Organizations: Human-Centered Technologies for Learning, Technical University of Munich, Munich, Germany · Munich Center for Machine Learning (MCML), Munich, Germany
AI assistants may carry forward not only observations but also interpretations of what a person is doing or may do next across time. Once retained, these interpretations become context for later predictions and system-triggered outputs. We study whether this representation step changes later behavior in gaze-aware streaming assistance. Holding retrieved events, their order, and the target input fixed, we find that Free-form Interpretation amplifies historical influence in both directions: correct histories produce more helpful changes and wrong histories more harmful changes. Unsupported and broader interpretations about activity or future behavior are associated with greater harmful influence. Evidence-Bounded Interpretation preserves distinctions between observation, inference, uncertainty, and temporal scope, reducing unsupported and overcertain content and lowering wrong-source harmful influence while retaining most helpful influence. The same effect structure appears with a second model. These findings identify persistent interpretation as a consequential design choice for history-aware human-AI systems.
Figures & tables
Figure 1. Illustrative example of persistent interpretation changing downstream influence. The same source-time evidence can be retained as a direct answer, rewritten as a fluent interpretation, or represented with explicit epistemic status before later reuse. The experiments use the original STREAMGAZE task formats. A conceptual teaser showing that fluent persistent interpretations can amplify both helpful and harmful downstream influence, while Evidence-Bounded Interpretation preserves distinctions between observed, inferred, uncertain, and temporally scoped content.
Figure 2. Controlled study setting. For each later target event, earlier source events are retrieved once and then reused across representation conditions. The retrieved events, their order, and the target input are held fixed; only the persistent representation changes. Wrong-source (WS) and correct-source (CS) analyses isolate histories in which all three retrieved direct source predictions are incorrect or correct, respectively. Study setting showing three earlier source events from the same video followed by a later target event. The same retrieved source events branch into Answer Only, Free-form Interpretation, and Evidence-Bounded Interpretation conditions before being used with the same target input. Wrong-source and correct-source populations are evaluated using harmful and helpful downstream flips, respectively.
Figure 3. Changing how the same retrieved history is represented changes its downstream influence. (a) In the frozen wrong-source population ( n=648 ), Free-form Interpretation increases harmful flips relative to Answer Only, while EB removes much of this additional harmful influence. (b) In the frozen correct-source population ( n=714 ), Free-form Interpretation increases helpful flips, while EB retains most of this increase. Exact paired percentage-point differences and video-level clustered bootstrap 95% confidence intervals are reported in the text. Two panels compare downstream flip rates across Answer Only, Free-form Interpretation, and Evidence-Bounded Interpretation. In wrong-source histories, harmful flips rise from Answer Only to Free-form and then fall under EB. In correct-source histories, helpful flips rise from Answer Only to Free-form and remain close under EB.
Figure 4. Cross-model robustness of the paired representation effects. Points show paired percentage-point differences and whiskers show video-level clustered bootstrap 95% confidence intervals. Free-form Interpretation increases wrong-source harmful flips and correct-source helpful flips relative to Answer Only for both evaluated models. EB reduces wrong-source harmful influence relative to Free-form for both models, while the GLM correct-source EB–Free contrast is smaller and uncertain. Horizontal forest plot comparing four paired representation effects for Qwen3.5-9B and GLM-4.6V-Flash. Both models show positive Free-form-minus-Answer effects for wrong-source harm and correct-source help, and negative EB-minus-Free effects for wrong-source harm. The GLM correct-source EB-minus-Free confidence interval crosses zero.
Figure 5. Properties of Free-form persistent interpretations associated with harmful downstream changes. (a) Within the frozen wrong-source population, unsupported interpretations have a higher harmful-flip rate than supported interpretations. (b) In the separate 500-case audit, harmful-flip rates are lowest for observable object or attribute statements and higher for more interpretive categories. Points and whiskers show group rates and Wilson 95% confidence intervals. The WS support contrast is reported in the text with a video-level clustered bootstrap interval; the scope endpoint contrast is descriptive. Two horizontal dot-and-whisker panels show harmful-flip rates. The first compares supported and unsupported interpretations within wrong-source histories. The second shows four inference-scope categories, from relatively observable object or attribute statements to prospective human inference.
