PaperlayerView exact revision on arXiv

Source-linked AI summary

Benchmarking Behavioral Steerability in Behavior Foundation Models

Minghe Gao, Zhanxi Yan, Jiahui Liu, Wendong Bu, Xiaoting Chen, Qizhou Wang, Yi Su, Siliang Tang, Jun Xiao, Yueting Zhuang, Tat-Seng Chua, Juncheng Li

arXiv:2610.10198v1cs.RO

TL;DR

BFMs can generate behaviors, but reliably steering them to satisfy user intentions remains an open question as they become interfaces for human–robot interaction. RoboSteer defines and benchmarks steerability across three levels using a large multimodal motion corpus. Across nine BFMs, steerability weakens as behavioral complexity increases, particularly for fine-grained control and simultaneous requirements.

  • Problem

    As BFMs become more general human–robot interfaces, it remains unclear whether they can reliably adapt behavior generation to diverse and evolving human intentions.

    Source: §1 · p. 1

  • Method

    RoboSteer defines behavioral steerability and evaluates it through Conditional, Constraint, and Compositional Steering using a multimodal motion corpus.

    Source: §1 · p. 2

  • Results

    Across 9 BFMs, the gap between behavior generation and steerability widens as steering complexity increases, with weaker fine-grained control and difficulty satisfying multiple requirements simultaneously.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Behavior generation alone does not establish intention realization, making explicit evaluation of how reliably models fulfill user intentions important.

    Source: §1 · p. 2, §6 · p. 8

    RoboSteer evaluates single-robot motion from single-subject data, not physical interaction, multi-robot coordination, or long-horizon planning across successive goals.

    Source: §A.3 · p. 13

Abstract

from arXiv · show

Behavior Foundation Models (BFMs) are emerging as a paradigm for translating human intentions into executable humanoid behaviors. As these models evolve beyond behavior generation toward general-purpose behavioral systems, a fundamental question arises: can they be reliably steered according to user intentions? In this paper, we introduce the concept of behavioral steerability, defined as the ability of BFMs to faithfully generate behaviors that satisfy user-specified intentions. To study this capability, we present RoboSteer, the first benchmark for behavioral steerability in BFMs. RoboSteer organizes behavioral steerability into a three-level hierarchy-Conditional Steering, Constraint Steering, and Compositional Steering-and establishes a unified evaluation framework supported by a large-scale multimodal motion corpus. Using RoboSteer, we conduct the first large-scale empirical study of behavioral steerability across 9 existing BFMs. We view behavioral steerability as more than a capability for controlling motion: it concerns how embodied systems translate human intentions into purposeful actions. We hope RoboSteer will advance research on intention realization as a foundation for general-purpose embodied intelligence.

1. Introduction

RoboSteer frames behavioral steerability as the ability to realize user intentions, not merely generate plausible behavior, and evaluates it across increasing levels of complexity. Its evaluation of nine BFMs finds that steerability weakens as complexity rises, especially for fine-grained control and simultaneous requirements.

1. Introduction

  • As BFMs become a more general human–robot interface, reliably adapting behavior generation to diverse and evolving intentions remains an important question.The paper presents intention realization as a potential foundation for more general embodied intelligence.Source: §1 · pp. 1–2
  • The paper draws a precedent from language models: ChatGPT helped shift foundation models from text generation toward steerable interactive systems.The authors argue that BFMs are entering a similar transition from behavior synthesis toward translating human intentions into intended behaviors.Source: §1 · p. 1
  • Behavioral steerability means generating behaviors that faithfully satisfy user intentions, including requirements about conditions, execution constraints, and combinations of instructions.Users often specify how a behavior should be performed, not just which behavior to produce.Source: §1 · p. 2
  • RoboSteer evaluates steerability in three levels: Conditional Steering, Constraint Steering, and Compositional Steering, from single conditions to multiple interleaved requirements.The benchmark also provides an open-source framework for simulation and real-world evaluation.Source: §1 · p. 2
  • The benchmark uses a multimodal humanoid motion corpus of 38,522 source motions, 77,044 human and rendered sequences, 91.58 hours, and 773,716 tasks.The corpus represents joint-level trajectories alongside video, text, audio, and keyframes.Source: §1 · p. 2
  • Across 9 BFMs, the gap between behavior generation and steerability widens with complexity, while fine-grained spatial and body-part control and simultaneous requirements remain difficult.The findings identify intention realization as a distinct capability beyond behavior generation.Source: §1 · p. 2

2. Existing Related Works

Prior BFM research has emphasized generating realistic and executable behaviors, while newer work explores ways to control generation. The field still lacks a unified formulation and evaluation framework for comparing behavioral steerability.

