Test-Time Compute for Humanoid Behavior Foundation Models

Weiji Xie; Junsong Wu; Zibo Zhou; Yueru Chen; Xingyi Wang; Zhenhua Xiong; Weinan Zhang; Chenjia Bai

cs.ROOct 10, 2026

Test-Time Compute for Humanoid Behavior Foundation Models

Authors: Weiji Xie, Junsong Wu, Zibo Zhou, Yueru Chen, Xingyi Wang, Zhenhua Xiong, Weinan Zhang, Chenjia Bai

Organizations: Institute of Artificial Intelligence (TeleAI), China Telecom · Shanghai Jiaotong University · Zhejiang University · Gamma Robotics

Abstract

Humanoid Behavior Foundation Models (BFMs) can execute impressive motions, but downstream applications such as loco-manipulation demand high whole-body positional accuracy. Existing BFMs suffer from a systematic command-execution gap, where executed motions deviate from desired ones. We show that additional test-time compute can improve execution without retraining the BFM. First, we introduce Test-Time Command Optimization (TTCO), which adapts the input command to the specific BFM to produce an executed trajectory closer to the desired motion. TTCO finds such commands by sampling-based optimization through closed-loop simulations of the BFM and robot dynamics. Across three BFMs on the AMASS dataset, TTCO reduces mean global body-position error by 79-84% in simulation. Second, we analyze test-time compute scaling and derive an approximate compute-optimal strategy that can sustain performance gains as the compute budget increases. Third, TTCO can generate BFM-specific action labels from recorded robot trajectories without requiring the original teleoperation body-motion commands. Simulation and real-robot experiments show that these labels are effective for training high-level loco-manipulation policies. This method also supports reusing existing demonstrations across BFMs.

Figures & tables

Figure 1: Changing the command can bring execution closer to the same reference. Directly using the reference as the command can leave a command–execution gap. With the BFM fixed, our command optimization adjusts commands to bring execution closer to the same reference. The adjustment of commands is exaggerated for demonstration.

Figure 2: Scaling test-time compute with an approximate compute-optimal strategy. Our proposed strategy achieves the best overall performance among four strategies.

Figure 3: Overview of test-time command optimization. We evaluate $N$ candidates through parallel $H$ -step rollouts from the same state and update the proposal using their local costs. After $I$ iterations, we execute the first $E$ steps of the lowest-cost candidate in simulation, then replan.

BFM	Mean cost $\downarrow$	Cost-win rate (%) $\uparrow$	Body-position error (m) $\downarrow$	Root-position error (m) $\downarrow$	Joint acceleration (rad/s 2 ) $\downarrow$
HEFT	$0.419\!\rightarrow\!0.046$	62.27	$0.271\!\rightarrow\!0.057$	$0.264\!\rightarrow\!0.048$	$10.104\!\rightarrow\!17.246$
SONIC	$6.047\!\rightarrow\!0.323$	86.00	$0.615\!\rightarrow\!0.127$	$0.609\!\rightarrow\!0.117$	$\phantom{0}8.832\!\rightarrow\!11.999$
TWIST2	$5.321\!\rightarrow\!0.573$	99.67	$1.095\!\rightarrow\!0.173$	$1.079\!\rightarrow\!0.164$	$43.280\!\rightarrow\!51.574$

Table 1: TTCO reduces mean cost and global position error at the price of higher joint acceleration, across three BFMs on AMASS. Each entry reports nominal

\rightarrow

optimized performance, averaged over all motions per BFM.

Figure 4: Real-robot demonstration of motion tracking with nominal and optimized commands. We compare nominal and optimized commands for left sidestep and backward walking executed open-loop on the real robot. In both examples, optimized commands bring the final displayed positions closer to the reference. Blue outlines show the reference motion.