Figure 6. Changes in annotated persistent-content properties under EB relative to Free-form Interpretation in the paired 500-case audit. Points show paired percentage-point differences and whiskers show video-level clustered bootstrap 95% confidence intervals. Gray labels show the corresponding Free-form-to-EB rates. These measures characterize specific properties of the persistent representations rather than overall representation quality. Horizontal forest plot showing EB-minus-Free-form differences for unsupported claims, contradictory claims, and overcertainty. All three estimates are negative and their confidence intervals remain below zero. Absolute Free-form-to-EB rates are displayed beside each property.
Historical intervention
Recovery
Neutral length-matched edit
11.11%
Delete selected entry
25.64%
Oracle historical correction
41.03%
Table 1. Recovery after intervening on 117 harmful cases. Oracle historical correction uses source-time evidence and source ground truth when needed, without access to future frames, target-time evidence, target labels, or downstream outcomes.
Task
No History
Free-form
Evidence-Bounded
OTP
41.9%
48.8%
48.0%
SR
53.6%
43.1%
50.9%
GSM
53.8%
79.6%
76.7%
FAP
29.8%
41.5%
39.2%
GTA
63.5%
63.8%
65.5%
OAA
69.4%
65.8%
69.0%
Table 2. Full-task accuracy under No History, Free-form Interpretation, and Evidence-Bounded Interpretation.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Population
n
Primary use
Stored source interpretations
4,513
Library-level context
Wrong-source histories
648
Harmful-flip analysis
Correct-source histories
714
Helpful-flip analysis
Detailed paired audit
500
Scope and representation quality
WS support annotation
648
Evidence support within WS
Harmful-case interventions
117
Counterfactual mechanism
Appendix
Table 3. Analysis populations used throughout the study.
Comparison
b
c
WS harm, Free − Answer
50
8
CS help, Free − Answer
74
9
WS harm, EB − Free
5
55
CS help, EB − Free
5
24
Appendix
Table 4. Raw paired discordance counts for the primary representation comparisons. These counts are descriptive only.
Wrong-source harm
Correct-source help
Task
n
Answer
Free
EB
ΔF−A
n
Answer
Free
EB
ΔF−A
OTP
118
4.2%
14.4%
5.9%
+10.17
124
45.2%
52.4%
50.0%
+7.26
SR
96
8.3%
11.5%
7.3%
+3.12
109
45.0%
55.0%
52.3%
+10.09
GSM
112
8.0%
15.2%
8.9%
+7.14
131
36.6%
44.3%
40.5%
+7.63
FAP
103
7.8%
15.5%
5.8%
+7.77
115
37.4%
47.0%
44.3%
+9.57
GTA
121
9.1%
10.7%
3.3%
+1.65
118
44.1%
59.3%
55.9%
+15.25
Appendix
Table 5. Detailed task-level paired results under the final frozen WS/CS analysis. Differences are reported in percentage points.
Intervention
Recovery
Neutral length-matched edit
11.11%
Delete selected entry
25.64%
Oracle historical correction
41.03%
Appendix
Table 6. Recovery rates under counterfactual interventions on 117 harmful cases.
Comparison
Δ (pp)
95% CI
Oracle − Delete
+15.38
[+8.55, +22.22]
Oracle − Neutral
+29.91
[+21.37, +39.32]
Delete − Neutral
+14.53
[+7.69, +22.22]
Appendix
Table 7. Paired differences between counterfactual interventions.
Figure 7. Paired proactive-task effects of EB relative to Free-form Interpretation. Circles denote Hurt and diamonds denote Gain. Points show percentage-point differences and whiskers show video-level clustered bootstrap 95% confidence intervals. EB reduces Hurt on both GTA and OAA, while the corresponding Gain differences are smaller and their intervals include zero. These are benchmark model-output metrics rather than measures of user harm or benefit. Horizontal forest plot showing EB-minus-Free-form effects for Hurt and Gain on GTA and OAA. The Hurt estimates for both tasks are negative with confidence intervals below zero. Both Gain confidence intervals cross zero.