2.1. Behavior Generation as Dominant Paradigm

  • Behavior-generation research spans motion generation, imitation, completion, editing, multimodal modeling, and interaction, with benchmarks focused on quality, fidelity, and task completion.These efforts implicitly treat successful behavior generation as a sufficient measure of BFM capability.Source: §2.1 · p. 3

2.2. Toward Steerable Behavior Systems

  • Recent control approaches include motion editing, instruction-conditioned generation, constraint-aware synthesis, and vision-language-action models, but the field lacks a unified steerability formulation and evaluation framework.These approaches primarily introduce control mechanisms rather than a measurable capability for systematic comparison across BFMs.Source: §2.2 · p. 3

3. Behavioral Steerability

Behavioral steerability extends behavior generation by assessing how well generated behavior realizes the intention conveyed by its input. The paper presents it as a graded capability that becomes more demanding as behavioral specifications grow in complexity.

3. Behavioral Steerability

  • Existing behavior-generation paradigms map behavioral inputs x to resulting behaviors B, often within a control space bounded by learned capabilities and patterns.Freeform inputs do not necessarily enable requirements beyond the behavioral patterns supported by the model.Source: §3 · p. 3
  • Steerability evaluates how well behavior B realizes the intention conveyed by input x, adding intention alignment to the objective of generating executable behavior.The formulation does not require models to explicitly represent or extract intentions; steerability can be quantified through task-specific measures.Source: §3 · p. 3
  • Steerability progresses from a single behavioral condition to explicit execution requirements and then multiple heterogeneous specifications that must be satisfied together.Examples of constraints include movement speed, body-part use, and target locations.Source: §3 · p. 4

4. RoboSteer

RoboSteer is a unified benchmark for behavioral steerability, organized into progressively demanding conditional, constraint, and compositional steering tasks.

4.1. Task Definition

  • Conditional Steering tests behavior realization under complete or partial conditions; Constraint Steering adds one explicit requirement, and Compositional Steering combines interleaved requirements from heterogeneous sources.The tasks span temporal and spatial completion, seven constraint types, and sequential instructions from sources such as video, text, audio, and keyframes.Source: §4.1 · pp. 4–5

4.2. Construction

  • RoboSteer is constructed from web-scale human motion videos using filtering, annotation, reconstruction, retargeting, and simulation-based tracking to retain stably executable motions.The three benchmark levels are instantiated as conditional tasks, corresponding tasks augmented with constraints, and temporally interleaved multimodal instructions.Source: §4.2 · p. 5

4.3. Evaluation Matrix

  • Behavior Steering Score extends Behavior Generation Score with intention realization, which measures whether requirements are achieved during robot execution rather than motion-reference similarity alone.Intention realization uses robot state and environment interaction feedback, with requirements accumulated across the three steering levels.Source: §4.3 · pp. 5–6

4.4. Dataset Statistics

  • RoboSteer covers three steering levels, multiple modalities, seven constraint types, and interleaved multi-source instructions across 77,044 sequences, 91.58 hours, and 773,716 tasks.The paper describes this coverage as broader than existing humanoid behavior benchmarks.Source: §4.4 · p. 6

5. Experiment

Across nine evaluated BFM families, behavior-generation capability does not guarantee behavioral steerability, and steerability declines as steering requirements become more complex. The experiments also identify spatial and body-part control and simultaneous multimodal requirements as persistent challenges.

