaboration Artificial Intelligence Alignment Human-Ai Interaction Interdisciplinary Research Artificial Intelligence Systems Agentic Workflows

Ask PDF