Figure 8. Illustrative case of harmful propagation and the EB response. The same three earlier source events are retained under each representation condition. The later target prediction is correct under Answer Only, changes to an incorrect prediction under Free-form Interpretation, and is correct again under EB. The Free-form history extends the earlier evidence into a fluent prospective interpretation, whereas EB keeps observation, inference, uncertainty, and temporal scope distinguishable. This example is illustrative and is not used as quantitative evidence. A qualitative appendix figure showing source evidence, a Free-form persistent interpretation, an EB persistent interpretation, and their later downstream consequences.
Current LLM assistants are powerful at answering questions, but they have limited access to the behavioral context that reveals when and where a user is struggling. We present a gaze-grounded multimodal LLM assistant that uses egocentric video with gaze overlays to identify likely points of difficulty and target follow-up retrospective assistance. We instantiate this vision in a controlled study (n=36) comparing the gaze-aware AI assistant to a text-only LLM assistant. Compared to a conventional LLM assistant, the gaze-aware assistant was rated as significantly more accurate and personalized in its assessments of users' reading behavior and significantly improved people's ability to recall information. Users spoke significantly fewer words with the gaze-aware assistant, indicating more efficient interactions. Qualitative results underscored both perceived benefits in comprehension and challenges when interpretations of gaze behaviors were inaccurate. Our findings suggest that gaze-aware LLM assistants can reason about cognitive needs to improve cognitive outcomes of users.
Valdemar Danry, Javier Hernandez, Andrew Wilson +2
Microsoft Research Cambridge, MA, USA · MIT Media Lab Cambridge, MA, USA · Microsoft Research Cambridge, Massachusetts, USA
AI assistants in human-AI collaboration often correct suboptimal human actions through behavioral feedback (e.g., alerts or steering-wheel nudges in assistive driving). Such interventions can mitigate immediate errors, but long-term improvement requires addressing the underlying misconceptions that cause repeated mistakes. We introduce SENSEI, a framework that infers user misconceptions from interaction behavior and provides targeted, minimal yet sufficient suggestions to correct them. Our approach departs from action- or trajectory-level interventions by operating over a structured knowledge representation to localize and correct the sources of erroneous behavior. Across three long-horizon tasks with diverse misconceptions and corresponding behaviors, SENSEI demonstrates zero-shot compositional generalization, disentangling multiple overlapping misconceptions despite training only on single-misconception cases. A user study further shows that our method identifies real human misconceptions and provides effective guidance that improves long-horizon task performance, successfully correcting 90% of student misconceptions. Code and project page are available at https://misoshiruseijin.github.io/SENSEI/.
Proactive Smart Assistants (PSAs) increasingly use general knowledge, learned patterns from group-level data, patterns from individual user's data, and recent sensor observations to anticipate which action a user may prefer. Existing systems typically select actions from predicted user preference likelihoods, but these likelihoods alone do not reveal how reliable the prediction is, which evidence supports it, or why ambiguity remains. In this vision paper, we argue that uncertainty estimation can provide this missing information and substantially change appropriate downstream behavior. Using the example of a home assistant, we introduce the questions that uncertainty estimation aims to answer, i.e. how sure is the assistant about the predicted preference likelihood, whether the confidence come from group-level assumptions or individual's data, and whether likelihood ambiguity reflect the user's intrinsic preference variability or insufficient data for the particular situation. We further introduce uncertainty estimation applications in neighboring domains, such as human trajectory prediction, personalized activity recognition, information-collection, and model-based control. Finally, we draw on insights from neighboring domains to envision how future uncertainty-aware proactive assistants may address key challenges in proactive interaction.
Zhiyu An, Ziyi Xuan, Yu Yang +1
University of California, Merced Merced, California, USA · Lehigh University Bethlehem, Pennsylvania USA