5.1. Experimental Setup

  • The evaluation retargets human-centric motions to a unified robot-centric space and uses simulation-based tracking to validate physical executability.Nine representative model families are evaluated across their supported conditioning modalities, using official resources where available.Source: §5.1 · p. 7

5.2. Main Results and Analysis

  • Behavior-generation scores remain relatively competitive, but behavioral-steerability scores are consistently lower, especially when additional steering requirements are introduced.The comparison indicates that generating valid, instruction-relevant behavior and modifying it reliably to meet user requirements are distinct capabilities.Source: §5.2 · p. 7
  • Behavioral steerability progressively degrades with complexity, as models struggle to preserve and coordinate multiple constraints during execution.The reported degradation is smaller at lower complexity, where models can translate individual requirements into motion.Source: §5.2 · p. 7
  • Spatial completion scores are lower than temporal completion scores across most models, indicating weaker spatial steerability.The authors hypothesize that motion-model pretraining emphasizes temporal evolution more than whole-body spatial dependencies among joints.Source: §5.2 · p. 7
  • Body Restrain is the most challenging Level 2 constraint, requiring instruction-to-body-part correspondence while maintaining balance.The authors connect this difficulty to limited localized modeling and control of individual joints and body regions.Source: §5.2 · pp. 7–8
  • Current BFMs struggle to satisfy multiple behavioral requirements simultaneously, and language-only specification can bottleneck complex spatial and body-part intentions.The authors propose multimodal, interleaved, and potentially streaming inputs as directions for specifying compositional intentions.Source: §5.2 · p. 8
  • RoboSteer examines temporal and spatial requirements, multiple execution constraints, and interleaved multimodal requirements, yet current BFMs still struggle to translate them into precise, coordinated behavior.The authors describe behavioral steerability as an open problem and a quantitative dimension for future BFMs.Source: §5.2 · p. 8

6. Conclusion

RoboSteer provides a benchmark for studying behavioral steerability in BFMs. Its evaluation shows that behavior generation does not necessarily translate into steerability, positioning intention realization as a starting point for broader embodied-intelligence challenges.

6. Conclusion

  • RoboSteer is a comprehensive benchmark for studying behavioral steerability, and its evaluations show that behavior generation does not necessarily translate into steerability.The authors position steerability as a starting point for broader challenges including interaction, simulation, and closed-loop feedback.Source: §6 · p. 8

A.1. Data Statistics

RoboSteer contains 773,716 benchmark instances across three levels of increasing control complexity. Its motion and media assets span video, audio, text, images, rotation packages, and trajectory images.

A.1. Data Statistics

  • 773,716 benchmark instances are distributed across 20 task families, with 38,522 unique human motion sequences paired with source and rendered-skeleton videos.The paired video assets span 91.58 hours and 6,683,715 frames in total.Source: §A.1 · p. 11, Figure 7 · p. 11
  • Level 1 contains 630,623 instances (81.51%), Level 2 contains 135,329 (17.49%), and Level 3 contains 7,764 (1.00%).The levels comprise 12 Level 1 task families, seven Level 2 behavioral constraints, and Level 3 Interleaved Multi-Source Steering.Source: §A.1 · p. 11, Figure 7 · p. 11
  • Level 3 uses 33,814 ordered modality components: 8,801 Text, 8,774 Audio, 8,475 Video, and 7,764 Image components.Figure 7 also reports unique Level 1 and 2 input assets by modality, counting repeated references to the same source asset once.Source: §A.1 · p. 11, Figure 7 · p. 11
  • Figure 7 reports that source human and rendered skeleton videos are concentrated between 2 and 5 s.The figure summarizes benchmark scale, video durations, task-family and constraint counts, and modality accounting.Source: §A.1 · p. 11, Figure 7 · p. 11

A.2. Dataset Biases

The source motions are concentrated in common daily activities, locomotion, and performance, with much thinner coverage of several other categories. Joint participation, source-video annotations, and task-specific filters further shape the benchmark’s coverage.

A.2. Dataset Biases

  • The 38,522 source motions span 133 action categories, but six frequent categories account for 65.34% and coverage is stronger for daily activities, locomotion, and performance than for several less-represented categories.Figure 8 summarizes category counts, joint participation, body-region motion, and source-video age and gender annotations.Source: §A.2 · p. 12, Figure 8 · p. 13
  • Of 133 motion types, 91 (68.42%) have median active-joint counts of 20–29; the waist has the lowest mean range of motion across every family.Sports, Dance, and Martial arts are dominated by broad multi-joint motion, while Gesture and Performance include more types with 10–19 active degrees of freedom.Source: §A.2 · p. 12
  • Source-video annotations assign 76.25% of subjects to ages 18–39, while male and female labels account for 66.76% and 33.24% of videos, respectively.The corresponding counts are 25,716 male-labeled and 12,806 female-labeled videos.Source: §A.2 · p. 12
  • Task-specific whitelists and filters tailor category selection to each constraint, including audio-content filtering for Rhythm-to-Motion Alignment and visibility requirements for human-video body completion.Direction and Trajectory emphasize root displacement; Speed and Amplitude select actions that can be modified while preserving action identity.Source: §A.2 · p. 13

A.3. Known Limitations

RoboSteer evaluates single-robot motion generation from single-subject motion data, leaving physical interaction, multi-robot coordination, and long-horizon planning outside its explicit scope. Its temporal text and semantic-audio conditions combine a global action summary with context from observed intervals.