Figure 5: Compute allocation governs performance through solver and time horizon parameters. We examine these two parameter groups separately. (a,b) Solver parameters

(\mathcal{A},I,N)

. (a) Each point reports the lowest measured cost over all tested

(I,N)

configurations for that algorithm at that budget. CEM achieves competitive performance overall, so we restrict the remaining analysis to it. (b) At fixed

O=IN

, its cost first falls and then rises as iterations increase and candidates per iteration decrease; the best measured configuration changes with the budget. (c,d) Time horizon parameters

(H,E)

. (c) The best measured cost over

E

levels off beyond a horizon ratio

K_{h}=H/E

of 4. (d) Increasing

E

can either lower or raise cost depending on

K_{h}

, motivating joint selection of

H

and

E

. In (b,d), model curves follow the measurements, supporting the proposed forms of the solver and horizon error terms. Error bars show the standard deviation of the three seed means; horizontal dashed lines mark nominal-command cost.

Figure 6: The proposed strategy generalizes to unseen motions and BFMs. Our strategy stays close to the empirical optimum and achieves the lowest average regret among the four strategies across three BFMs and five compute budgets. Faint lines and bands show median costs and 25th–75th percentiles across random configurations; nominal-command costs are labeled above the panels. Costs are averaged over five motions and three seeds; error bars show standard deviations of seed means. Each panel uses logarithmic axes.

Figure 7: Illustration of real-robot loco-manipulation tasks.

SONIC $\rightarrow$ SONIC
Labels	Boxing	Open Door	Pick & Place Box	Sit Sofa	Vision Navigation	Mean
Original	$83.9_{\pm 1.0}$	$84.4_{\pm 1.9}$	$61.1_{\pm 1.0}$	$88.9_{\pm 1.0}$	$60.6_{\pm 1.0}$	$75.8_{\pm 0.5}$
Ours	$63.9_{\pm 1.0}$	$51.1_{\pm 1.0}$	$54.4_{\pm 1.0}$	$66.7_{\pm 1.7}$	$47.2_{\pm 1.0}$	$56.7_{\pm 0.3}$
State	$0.0_{\pm 0.0}$	$0.0_{\pm 0.0}$	$0.0_{\pm 0.0}$	$0.0_{\pm 0.0}$	$0.0_{\pm 0.0}$	$0.0_{\pm 0.0}$
SONIC $\rightarrow$ TWIST2
Original	$48.9_{\pm 1.9}$	$0.0_{\pm 0.0}$	$0.0_{\pm 0.0}$	$17.8_{\pm 1.9}$	$11.7_{\pm 1.7}$	$15.7_{\pm 0.6}$

Table 2: Simulation task success rates (%) on HumanoidArena. Each arrow points from the source BFM to the target BFM. Results are reported as mean

\pm

sample standard deviation across three training seeds, with 60 evaluation rollouts per task and seed.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Motion-level tracking performance on AMASS. Empirical cumulative distributions of optimized-to-nominal global body-position error (top) and full-sequence cost per time step (bottom) across 12,273 motions per BFM. Ratios below one indicate improvement. Horizontal axes are logarithmic; reference lines mark a ratio of one and the 50th percentile. Panels report win rates and median percentage reductions.

BFM	Metric	Median nominal $\to$ optimized	Median reduction (%)	Win (%)	Increase $P_{95}$
HEFT	Body-position error (cm)	$13.88\to 4.80$	62.65	93.86	1.20
	Cost	$0.0529\to 0.0367$	24.59	62.27	0.0148
SONIC	Body-position error (cm)	$11.51\to 4.86$	54.04	90.30	1.09
	Cost	$0.0677\to 0.0416$	30.35	86.00	0.0054
TWIST2	Body-position error (cm)	$62.68\to 13.15$	77.43	95.97	23.76
	Cost	$1.8983\to 0.4709$	73.23	99.67	6.4616

Appendix

Table 4: Motion-level reductions and increases in tracking error and cost. Global body-position error and full-sequence cost per time step for 12,273 motions per BFM. Median reduction is the median of the per-motion percentage reductions,

100(1-X_{m}^{\mathrm{opt}}/X_{m}^{\mathrm{nom}})

, for metric

X

. Win rate is the percentage of motions with a lower error or cost after optimization.

