PaperlayerView exact revision on arXiv

Source-linked AI summary

UniPath: Adaptive Coordination of Understanding and Generation for Unified Multimodal Reasoning

Hayes Bai, Yinyi Luo, Wenwen Wang, Qingsong Wen, Jindong Wang

arXiv:2605.11400v1cs.MM

TL;DR

Different inputs may benefit from different ways of coordinating understanding and generation, but how to represent and select those paths remains a challenge. UniPath uses a path-conditioned executor and input-dependent planner; experiments report improved accuracy with lower token cost and interpretable reasoning traces.

  • Problem

    How to represent multiple understanding-generation coordination patterns and select one per input remains a challenge, as different inputs may benefit from different paths.

    Source: §1 · p. 2, §3.1 · p. 4

  • Method

    UniPath models reasoning as path selection and execution, training a role-aligned path-conditioned executor and using a planner to select a path for each input.

    Source: §1 · p. 3, §5 · p. 10

  • Results

    UniPath's planner-executor system improves accuracy with lower token cost while producing interpretable reasoning traces.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The results support treating coordination policy as a first-class component of unified multimodal reasoning.

    Source: §5 · p. 11

    The deployable planner remains far from oracle routing, and robust path-selection generalization across domains remains challenging.

    Source: §5 · p. 11

Abstract

from arXiv · show

Unified multimodal models (UMMs) aim to integrate understanding and generation within a single architecture. However, it remains underexplored how to effectively coordinate these two capabilities for more effective and efficient reasoning. Existing coordination approaches either perform coupling during training, without explicit inference-time coordination, or impose a fixed coordination pattern for all inputs. In this work, we show that multimodal tasks exhibit substantial coordination-path diversity: different inputs favor different coordination paths. This suggests that exploiting such diversity is key to improving performance. We propose UniPath, a framework for adaptively modeling and exploiting coordination-path diversity. Instead of enforcing a single coordination pattern, we represent task solving as the selection and execution of a path, ranging from direct answering to textual inference, visual-thought construction, and hypothesis-based exploration. We construct role-aligned trajectories to train a path-conditioned executor and introduce a lightweight planner mechanism to enable input-dependent path selection. Experiments show that leveraging coordination-path diversity improves performance over fixed coordination strategies while providing interpretable intermediate behaviors. The code is available at:https://github.com/AIFrontierLab/TorchUMM/tree/main/src/umm/post_training/unipath.

1 Introduction

UniPath argues that unified multimodal models need input-dependent coordination because different questions benefit from different reasoning paths. It proposes a planner-executor framework to select and execute those paths, aiming to improve accuracy while reducing token cost and producing interpretable traces.

Motivation and path diversity

  • Suitable paths can reserve deeper multimodal reasoning for inputs that need it, avoiding overlooked problems, unnecessary tokens, wasted computation, and added error risk.The paper also motivates path selection as a way to provide a readable explanation of the chosen solving strategy.Source: §1 · p. 1

Prior coordination approaches

Motivation and path diversity

  • The best coordination path varies by subject and instance, and selecting the best path per input substantially outperforms any fixed path on MMMU.Correctness varies sharply across paths, with many inputs solved by only a subset of paths.Source: §1 · p. 2, Figure 1 · p. 2

Challenges and framework

  • Practical adaptive coordination requires categorizing roles, training one model to execute distinct paths with suitable visual supervision, and selecting paths accurately despite scarce labels and dataset biases.Visual intermediate states may be abstract constructions or hypotheses, so supervision should not require every step to become a complete image.Source: §1 · pp. 2–3
  • UniPath defines five roles and five representative paths, trains a path-conditioned executor on role-aligned trajectories, and uses a lightweight planner with query-form calibration to select paths per input.For visual roles, aligned visual thought keeps traces readable as text while supervising hidden states of visual-thought spans with visual summaries; the system is reported to improve accuracy with lower token cost and interpretable traces.Source: §1 · p. 3

2 Related Work

Prior unified multimodal models combine multiple capabilities but often leave understanding–generation coordination implicit, while existing coordination methods provide limited input-dependent control. UniPath instead selects a coordination path and executes its corresponding role sequence.