A.3. Known Limitations

  • The benchmark does not explicitly evaluate physical interaction with people or objects, coordination among multiple robots, or long-horizon planning across successive goals.Its source collection contains more daily activities and locomotion than Yoga and Gymnastics.Source: §A.3 · p. 13
  • Temporal text and semantic-audio conditions pair a global action summary from the complete sequence with detailed context from observed intervals.Semantic-audio conditions use synthesized speech with specified voices and synthesis parameters.Source: §A.3 · p. 13

B.1. Data Collection

This section describes how RoboSteer constructs filtered steering tasks from web videos, motion, and associated media, aligning text, audio, and visual conditions with target motions. It specifies task inputs and selection rules for Level 1 completion tasks and Level 2 behavioral constraints.

B.1. Data Collection

  • A five-stage pipeline segments and filters web videos, extracts multimodal information and human motion, then optimizes, retargets, and filters motions for physical executability.The stages produce the motion data used to construct the benchmark.Source: §B.1 · pp. 13–14, Figure B.1 · p. 14

B.2. Motion Caption and Semantic-Audio Construction

  • Motion captions use temporally ordered, action-aligned segments, while task-specific polishing and validation preserve body-region omissions and missing-middle context.Caption fields support full-body, upper/lower-body completion, prediction, retrodiction, and interpolation; invalid outputs are regenerated up to three times and then rejected.Source: §B.2 · pp. 14–15

Semantic-Audio Construction

  • Semantic audio verbalizes task conditions without adding actions, combining base actions with modifiers when present and retaining required visual or trajectory assets.Level 1 variants derive speech from their selected text or action summary; Level 2 modifiers specify properties such as speed, direction, restraint, order, or repetition.Source: Semantic-Audio Construction · p. 16

B.3.1. Level 1: Basic Conditional Generation and Local Completion

  • Task construction selects inputs and temporal segments, then links sample identifiers and boundaries to motion assets; Level 1 comprises full-conditioning reproduction, temporal completion, and spatial completion with target constraints.Text-to-motion, video-to-motion, and audio-to-motion each use all 38,522 samples; rhythm alignment filters by motion type and audio content.Source: §B.3 · p. 16, §B.3.1 · p. 17, Full-Conditioning Reproduction · p. 17

Temporal Completion

  • Temporal completion constructs prediction, retrodiction, and interpolation inputs from partial segments, with synchronized motion targets covering the withheld intervals.Interpolation conditions use endpoint context and explicitly mark the missing middle; key-frame conditioning pairs transition images and timestamps with the complete motion target.Source: Temporal Completion · pp. 17–18

Spatial Completion and Target Constraints

  • Spatial completion removes one body region from text or video conditions while retaining the complete motion as ground truth; Target Reaching separates a base action from a local goal.Target Reaching retains 8,193 instances after filtering for clear, visible goals that are separable from the primary action.Source: Spatial Completion and Target… · p. 18

B.3.2. Level 2: Controllable Motion Modification and Composition

  • Level 2 constraint tasks compare generated motion with unmodified references; speed variants select within-type percentile bands, while direction variants filter for straight, forward locomotion before redirecting it left or right.Speed instructions request twice or half the original speed; direction filtering applies explicit trajectory and facing-consistency thresholds.Source: §B.3.2 · pp. 18–19, Speed · p. 19, Direction · p. 19

Body Restrain

  • Body Restrain selects actions with a suitable secondary body region and retains only moderate-activity samples before assigning an arms-or-legs restraint target.Within each motion type, the 35th–65th percentile band is retained so the region is neither nearly stationary nor too active to restrain without affecting the main action.Source: Body Restrain · pp. 19–20
  • Body Restrain asks the model to preserve the primary action while keeping the specified arms or legs still, with Text, human-video, rendered-skeleton-video, and Audio variants.The whitelist maps hand constraints to arms and foot constraints to legs to avoid end-effector targets that are too small for reliable evaluation.Source: Body Restrain · p. 20, Motion-type Semantic Whitelists · p. 23, Table B.4 · p. 23