P_{95}

is the 95th percentile of

X_{m}^{\mathrm{opt}}-X_{m}^{\mathrm{nom}}

among motions with increased error or cost, in the metric’s units.

				Contribution to total (%)
BFM	Group	Nominal error range (cm)	Median error reduction (cm)	Error reduction	Error increase	Cost reduction
HEFT	Q1	1.89–6.66	0.57	0.86	78.53	0.01
	Q2	6.66–13.88	5.11	6.00	21.33	0.26
	Q3	13.88–27.78	13.78	16.60	0.14	2.92
	Q4	27.79–843.14	38.04	76.53	0.00	96.81
SONIC	Q1	1.89–5.63	0.22	0.23	43.60	0.02

Appendix

Table 5: Tracking performance grouped by nominal body-position error. Quartiles run from lowest (Q1) to highest (Q4) nominal global body-position error within each BFM. Error reduction is nominal minus optimized error. For each contribution column, we sum only changes in the indicated direction: error decreases, error increases, or cost decreases. We divide each group sum by the corresponding total over all motions of the same BFM and express the result as a percentage.

Source sequence	Frames	Duration (s)
CMU/63/63_01_poses	200	3.98
BMLmovi/Subject_22_F_MoSh/Subject_22_F_3_poses	159	3.16
KIT/883/wipe_leg_bigcircle03_poses	267	5.32
BMLmovi/Subject_45_F_MoSh/Subject_45_F_19_poses	199	3.96
BioMotionLab_NTroje/rub061/0014_knocking2_poses	507	10.12
CMU/102/102_20_poses	265	5.28

Appendix

Table 6: Reference motions used in the compute-allocation and strategy experiments.

Optimizer	Settings
CEM	Initial standard deviation $1.6$ ; minimum standard deviation $0.075$ ; $\max(1,\lfloor 0.1N\rfloor)$ elites; exploration fraction $0$ .
PS	Gaussian noise scale $1.2$ .
MPPI-CMA	Initial noise scale $1.6$ ; minimum noise scale $0.075$ ; temperature $0.1$ ; covariance adaptation rate $0.1$ . Covariance eigenvalues are floored at $0.075^{2}$ .
CMA-ES	Separable CMA-ES; initial scale $1.6$ ; minimum scale $0.05$ ; maximum scale $2.0$ .
CBO	Initial noise scale $1.0$ ; temperature $0.1$ ; consensus weight $1.0$ ; noise weight $0.3$ ; step size $0.1$ .

Appendix

Table 7: Fixed optimizer settings in the solver sweep. Noise scales apply to the normalized knot coordinates before clipping and physical scaling. The CEM settings also apply to the horizon and strategy experiments.

Solver parameter	Value	Horizon parameter	Value
$c$	$0.036137$	$J_{\infty}$	$0.035463$
$a$	$0.44596$	$C_{H}$	$1.0288$
$b$	$0.58164$	$\alpha_{H}$	$0.10719$
$u$	$0.19475$	$C_{E}$	$0.00042258$
$v$	$0.35423$
$g$	$3.9471$

Appendix

Table 8: Parameters of the separate component fits in Fig. 5 (b,d), rounded to five significant figures for display.

$(J_{\infty},C_{H},\alpha_{H},C_{E})$	$(C_{I},C_{N},\alpha_{I},\alpha_{N})$
$(0.02,0.5,0.10,0.0001)$	$(0.03,0.10,0.10,0.30)$
$(0.02,2,0.20,0.0005)$	$(0.03,0.25,0.30,0.60)$
$(0.035,0.5,0.10,0.0001)$	$(0.10,0.10,0.60,0.30)$
$(0.035,2,0.20,0.0005)$	$(0.10,0.25,0.30,0.60)$
$(0.04,1,0.30,0.001)$	$(0.25,0.25,0.60,0.60)$

Appendix

Table 9: Initial values for the horizon fit and the solver coordinates of the joint fit. Each column lists five values in the stated parameter order.