Unified Multimodal Models

  • UMMs span unified-token, diffusion- or flow-based, and efficiency-oriented designs, yet often handle reasoning coordination implicitly, risking inconsistencies between understanding and generation [Luo et al., 2026].These models share the goal of equipping one model with multiple multimodal functionalities.Source: §2 · p. 3

Coordinating Understanding and Generation

  • Training-time coupling [Su et al., 2025; Xie et al., 2025a] and multimodal chain-of-thought [Qin et al., 2025] do not adapt coordination to each input, while inference-time loops [Wu et al., 2026; Huang et al., 2025] use fixed patterns.Training-time approaches do not specify inference coordination; prompting or training formats largely predetermine multimodal chain-of-thought structure.Source: §2 · p. 3
  • UniPath models understanding–generation coordination as selecting a path and executing its role sequence, shifting from a universal protocol to adaptive exploitation.The approach explicitly distinguishes which functional roles are needed for different inputs.Source: §2 · p. 3

3 Methodology

UniPath formulates understanding-generation coordination as input-dependent path selection followed by execution, rather than assuming one fixed pattern. It defines compact role-based paths and trains a path-conditioned UMM executor with a lightweight planner to select among them.

3.1 Problem Formulation

  • Each input x = (q, I) is routed to one coordination path before a path-conditioned UMM executor produces intermediate states and the final output.This formulation provides a unified interface for multiple coordination strategies without assuming a fixed pattern.Source: §3.1 · p. 4

3.2 Coordination Categorization

  • UniPath defines five functional roles—understanding, reasoning, construction, hypothesis, and answer—and uses them to form a compact set of representative paths.Construction creates a visual thought for a later step, while hypothesis maintains candidate visual thoughts for comparison; the role set is not exhaustive.Source: §3.2 · pp. 4–5

3.3 Planner-Executor Framework

  • Role-aligned trajectories supervise path following, while visual-thought spans remain readable text and their hidden states are aligned to visual summaries for subsequent reasoning.The executor is trained with a four-stage LoRA curriculum; role-weighted language modeling supervises textual roles, and a projection head aligns visual-thought states to reference-image summaries.Source: §3.3 · pp. 5–6
  • The planner predicts success for each path as a multi-label task, then query-form calibration adjusts its scores using bucket-specific temperatures and path biases before selection.Samples with fewer successful paths receive greater weight, and positive labels outside the answer-only path are mildly upweighted; selection requires an advantage over a default path to exceed a margin.Source: §3.3 · p. 6

4 Experiments

Experiments report improvements across understanding, generation, and understanding-generation consistency, and analyze how planner routing and path-specific execution affect accuracy and token cost.

4.1 Experimental Setup

  • BAGEL with LoRA adapters executes role-aligned trajectories, while planner supervision records which of five candidate paths solve roughly 8k calibration queries.The executor uses a comparatively smaller training set, and evaluation covers understanding, generation, and consistency benchmarks. Generation and consistency run the executor on the original query, isolating executor training rather than input-dependent routing.Source: §4.1 · pp. 6–7

4.2 Main Results

  • UniPath gains +4.3% on MMMU, +4.4% on MMBench-EN, +7.7% on MMStar, and +0.8% on MathVista over BAGEL; GenEval improves from 78.81 to 80.00 and WISE from 0.3989 to 0.4100.It achieves the best results on MMMU, MMBench-EN, and MMBench-CN among the compared BAGEL-based and unified-reasoning variants, while remaining competitive on MathVista and MMStar. Generation runs without planner routing, so these gains reflect executor training and aligned visual-thought supervision.Source: §4.2 · p. 7, Table 1 · p. 8

4.3 Understanding-Generation Consistency

  • On UnifiedBench, executor training improves the BAGEL backbone from 0.8346 to 0.8380 overall.The consistency diagnostic tests whether information survives a reconstruction loop; the authors attribute the gain to aligned visual-thought supervision carrying useful visual information into image regeneration.Source: §4.3 · p. 8, Table 2 · p. 8

4.4 Planner Analysis

  • Planner path distributions vary by benchmark, and its final version achieves the highest validation utility and routed MMMU accuracy among five versions.MMMU is mostly routed to pC, MMBench-EN and MMStar to pA, and MathVista has a more balanced distribution. Model scores and query-form buckets are each useful but insufficient alone; the planner is better or tied on four of five benchmarks against each and has the best average among deployable planners. Feature-space analysis also identifies domain shift and overlapping path labels as generalization challenges.Source: §4.4 · p. 9, Figure 3 · p. 9, Table 3 · p. 10