Amplitude

  • Amplitude selects motion types amenable to spatial scaling, excludes near-limit or unsuitable actions, and retains samples in the 20th–40th and 60th–80th percentile bands.Low-amplitude samples are instructed to use twice their original amplitude and high-amplitude samples half; the task has Text, human-video, rendered-skeleton-video, and Audio variants.Source: Amplitude · p. 20

Trajectory

  • Trajectory tasks filter semantic motion groups for reusable root paths, require paths at least 0.8 m long, then normalize, resample, smooth, and simplify them.Rendered paths are classified into S-shaped, back-and-forth, or clockwise/counterclockwise categories; text inputs remove global path shape and displacement, while targets retain motion and normalized trajectory points.Source: Trajectory · p. 20

Order

  • Order tasks use distinct, contiguous BABEL actions separated by a 0.15–1.65 s transition; Text, Audio, and Video each contain 15,884 instances.Inputs express A then B or do B after doing A; video uses two clips sharing the transition interval, and labels require no language-model annotation.Source: Order · p. 21

Times

  • Times repeats eligible single-action motions two, three, or four times, with continuous root translation; Text, Video, and Audio each contain 9,738 instances.Filtering requires stable standing poses at both endpoints, including limits of 0.25 rad root-orientation difference, 0.08 m root-height difference, 0.30 standing-joint pose difference, and 0.45 full-body pose difference.Source: Times · p. 21

Motion-type Semantic Whitelists

  • Motion-type whitelists define candidate pools for rhythm and Level 2 tasks, but samples must still pass AST-based audio detection or subsequent motion-statistical or geometric filters.Rhythm candidates include actions that may exhibit musical beats, repeated cycles, or performance rhythm; Level 2 groups include trajectory-producing motion types.Source: Motion-type Semantic Whitelists · pp. 21–22, Table B.2 · p. 21, Table B.3 · p. 22

B.3.3. Level 3: Interleaved Multi-Source Steering

  • 7,764 interleaved tasks combine ordered Text, Audio, and Video segments with an internal-boundary Image, targeting the complete source motion.Each task includes at least one Text, Audio, and Video segment; the released set contains 33,814 modality components: 8,801 Text, 8,774 Audio, 8,475 Video, and one Image per task.Source: §B.3.3 · p. 23

B.3.4. Released Task and Asset Organization

The released task and asset organization specifies how benchmark tasks reference multimodal inputs and targets. The evaluation overview separates basic behavior generation from realization of steering requirements, with factors corresponding to the three steering levels.

B.3.4. Released Task and Asset Organization

  • Level 1 and 2 task files store metadata, inputs, and ground truth, while assets are organized by level and shared-resource directories.Task inputs include prompts and modality fields; asset references are relative to the dataset root.Source: §B.3.4 · p. 24
  • Table B.5 summarizes released task families, input conditions, and benchmark-instance counts; its counts refer to task JSON files.Textual instructions may accompany non-text conditions, while the table lists substantive conditioning signals.Source: §B.3.4 · p. 24, Table B.5 · p. 24

C.1. Evaluation Overview

  • The evaluation separates behavior-generation quality (BG) from steering realization, and Level 3 multiplies BG by IR1, IR2, and IR3.Each IR factor measures realization of a requirement introduced at its corresponding steering level.Source: §C.1 · pp. 24–25
  • IR1, IR2, and IR3 measure realization of conditional, constraint, and interleaved multi-source requirements, respectively.IR values are in [0, 1]; factor definitions and aggregation are task-dependent.Source: §C.1 · pp. 24–25

C.2. Behavior Generation Metrics

This section specifies behavior-generation metrics and task-specific intention-realization rules for Conditional, Constraint, and Compositional Steering, covering motion quality, input-condition alignment, requested constraints, body-region stillness while other regions move, and modality-specific requirements interleaved over time.