Parameter	Bounds	Fitted value
$J_{\infty}$	$[10^{-8},1]$	0.034123
$C_{H}$	$[10^{-8},20]$	1.8011
$\alpha_{H}$	$[10^{-4},2]$	0.094838
$C_{E}$	$[10^{-8},0.1]$	0.00047245
$C_{I}$	$[10^{-8},2]$	0.49438
$C_{N}$	$[10^{-8},2]$	0.83919

Appendix

Table 10: Parameter bounds and fitted values of the complete CEM surrogate. Fitted values are rounded to five significant figures.

Budget $K$	$I$	$N$	$H$	$E$	Actual $INH/E$	$\widehat{J}$
256	8	32	40	40	256	0.11652350
512	8	32	36	18	512	0.07089407
1024	9	31	33	9	1023	0.05898193
2048	10	48	34	8	2040	0.05236377
4096	13	54	35	6	4095	0.04772997
8192	15	78	35	5	8190	0.04439349

Appendix

Table 11: CEM allocations selected by the surrogate cost model for Sec. 4.3 .

$K$	Random forest	RBF	Gaussian process
256	$(8,32,35,35)$	$(128,2,40,40)$	$(8,32,40,40)$
512	$(8,62,35,34)$	$(256,2,40,40)$	$(8,32,80,40)$
1024	$(14,71,35,34)$	$(512,2,40,40)$	$(14,37,79,40)$
2048	$(15,78,35,20)$	$(512,2,40,40)$	$(6,46,191,26)$
4096	$(18,130,35,20)$	$(512,2,40,40)$	$(7,73,191,24)$
8192	$(30,140,39,20)$	$(512,2,40,40)$	$(16,105,39,8)$

Appendix

Table 12: Configurations

(I,N,H,E)

selected by each regression baseline. The same configuration applies to both motion sets and all BFMs at a given budget. The new-motion evaluation uses the first five budgets.

$K$	Ours	Random forest	RBF	Gaussian process	Empirical Optimum
Ten fitting motions, SONIC
256	$0.1048\pm 0.0063$	$0.1226\pm 0.0310$	$0.2096\pm 0.0260$	$0.1053\pm 0.0068$	$0.1027\pm 0.0028$
512	$0.0715\pm 0.0034$	$0.0923\pm 0.0073$	$0.1554\pm 0.0360$	$0.0855\pm 0.0140$	$0.0715\pm 0.0034$
1024	$0.0607\pm 0.0052$	$0.0783\pm 0.0141$	$0.1461\pm 0.0350$	$0.0799\pm 0.0065$	$0.0570\pm 0.0017$
2048	$0.0557\pm 0.0021$	$0.0585\pm 0.0093$	$0.1461\pm 0.0350$	$0.0868\pm 0.0110$	$0.0528\pm 0.0049$
4096	$0.0484\pm 0.0010$	$0.0520\pm 0.0049$	$0.1461\pm 0.0350$	$0.0741\pm 0.0015$	$0.0461\pm 0.0028$

Appendix

Table 13: Measured costs on the ten SONIC fitting motions at matched compute budgets; lower is better. Entries report mean cost

\pm

one sample standard deviation across the three seed means. Empirical Optimum uses the configuration with the lowest mean cost among the four strategies and 200 random configurations.

Figure 1 : ten fitting motions
BFM	$K$	Ours	Random forest	RBF	Gaussian process
SONIC	256	2.10	19.36	104.13	2.60
	512	0.00	29.08	117.26	19.52
	1024	6.40	37.27	156.16	40.12
	2048	5.38	10.65	176.43	64.25
	4096	5.15	12.90	217.10	60.90

Appendix

Table 14: Allocation regret (%) on the ten SONIC fitting motions. The reference is Empirical Optimum in Table 13 . The Mean row weights the six budgets equally. Bold marks the lowest regret in each row.