4.5 Token-Accuracy Tradeoff

  • Across five understanding benchmarks, UniPath uses substantially fewer output tokens while matching or improving accuracy on most, favoring appropriate paths over long reasoning traces.The comparison reports token reductions relative to IRG and UniCoT; AD-Loop is omitted because per-example outputs for token accounting are unavailable.Source: §4.5 · pp. 9–10, Figure 4 · p. 10

4.6 Qualitative Analysis

  • In a bridge-circuit example, pC alone gives the correct answer by constructing a visual-thought state that exposes the resistor relation and yields R = 6Ω.The example makes path-specific execution visible at the question level; additional examples for other paths are provided in the appendix.Source: §4.6 · p. 10

5 Conclusion and Limitations · Appendix · A Additional Path-Diversity Visualizations

UniPath treats coordination-path selection as a core part of unified multimodal reasoning: inputs favor different paths, while deployable planning remains well short of oracle routing. Full-resolution MMMU visualizations show this path diversity across subjects and individual questions.

5 Conclusion and Limitations

  • UniPath uses a compact path space, a path-following executor, and a query-form calibrated planner to select a path for each input.The executor follows paths through a unified interface.Source: §5 · p. 10
  • Raw planner scores, query-form rules alone, and BAGEL’s own path choice are insufficient, making path selection a central modeling problem.The conclusion frames coordination policy as a first-class component of unified multimodal reasoning.Source: §5 · pp. 10–11
  • The main limitation is that the deployable planner remains far from oracle routing, and robust cross-domain planner generalization remains challenging.The paper points to Appendix J for fuller discussion.Source: §5 · p. 11

A Additional Path-Diversity Visualizations

  • Figures 5 and 6 provide full-resolution versions of Figure 1’s MMMU visualizations, with rows labeled using the role-sequence notation from Section 3.1.The labels map (A) to pA, (U, A) to pU, (U, R, A) to pR, (U, R, C, R, A) to pC, and (U, R, H, R, A) to pH.Source: A Additional Path-Diversity… · p. 15

5 Conclusion and Limitations

  • Different examples favor different paths, and oracle path selection remains substantially above any fixed path, revealing complementarity among coordination patterns.Figure 5 shows differing subject-level affinities and a substantially higher oracle row than any fixed path.Source: §5 · p. 10, Figure 5 · p. 15

A Additional Path-Diversity Visualizations

  • MMMU subject domains favor different coordination paths, while the oracle row remains substantially higher than any fixed path.Figure 5 provides the full subject-level path-affinity visualization.Source: Figure 5 · p. 15
  • Sparse, non-identical path-correctness patterns across MMMU examples show that path complementarity also occurs at the individual-question level.Figure 6 marks correct path-example pairs with colored cells and failures with gray cells.Source: Figure 6 · p. 16

B Training Data Construction

UniPath builds role-aligned training trajectories for different understanding–generation paths using datasets and annotations suited to each path. It also trains a lightweight planner from path outcomes on calibration examples.

B Training Data Construction

  • Perception-only trajectories use VQAv2, while textual-reasoning trajectories pair ScienceQA understanding data with GoT generation data.ScienceQA answer indices are mapped to option text, with lecture and solution fields used as reasoning; GoT prompts specify generation targets and annotations provide reasoning traces.Source: B Training Data Construction · p. 16
  • Visual-thought and hypothesis paths draw on CoMT, CoF-SFT, FLUX-Reason-6M, GoT generation data, and VL-PRM300K, with approximately 5k examples per nontrivial level and source group.Visual-thought examples use visual references or a synthesized partial image; hypothesis examples compare multiple candidate visual hypotheses.Source: B Training Data Construction · p. 16
  • When intermediate annotations are unavailable, GPT-5.4 and Claude Code Opus 4.6 generate role-aligned text segments, and BAGEL synthesizes reference images for visual-thought supervision.Source: B Training Data Construction · p. 16
  • Planner supervision records which of five candidate paths solve roughly 8k calibration queries; the executor uses a comparatively small training set, with gains attributed to coordination rather than corpus size.The training-data scale and sources are summarized in Table 4.Source: B Training Data Construction · p. 16, Table 4 · p. 16

C Benchmark and Metric Details

This section defines how UniPath is evaluated across understanding, generation, and understanding–generation consistency. It specifies benchmark scopes, reported metrics, and what each evaluation protocol measures.