C.2. Behavior Generation Metrics

  • Behavior-generation evaluation combines learned metrics (FID, Diversity, MM-Distance, Recall@K) with direct motion and rhythm metrics, using frozen evaluators during benchmark evaluation.FID and Diversity use the OMG Motion Encoder; MM-Distance and Recall@K use modality-specific cross-modal evaluators trained for RoboSteer and then frozen with shared parameters across models.Source: §C.2 · p. 25
  • Motions are converted to a unified G1 representation, with temporal processing selected according to each metric’s evaluation objective.FID and Diversity use 60 endpoint-inclusive phase samples; physical-execution metrics retain the original 50 Hz scale, and MPJPE, g-MPJPE, and Evel align generated motion to the effective reference sequence.Source: §C.2 · pp. 25–26
  • Behavior Generation score combines FID for generated–reference consistency with modality-specific MM-Distance for generated–condition consistency.Both component metrics are lower-is-better; the resulting BG score is the common behavior-generation component of Behavior Steering scores at all three steering levels.Source: §C.2 · p. 26

C.3. Level 1: Conditional Steering

  • Conditional Steering defines intention realization by task: full conditioning uses BG, while completion and sparse-pose tasks use temporal, joint-range, or pose matching.Temporal Completion uses tIoU, which scores identical generated and reference temporal extents as one; Spatial Completion evaluates only the body region missing from the input using JRO.Source: §C.3 · pp. 26–28

C.4. Level 2: Constraint Steering

  • Constraint Steering macro-averages binary satisfaction across seven families: Speed, Amplitude, Direction, Order, Times, Trajectory, and Body Restrain.Level 2 evaluates whether an added constraint is satisfied while preserving the underlying Level 1 behavior.Source: §C.4 · p. 28
  • Order and Times constraints are evaluated from action sequences recognized by a fixed VideoLLM, requiring exact sequence order or the requested action count.The evaluator receives only rendered motion video and a fixed action vocabulary, not the target order or repetition count; the same recognized sequence serves both evaluations.Source: §C.4 · pp. 29–30
  • Trajectory constraints require the generated root path to satisfy the requested spatial pattern, while Body Restrain checks that a specified region stays stationary as the rest moves.Trajectory paths are smoothed before geometric checks; Body Restrain uses root-aligned link positions to remove global root motion before computing velocities at 50 Hz.Source: §C.4 · pp. 30–31
  • Constraint satisfaction requires restricted-link RMS velocity ≤0.10 m/s and remaining-link RMS velocity ≥0.12 m/s, pairing stillness with continued motion.The root is excluded; arm and leg restrictions use different link sets, and seven family-level scores measure reliability in following constraints while preserving behavior.Source: §C.4 · p. 31

C.5. Level 3: Compositional Steering

  • Compositional Steering duration-weights segment realization scores and combines behavior grounding with IR1, IR2, and IR3 to assess temporally interleaved requirements.This targets coherent behavior while switching requirements, separately from lower-level completion and explicit-constraint failures.Source: §C.5 · p. 31

D. Experiment Detail

The experiment detail standardizes heterogeneous conditions and model outputs for evaluation, while adapting each baseline through its native interface. It documents preprocessing, inference, and conversion into a shared G1-based motion representation.

D.1. Data Preprocessing

  • A hierarchical preprocessing pipeline converts text, audio, and image conditions into unified inputs, normalizing instructions while retaining task-relevant actions and constraints.Rule-based checks and iterative refinement validate format, length, temporal consistency, and task-specific semantics; unrecoverable samples are excluded.Source: §D.1 · p. 32

D.2. Model Adaptation and Inference Protocol

  • Baselines retain their native interfaces, inference procedures, and output parameterizations; adaptations address those interfaces without modifying pretrained weights.Failed samples are recorded by failure type rather than silently replaced, so statistics reflect each baseline’s runnable subset.Source: §D.2 · pp. 32–33

D.3. Post-processing and Unified Motion Representation

  • All evaluated outputs are converted to a shared 29-DoF G1-based SONIC representation, with model-specific processing and validation before evaluation.Human-motion outputs are retargeted and resampled; native simulator and robot outputs use corresponding mappings or preserve compatible states.Source: §D.3 · pp. 33–34

E. Detailed Experimental Supplement

The supplement presents task-level result tables for Level 1 conditional steering and Level 2 constraint steering.

Level 1: Conditional Steering

  • Table B.7 reports Level 1 conditional-steering results as BG/BS1 for applicable tasks, with video entries weighted by evaluated sample counts.The table caption does not state specific result values.Source: Table B.7 · p. 35

Level 2: Constraint Steering

Related papers

Loading 2610.10198v1…