$K$	$\tau\uparrow$	$\rho\uparrow$	NDCG@5 $\uparrow$	Regret (%) $\downarrow$
256	0.722	0.901	0.954	2.10
512	0.731	0.902	0.987	0.00
1024	0.754	0.917	0.935	6.40
2048	0.562	0.730	0.798	5.38
4096	0.552	0.711	0.981	5.15
8192	0.572	0.736	0.944	5.10

Appendix

Table 15: Additional ranking analysis in the SONIC scaling experiment. Ranking metrics use the 200 random configurations at each budget; regret refers to the proposed configuration. The Mean row weights the six budgets equally.

Source sequence	Frames	Duration (s)
CMU/143/143_14_poses	617	12.32
BMLmovi/Subject_57_F_MoSh/Subject_57_F_2_poses	277	5.52
EKUT/300/PushBeamPedar_15_poses	247	4.92
BioMotionLab_NTroje/rub012/0009_normal_jog1_poses	109	2.16
CMU/74/74_01_poses	400	7.98

Appendix

Table 16: New motions used to evaluate strategy generalization. Durations count executed control steps at

0.02

s per step.

Five new motions, SONIC
BFM	$K$	Ours	Random forest	RBF	Gaussian process	Empirical Optimum
SONIC	256	$0.1621\pm 0.0166$	$0.1553\pm 0.0318$	$0.2653\pm 0.0167$	$0.1617\pm 0.0171$	$0.1553\pm 0.0318$
SONIC	512	$0.0875\pm 0.0040$	$0.1174\pm 0.0104$	$0.2346\pm 0.0225$	$0.0786\pm 0.0053$	$0.0752\pm 0.0045$
SONIC	1024	$0.0660\pm 0.0026$	$0.0887\pm 0.0065$	$0.2014\pm 0.0141$	$0.0722\pm 0.0028$	$0.0647\pm 0.0033$
SONIC	2048	$0.0559\pm 0.0004$	$0.0728\pm 0.0067$	$0.2014\pm 0.0141$	$0.0849\pm 0.0035$	$0.0559\pm 0.0004$
SONIC	4096	$0.0542\pm 0.0011$	$0.0603\pm 0.0023$	$0.2014\pm 0.0141$	$0.0827\pm 0.0016$	$0.0537\pm 0.0022$

Appendix

Table 17: Measured costs on five new motions at matched compute budgets; lower is better. Entries report mean cost

\pm

one sample standard deviation across the three seed means. Empirical Optimum uses the configuration with the lowest mean cost among the four strategies and 50 random configurations on each BFM.

Figure 6 : five new motions
BFM	$K$	Ours	Random forest	RBF	Gaussian process
SONIC	256	4.34	0.00	70.81	4.07
	512	16.27	56.07	211.91	4.52
	1024	2.07	37.10	211.23	11.52
	2048	0.00	30.22	260.27	51.91
	4096	0.91	12.36	275.05	53.99

Appendix

Table 18: Allocation regret (%) on five new motions. The reference is Empirical Optimum in Table 17 . Each Mean row weights the five budgets equally; the Figure 6 mean also weights BFMs equally. Bold marks the lowest regret in each row.