Metric conventions

  • Table 1 reports understanding and GenEval as accuracies in percent, WISE as WiScore, and Avg. as the mean across five understanding benchmarks when all scores are available.Appendix F.1 gives a separate relative-gain comparison with AD-Loop to avoid mixing potentially different backbone and evaluation settings in the main table.Source: C Benchmark and Metric Details · p. 17

Understanding benchmarks

Generation benchmarks

  • GenEval [Ghosh et al., 2023] measures text-to-image object binding and compositional accuracy, while WISE [Niu et al., 2025] assesses broader cross-domain text-to-image alignment.Generation scores use the trained executor under the same prompt protocol and therefore reflect executor training rather than planner routing.Source: C Benchmark and Metric Details · p. 17

Understanding-generation consistency

  • UnifiedBench tests image-information preservation by comparing an image reconstructed from an image-to-text-to-image loop with the original using CLIP, DINOv2, DINOv3, and LongCLIP similarities.It probes executor training and aligned visual-thought supervision because later generation depends on information retained during understanding.Source: C Benchmark and Metric Details · p. 17

D Aligned Visual Thought Analysis · E More Details and Analysis on Planner · E.1 Planner Training Details

Aligned visual thought outperforms explicit latent or image feedback on routed MMMU while reducing runtime. The planner uses path-aware features and multi-label supervision to select among five paths.

D Aligned Visual Thought Analysis

  • The comparison tests readable visual-thought traces against latent feedback and image feedback in fixed pC/pH execution and routed MMMU.Latent feedback replaces the visual-thought step with generated visual latents; image feedback generates an intermediate image for subsequent reasoning.Source: D Aligned Visual Thought Analysis · p. 17
  • Aligned visual thought keeps the intermediate step readable as text while image-derived supervision shapes its hidden representation.Image feedback can be costly and brittle, while pure visual latents can disrupt semantic continuity in text reasoning.Source: D Aligned Visual Thought Analysis · p. 17
  • 4.44 points lower routed MMMU accuracy with latent feedback and 3.44 points lower with image feedback show aligned visual thought performs better.The comparison also finds the same accuracy trend for fixed pC and pH execution.Source: D Aligned Visual Thought Analysis · p. 17, Table 5 · p. 18
  • Aligned visual thought reduces per-sample runtime by 27.3–30.3% versus latent feedback and 24.4–28.6% versus image feedback.The runtime comparison is for fixed pC or pH execution.Source: D Aligned Visual Thought Analysis · p. 17, Table 5 · p. 18

E.1 Planner Training Details

  • The planner scores five paths from a 39,424-dimensional feature vector using a two-hidden-layer MLP with width 768 and five output logits.The vector concatenates an image-summary feature with last-token and mean text features extracted under each candidate path prompt; sigmoid scores are followed by query-form calibration.Source: §E.1 · p. 18
  • Planner supervision is multi-label weighted binary cross-entropy, and samples for which all five paths are incorrect are excluded from training.More than one path may solve the same input.Source: §E.1 · p. 18
  • Training weights positive paths outside pA by 1.3 and samples with one, two, or at least three positive paths by 3.0, 2.0, or 1.0, respectively.AdamW weight decay is 5 × 10−5, with batch size 256 and learning rate 5 × 10−4.Source: §E.1 · p. 18

E.2 Planner Training Transfer … F.2 Path Format Compliance on MMMU

Planner analyses show that path selection remains challenging because global features reflect domain more clearly than path, while bucket-conditioned views reveal more localized structure. On fixed-path MMMU evaluation, the executor generally follows the requested format, pointing to path selection rather than formatting as the larger remaining bottleneck.

E.2 Planner Training Transfer

  • Lower-utility planner variants often collapse to narrow path preferences, while the final planner retains broader routing, supporting learned path selection over query-form calibration alone.Examples of collapse include routing 856/900 cases to pR or 690/900 to pU.Source: §E.2 · p. 18, Figure 7 · p. 18

E.3 Planner Feature-Space Analysis

  • Global planner features cluster more clearly by dataset or domain than by oracle-correct path, while bucket-conditioned views expose more localized path structure.For examples with multiple correct paths, the global visualization randomly samples one correct label; bucket-conditioned panels instead show the planner-selected successful path.Source: §E.3 · p. 19, Figure 8 · p. 19
  • Query-form buckets act as a lightweight inductive bias by partitioning inputs before learned planner scoring, reducing path-selection entanglement without solving selection by themselves.The analysis reports more homogeneous local views in three representative buckets than in the global path view.Source: §E.3 · p. 19

