Altered Thoughts, Altered Actions: Reasoning Chain as Control Surface for a Vision-Language-Action Policy

Tuan Duong Trinh; Basim Azam; Mohammed Ishaq Ansari; Mohammed Yaqoob Ansari; Naveed Akhtar

cs.ROMar 13, 2026

Altered Thoughts, Altered Actions: Reasoning Chain as Control Surface for a Vision-Language-Action Policy

Authors: Tuan Duong Trinh, Basim Azam, Mohammed Ishaq Ansari, Mohammed Yaqoob Ansari, Naveed Akhtar

Organizations: University of Melbourne Parkville, VIC, Australia

Abstract

Vision-language-action policies map camera images and natural-language instructions to a robot's motor actions. Some of these policies are designed to reason in text before acting, generating a reasoning chain and decoding actions conditioned on that chain. The works introducing this design offer the reasoning chain as an oversight interface: text a person can read and edit to correct the policy. What an edited reasoning chain does to the policy's motor actions, whether it repairs them or corrupts them, remains an open question. We measure both directions, repair and corruption, with our deterministic entity swap applied to the instruction the policy receives and to the reasoning chain it generates. A forty-task observed backdrop across all four LIBERO simulation suites reveals that the cost of corrupting the reasoning chain concentrates where language alone determines the goal. There, on LIBERO-Goal, we run the decisive counterfactual intervention with DeepThinkVLA, chosen because its reasoning chain is exposed as plain text. The policy receives a corrupted instruction, paired with the reasoning chain it generates when that instruction is clean. This counterfactually correct reasoning chain recovers 47.8 pp of the lost success, our pre-registered confirmatory test. Had the chain merely restated what the camera image already determines, the injection could have changed nothing. Instead, all 10 tasks move in the predicted direction. The reasoning chain is therefore a working control surface: text written into it steers the robot, repairing behaviour when the text is right and corrupting it when the text is wrong. Whether to expose such a control surface is a real deployment tradeoff, and it can now be measured.

Figures & tables

Figure 1: (a) A non-reasoning VLA policy consumes one text input, the instruction. (b) DeepThinkVLA first generates a reasoning chain in the same context and then decodes actions conditioned on it, so a second text input reaches the same decode; our entity swap can edit either channel, and which channels it edits separates the four conditions. (c) Two rollouts under condition A (both texts intact) and condition B (intact instruction, corrupted reasoning chain); the seeded per-episode draw here rewrites stove as cream cheese box , and the policy lifts the object the swap names.

condition	instruction	reasoning chain
A	intact	generated from the clean instruction
B	intact	generated, then entity-swapped
C	entity-swapped	regenerated from the corrupted instruction
D	entity-swapped, identical to C	clean reasoning chain injected at every query

Figure 2: The entity swap and the four conditions, shown for the LIBERO-Goal task turn on the stove . The swap maps each object a text mentions to a seeded replacement, the same at every occurrence, in whichever channel the condition edits. The draw shown is cream cheese box, as in Figure 1 (c). C and D receive the identical corrupted instruction; D pairs it with the clean reasoning chain, injected at every query. The reasoning chain shown is the first query’s.

suite	scenes	tasks	goal structure	language must supply
LIBERO-Goal	1	10	10 distinct goals, 4 predicate types	the goal itself
LIBERO-Spatial	1	10	one goal: On (bowl, plate)	which of two identical bowls
LIBERO-Object	10	10	one schema: In ( $x$ , basket)	which object $x$
LIBERO-Long	9	10	10 distinct goals, mostly compound	little: the scene fixes the task

Table 1: Ground-truth structure of the four LIBERO suites, parsed from their .bddl task specifications; the per-task parse appears in Appendix D . Only in the Goal suite does the text determine what to do rather than which of several similar objects to act on.

condition	status	instruction	reasoning chain	success rate	degradation
A	observed	intact	generated	97.3%	–
B	observed	intact	entity-swapped	76.5%	20.8 pp
C	observed	entity-swapped	regenerated from corrupted	11.5%	85.8 pp
D	confirmatory	entity-swapped	clean, injected	59.3%	38.0 pp

Table 2: LIBERO-Goal results (

n=10

tasks, 600 episodes per condition): conditions above, pre-specified comparisons below; the descriptive comparison carries no test (§ 4.2 ).

Figure 3: The confirmatory contrast, task by task: C (instruction corrupted) against D (identical instruction, clean reasoning chain injected), with the clean reference as a tick, tasks sorted by rescue size. All 10 tasks move in the predicted direction; each pair is annotated with its rescue.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

instruction	goal predicate
open the middle drawer of the cabinet	Open (cabinet middle drawer)
open the top drawer and put the bowl inside	In (bowl, cabinet top drawer)
push the plate to the front of the stove	On (plate, region in front of stove)
put the bowl on the plate	On (bowl, plate)
put the bowl on the stove	On (bowl, stove cook region)
put the bowl on top of the cabinet	On (bowl, cabinet top)

Appendix

Table 3: The 10 LIBERO-Goal tasks, parsed from their .bddl specifications: one shared scene, 10 distinct goals over four predicate types ( Open , In , On , Turnon ). Object and region identifiers are shortened for readability; the specification files carry the full names.