BFM	$K$	$J_{\mathrm{prop}}\downarrow$	$J_{\mathrm{best}}\downarrow$	Regret (%) $\downarrow$	$\tau\uparrow$	$\rho\uparrow$	NDCG@5 $\uparrow$
SONIC	256	0.1621	0.1553	4.34	0.726	0.886	0.985
SONIC	512	0.0875	0.0752	16.27	0.758	0.916	0.956
SONIC	1024	0.0660	0.0647	2.07	0.678	0.842	0.902
SONIC	2048	0.0559	0.0559	0.00	0.479	0.631	0.691
SONIC	4096	0.0542	0.0537	0.91	0.527	0.703	0.945
SONIC	Mean	0.0851	0.0810	4.72	0.634	0.795	0.896

Appendix

Table 19: Additional ranking analysis on new motions and BFMs. Costs average five motions and three seeds. Ranking metrics use the 50 random configurations per budget; regret refers to the proposed configuration. Each Mean row averages five budgets, and each Max row gives the largest regret among those budgets.

Parameter	Distribution
Contact-friction multiplier	$\mathcal{U}(0.8,1.2)$
Joint-armature multiplier	$\mathcal{U}(0.8,1.2)$
Joint-friction-loss multiplier	$\mathcal{U}(0.8,1.2)$
Actuation-strength multiplier	$\mathcal{U}(0.9,1.1)$
Root linear velocity offset	$\mathcal{N}(0,0.02^{2})$ per component
Root angular velocity offset	$\mathcal{N}(0,0.05^{2})$ per component

Appendix

Table 20: Randomization distributions used in downstream label optimization. Multipliers scale nominal dynamics parameters; velocity perturbations are additive, zero-mean Gaussian offsets.

Task	Success condition	Max steps
Boxing	The hand/fist proxy enters the target sphere around the punching bag.	900
Open Door	The latch is unlocked and the door opens at least $60^{\circ}$ , with root height $\geq 0.45$ m and up-axis $z\geq 0.60$ .	1,800
Pick & Place Box	The box rests on the shelf support surface.	1,450
Sit Sofa	Sitting-related links enter the seat region with contact, or satisfy the geometric occupancy criterion.	2,000
Vision Navigation	Both feet reach the target region while the robot remains standing.	1,800

Appendix

Table 21: Success conditions and episode limits for the five HumanoidArena tasks.

Signal	Size	Representation
Language	Text	Task description
Egocentric frame	$376\times 672\times 3$	ZED-M left eye (H $\times$ W $\times$ C)
Body state	29	Joint values
Hand state	$6+6$	BrainCo Revo2, normalized to $[0,1]$
Projected gravity	3	Computed from IMU orientation
Body action	64	SONIC motion token

Appendix

Table 22: Real-robot observations and actions at the policy interface. The body-and-hand state contains 41 values, with three additional projected-gravity components computed from the IMU orientation.

Parameter	Value
Optimizer	AdamW
Learning rate	$10^{-4}$
Global batch size	160
Learning-rate schedule	Cosine, 30,000 updates
Warmup	5% (1,500 updates)
Weight decay	$10^{-5}$

Appendix

Table 23: Shared GR00T fine-tuning configuration for the real-robot experiments with SONIC execution.

Task	Execution sequence	Success condition
Cup in basket	Approach the table with one or two short steps, grasp the cup with the right hand, and place it in the basket to its right.	The cup is picked up and placed in the basket.
Parcel bag on shelf	Grasp the bag from the table, move and turn left toward the shelf, and place it at the target location.	The bag is picked up and placed at the target shelf location.
Basket on shelf	Squat to grasp the low basket with the right hand, rise while lifting it, then move and turn left to place it on the shelf.	The basket is picked up and placed at the target shelf location.
Bottle in drawer	Grasp the bottle from the cabinet top, lower it into the initially open drawer, and push the drawer closed.	The bottle is inside the drawer and the drawer is closed.

Appendix

Table 24: Real-robot task sequences and success conditions. Shelf placements use the designated target location.