F.1 Relative Gains of BAGEL-based Post-training Methods

  • Table 6 reports relative gains over corresponding BAGEL baselines for RecA, UniGame, UniCoT, Ours, and AD-Loop, with missing entries marking unavailable reports.The first four methods' gains are computed from Table 1; AD-Loop's are based on values reported in its paper.Source: §F.1 · p. 19, Table 6 · p. 19

F.2 Path Format Compliance on MMMU

  • Path-format checks assess required role sections, legal parsed answers, and whether raw answer spans follow the expected answer surface form.Required sections can include Understanding, Reasoning, Visual, Hypothesis, and Answer; legal parsed answers are valid option letters or nonempty open-question answers.Source: §F.2 · p. 20
  • Format checks are high across all fixed-path MMMU paths, indicating that the executor usually follows the requested coordination template.The reported checks cover section structure, parsed-answer validity, and strict answer-text format.Source: Table 7 · p. 20, §F.2 · p. 20
  • The executor follows the requested section structure for nearly all fixed-path MMMU examples and always produces a parsable final answer, making template errors unlikely to dominate downstream failures.The analysis suggests that selecting the right path for each query is a larger remaining bottleneck than emitting the requested format.Source: §F.2 · p. 20

F.3 Different Backbone Experiments … G.1 Executor Data Statistics

With a Harmon-1.5B backbone, the five-path oracle reaches 67.78% on MMMU, while routed gains are smaller. The appendix also reports category-level generation results and details executor training and data.

F.3 Different Backbone Experiments

  • The five-path oracle reaches 67.78% on MMMU, 33.45 points above raw Harmon.The oracle selects the correct answer whenever any of the five path-conditioned executions is correct.Source: §F.3 · p. 20, Table 8 · p. 20
  • The routed system improves over raw Harmon on four of five understanding benchmarks, but gains are much smaller than the oracle gap.The results suggest that converting path complementarity into reliable routed gains is harder with a weaker executor.Source: §F.3 · p. 20, Table 9 · p. 20

F.4 Additional Generation Results

  • Tables 10 and 11 report category-level GenEval and WISE scores for the Table 1 models, tracing overall generation results to their subcategories.Table 10 reports accuracies; Table 11 reports WiScore.Source: §F.4 · p. 21, Table 10 · p. 21, Table 11 · p. 21
  • Algorithm 1 trains the path-conditioned executor in four stages, applying role-weighted text, visual-alignment, and image-latent losses when applicable, then selecting checkpoints by validation and format-preservation criteria.Visual-alignment loss is used for batches with Visual or Hypothesis spans, and image-latent reconstruction loss for batches with final image answers.Source: §F.4 · p. 22

G Executor Training Details

  • The final executor freezes BAGEL-7B-MoT and trains rank-16 LoRA adapters with a visual-thought projection head through four checkpoint-chained stages.The stages cover text understanding, understanding with aligned visual thoughts, generation without aligned visual-thought supervision, and generation with construction or hypothesis supervision. LoRA alpha is 32 and dropout is 0.05.Source: G Executor Training Details · p. 22
  • Training uses batch size 1 and bf16; S1–S2 run one epoch, while S3–S4 run 3000 steps, with format-accuracy degradation tolerances of 0.01 for S1 and 0.03 for S2–S4.S1/S2 maximum sequence lengths are 3072/6144; S1 uses patience 6 and a 1200-step warmup, while S2–S4 use patience 8 and a 1600-step warmup.Source: G Executor Training Details · p. 22
  • S1/S2 use token-level weights for text loss, whereas S3/S4 use unit target-text weights and a global CE coefficient; CE, Visual, and MSE denote text, visual-thought alignment, and image-latent reconstruction losses.The text, MSE, and visual loss coefficients are λtext, λmse, and λvis; absent losses are set to 0. Format accuracy checks role structure rather than reasoning correctness.Source: G Executor Training Details · pp. 22–23, Table 12 · p. 23

G.1 Executor Data Statistics

  • Table 13 gives post-filter executor-stage split sizes for the four validation splits in Table 14, excluding the separate path-outcome runs used for planner supervision.The supplied passage describes the table’s scope but gives no split counts.Source: §G.1 · p. 23, Table 13 · p. 23

G.2 Executor Training Diagnostics · H Path Prompt Templates

Executor diagnostics favor staged LoRA because format compliance alone does not ensure useful path execution. The prompt templates specify role-ordered outputs after the planner selects a path.