Figure 4: The four rendered condition-B episodes not shown in the body: the LIBERO-Goal task turn on the stove at one run seed, episodes 0 to 4 except episode 1, which is the first in which the policy acts on the object the swap names and appears in Figure 1 (c). Each row label gives the object the swap wrote into the reasoning chain, the outcome, and the episode length; frame indices are printed per row because episode lengths differ. Episodes 0, 2, and 3 succeed despite the edit. Episode 4 fails at the step cap, having acted on the object the swap named rather than turning the stove knob.

scene	instruction	goal predicates
kitchen 3	turn on the stove and put the moka pot on it	Turnon (stove) $\wedge$ On (moka pot, stove)
kitchen 4	put the black bowl in the bottom drawer of the cabinet and close it	Close (cabinet bottom drawer) $\wedge$ In (bowl, cabinet bottom drawer)
kitchen 6	put the yellow and white mug in the microwave and close it	In (mug, microwave) $\wedge$ Close (microwave)
kitchen 8	put both moka pots on the stove	On (moka pot 1, stove) $\wedge$ On (moka pot 2, stove) $\wedge$ Turnon (stove)
living room 1	put both the alphabet soup and the cream cheese box in the basket	In (alphabet soup, basket) $\wedge$ In (cream cheese, basket)
living room 2	put both the alphabet soup and the tomato sauce in the basket	In (alphabet soup, basket) $\wedge$ In (tomato sauce, basket)

Appendix

Table 4: The 10 LIBERO-Long tasks, parsed from their .bddl specifications: nine scenes, mostly compound goals, four predicate types. Only living room 2 hosts two tasks.

Figure 5: Six corruptions of the reasoning chain, observed in the original environment on the forty tasks of the four LIBERO suites; these data predate the design of § 3.3 and are observed, not tested. Each cell shows the absolute success rate under the corruption and its change from that suite’s clean reference. The entity swap is the most damaging corruption in the matrix, with its largest cost on Goal (19.3 pp).

asset	role	provenance	licence
LIBERO	benchmark: scenes, initial states, success checks	public repository; specifications verified byte-identical to the public distribution	MIT (codebase); the demonstration datasets are CC BY 4.0 and are not used here
DeepThinkVLA code	evaluation and policy code	public repository, shallow clone, commit unrecorded	MIT
DeepThinkVLA checkpoint	the policy under study	released checkpoint, 6.9 GB in bfloat16, refactored from the public pi0-FAST checkpoint	Gemma Terms of Use, inherited through the PaliGemma backbone
PaliGemma backbone	SigLIP vision encoder and Gemma decoder inside the checkpoint	distributed with the checkpoint	Gemma Terms of Use (source-available, not OSI-approved)
robosuite 1.4.0	simulation stack	pinned package version	MIT
bddl 1.0.1	task-specification parser	pinned package version	MIT

Appendix

Table 5: External assets and their licences, verified against each distribution 2026-08-31. Code and model weights are listed separately where their licences differ.

Explore similar work

When Reasoning Helps Action: Monitoring and Steering Chain-of-Thought in Vision-Language-Action Policies
Sep 30, 2026Sathwik Karnik, Joseph JR. Lee, Aryaman Gupta +1Reasoning Trajectory Chain-of-Thought Reasoning
Safe and Intelligent Autonomy Lab, Stanford University
Continuous Reasoning for Vision-Language-Action
May 29, 2026Yueh-Hua Wu, Tatsuya Matsushima, Kei OtaAction Generation Reasoning Skills
AIRoA
In-Context VLA: Endowing Vision-Language-Action Models with Language via In-Context Post-Training and Agentic Tool Use
Aug 6, 2026Jiarui Yang, Wen Huang, Jiale Zhang +2Behavior Cloning
Nankai University · Tsinghua University · Swiss Federal Institute of Technology in Lausanne (EPFL)

Sep 30, 2026, cs.ROJ/K move · Enter open · S save

When Reasoning Helps Action: Monitoring and Steering Chain-of-Thought in Vision-Language-Action Policies

Sathwik Karnik, Joseph JR. Lee, Aryaman Gupta, Somil Bansal

Safe and Intelligent Autonomy Lab, Stanford University

Reasoning Trajectory Chain-of-Thought Reasoning Closed Loop Behavioral Divergence

Open paper AskPDF arXiv

Reasoning-enabled VLA policies expose chain-of-thought (CoT) traces that appear to explain and guide their actions, creating a potential interface for runtime safety through reasoning monitoring and correction. In this work, we define and operationalize two evaluation axes for assessing when this interface can improve embodied behavior: correctability, which measures whether unreliable reasoning can be detected and improved during generation, and actionability, which measures whether reasoning corrections produce behaviorally meaningful changes in the intended direction. To enable correctability, we introduce Token-level Reward for Utility-Steered Chain-of-Thought (TRUST), an offline-trained value model that predicts eventual reasoning correctness from partial prefixes and uses these estimates to monitor and selectively steer reasoning generation in frozen VLA policies. On the Alpamayo 1.5 driving VLA, TRUST monitors correctness with 88.9% accuracy and improves reasoning correctness from 75.9% to 90.0%. On a baseline-defined challenging subset in AlpaSim, TRUST reduces collision rate by 30.4% and maximum trajectory error by 11.5% relative to the unsteered policy, outperforming a compute-matched Best-of-4 baseline. On the DeepThinkVLA manipulation VLA, TRUST improves the correctness of grasp-state claims from 69.3% to 90.2% and action-choice claims from 68.8% to 85.9%, yet closed-loop task performance on LIBERO-Plus remains largely unchanged. Empirical analysis reveals intent-consistent behavioral effects in Alpamayo 1.5 but limited effects in DeepThinkVLA, helping interpret these different task-level outcomes. Together, our results show that gains in reasoning correctness do not automatically imply gains in embodied performance, motivating evaluation of correctability and actionability when using CoT as a runtime safety interface.

Abstract

Explore similar work

When Reasoning Helps Action: Monitoring and Steering Chain-of-Thought in Vision-Language-Action Policies

Continuous Reasoning for Vision-Language-Action

In-Context VLA: Endowing Vision-Language-Action Models with Language via In-Context Post-Training and Agentic Tool Use