Source	Target	Labels	Boxing	Open Door	Pick & Place Box	Sit Sofa	Vision Navigation	Mean
SONIC	SONIC	Original	$83.9\pm 1.0$	$84.4\pm 1.9$	$61.1\pm 1.0$	$88.9\pm 1.0$	$60.6\pm 1.0$	$75.8\pm 0.5$
SONIC	SONIC	State	$0.0\pm 0.0$	$0.0\pm 0.0$	$0.0\pm 0.0$	$0.0\pm 0.0$	$0.0\pm 0.0$	$0.0\pm 0.0$
SONIC	SONIC	Ours	$63.9\pm 1.0$	$51.1\pm 1.0$	$54.4\pm 1.0$	$66.7\pm 1.7$	$47.2\pm 1.0$	$56.7\pm 0.3$
SONIC	TWIST2	Original	$48.9\pm 1.9$	$0.0\pm 0.0$	$0.0\pm 0.0$	$17.8\pm 1.9$	$11.7\pm 1.7$	$15.7\pm 0.6$
SONIC	TWIST2	State	$33.9\pm 1.0$	$0.0\pm 0.0$	$0.0\pm 0.0$	$14.4\pm 1.0$	$1.7\pm 1.7$	$10.0\pm 0.3$
SONIC	TWIST2	Ours	$50.0\pm 1.7$	$0.0\pm 0.0$	$0.0\pm 0.0$	$21.7\pm 1.7$	$7.2\pm 1.9$	$15.8\pm 0.4$

Appendix

Table 25: HumanoidArena task success rates (%) for all evaluated source–target pairings in the five-task subset. Source is the demonstration-collection BFM and target is the execution BFM. Original, State, and Ours denote the three label conditions. Results are reported as mean

\pm

sample standard deviation across three training seeds, with 60 evaluation rollouts per task and seed.

Explore similar work

Task Inference Beyond Least Squares in Behavioral Foundation Models
Oct 4, 2026Kuan-Hsun Tu, Chien-Sheng Chiang, Hsin-Wei Chen +2Reinforcement Learning Multi-Task RL
National Taiwan University
Scaling Behavior Foundation Model for Humanoid Robots
Jul 16, 2026Weishuai Zeng, Kangning Yin, Xiaojie Niu +15Robot Foundation Models Humanoid Robot Control
The Chinese University of Hong Kong · Galbot · Shanghai Artificial Intelligence Laboratory +4
BFMTrack: Latent Sequence Optimization for Physics-Based Motion Tracking with Behavioral Foundation Models
Jun 23, 2026Thomas Rupf, Agon Serifi, David Müller +4Humanoid Robot Control
Disney Research, Zurich, Switzerland

Oct 4, 2026J/K move · Enter open · S save

Task Inference Beyond Least Squares in Behavioral Foundation Models

Kuan-Hsun Tu, Chien-Sheng Chiang, Hsin-Wei Chen, Ping-Chun Hsieh+1

National Taiwan University

Reinforcement Learning Multi-Task RL

Open paper AskPDF arXiv

Behavioral Foundation Models (BFMs) aim to solve a wide range of downstream tasks without test-time policy learning by inferring a task vector from the reward function. While efficient, the retrieved policies are often suboptimal because of how this task vector is inferred, typically with ordinary least squares (OLS). OLS minimizes reward reconstruction error but leaves the ordering of rewards unconstrained, which can bias the successor measure of the retrieved zero-shot policy away from that of the optimal policy. In this work, we propose BLS, an efficient test-time inference method that balances minimizing reward reconstruction error with reducing successor-measure mismatch. Theoretically, we provide a suboptimality gap upper bound characterized by both successor-measure and reward-function residuals. Empirically, we evaluate BLS on top of state-of-the-art BFMs across benchmarks for locomotion, manipulation, and humanoid control. BLS outperforms existing task inference baselines with negligible computational overhead. Project page: https://embodiedai-ntu.github.io/BLS

Abstract

Explore similar work

Task Inference Beyond Least Squares in Behavioral Foundation Models

Scaling Behavior Foundation Model for Humanoid Robots

BFMTrack: Latent Sequence Optimization for Physics-Based Motion Tracking with Behavioral Foundation Models