G.2 Executor Training Diagnostics

  • Figure 9 shows CE decreasing on S1/S2, while visual and MSE losses indicate optimization of aligned visual-thought and image-latent objectives where enabled.The figure reports smoothed training losses for the staged LoRA executor.Source: §G.2 · p. 23, Figure 9 · p. 24
  • Staged LoRA offers the best balance of stage-level learning, format preservation, checkpoint size, and downstream path behavior, so it is used in the main experiments.It trains one stage at a time, passing the previous adapter forward.Source: §G.2 · pp. 23–24
  • Raw BAGEL often preserves requested formats, but low S1/S2 answer accuracy and high visual-summary loss indicate that format compliance alone does not ensure path execution.Partial-SFT and multitask training often preserve perfect formatting while improving task accuracy only weakly; Partial-SFT unfreezes part of the language model.Source: §G.2 · pp. 23–24, Table 14 · p. 24

H Path Prompt Templates

  • Execution-time wrappers are prepended for pU, pR, pC, and pH, while pA receives the benchmark query directly; the query follows the wrapper after “Query:”.These prompts specify executor role order and output format after planner path selection; they do not replace the learned selector.Source: H Path Prompt Templates · pp. 24–25
  • The pU path requires one-sentence visual understanding followed by an answer based only on perception, while pR adds a short textual reasoning step before answering.The pU prompt limits the answer to what is explicitly visible or stated; pR reasoning uses only provided information.Source: H Path Prompt Templates · p. 25
  • The pC path inserts a visual construct between two reasoning steps, whereas pH inserts candidate visual hypotheses before reasoning again and answering.Both paths begin with one-sentence understanding and use a final concise answer.Source: H Path Prompt Templates · p. 26

I Additional Qualitative Examples … Broader Impact

The qualitative examples illustrate input-dependent coordination paths and role-aligned trajectories, while identifying planner accuracy and cross-domain calibration as limitations and discussing potential benefits and misuse risks.

I Additional Qualitative Examples

  • Across additional MMMU cases, the correct path varies: direct answering works for the potato question, explicit perception for the velocity graph, reasoning for cisplatin, and hypothesis exploration for the power graph.The examples show that extra decomposition can hurt when the answer is already apparent, while visual anchors, textual reasoning, or comparison of candidate graph shapes can be decisive in other cases.Source: I Additional Qualitative Examples · p. 27

J Role-Aligned Trajectory Examples

  • Staged executor trajectories illustrate direct visual answering, shared-property reasoning, row-focused visual thought, and hypothesis comparison, with tagged textual traces emitted at inference.For pC and pH, visual references supervise aligned visual thoughts during training; inference emits the tagged textual trace rather than inserting those images into context.Source: J Role-Aligned Trajectory Examples · pp. 28–29

Limitations

  • The deployable planner remains far behind oracle path selection, leaving path-space potential unused; improving it requires more accurate instance-level choices and robustness across domains.Path-outcome supervision is costly because it requires running and comparing multiple candidate paths, especially for visual-thought construction and hypothesis exploration.Source: Limitations · p. 29
  • Cross-dataset robustness is limited because query-form buckets are calibrated on auxiliary data, implicitly assuming the target distribution is aligned; dataset-adapted results can improve when that distribution is known.The paper calls for cheaper path-outcome supervision and planner objectives that generalize better under domain shift.Source: Limitations · p. 29

Broader Impact

  • Potential benefits include more efficient reasoning, lower inference cost when a short path suffices, and interpretable path traces that may help developers analyze failures and allocate computation selectively.Source: Broader Impact · p. 29
  • Misuse risks include deceptive synthetic visuals and confident answers from inappropriate paths; proposed safeguards include diverse evaluation, uncertainty or trace reporting, provenance measures, monitoring, access controls, and human oversight for high-stakes uses.The planner's auxiliary-data calibration may also cause performance to vary across domains, languages, or user groups.Source: Broader Impact · pp. 29–30

Related papers

Loading 2605.11400v1…