PaperlayerView exact revision on arXiv

Source-linked AI summary

Memory That Changes Action Is Not Memory That Guides It: Counterfactual Auditing of History-Conditioned Robot Policies

Jiajie Zhang, Yankai Xiang, Changhao Chen

arXiv:2609.27247v1cs.RO

TL;DR

Memory evaluations often measure task success or action changes without establishing whether memory guides a particular decision. CMA compares a frozen policy’s decisions under different histories at an identical present, then evaluates the actions across both worlds. The results show that memory can change behavior without reliably selecting the warranted branch, while native-memory interventions can also redirect behavior and recover success after faults.

  • Problem

    Task-level outcomes, memory accessibility, and action changes do not establish whether remembered content caused the warranted decision at a fixed present.

    Source: §I · p. 1

  • Method

    CMA crosses two lawful histories at a verified-identical present, queries a frozen policy under common randomness, and evaluates each saved action in both worlds.

    Source: §I · p. 2, A. Audit Protocol · p. 2

  • Results

    The audit finds that memory can change actions without changing the selected branch, while restoring a protected component recovers 38.9 points of success lost to injected bank faults.

    Source: B. When Action Change Misleads (RQ1) · p. 5, §V · p. 8

  • Takeaways & Limitations

    CMA complements episode-scored evaluation by testing sensitivity, warranted choice, matched-world value, and reliability at a fixed present.

    Source: §V · p. 8

    Conclusions are bounded to the audited checkpoints and tasks; two held-out comparisons showed no exclusive predictive gain over strong simple baselines.

    Source: §V · p. 8

Abstract

from arXiv · show

A robot returning a block to its origin tray may encounter two task-consistent pasts that reconverge to the same current input but warrant different actions. Yet memory-policy evaluations often rely on task success or action change under memory perturbation, neither of which establishes that memory guides the decision. We propose the \textbf{Counterfactual Memory Audit (CMA)}, an evaluation protocol that crosses two histories at a verified-identical present, queries a frozen policy under common randomness, and evaluates each saved action under both pasts. This separates memory sensitivity, warranted choice, matched-world physical value, and per-pair reliability. On Mem-0, every audited Put Back pair changes action, but only $20/64$ pairs are fully reliable; at a later Swap decision, all paired actions change while both memories select the same branch. Native interventions further show closed-loop influence: replacing the history bank redirects behavior toward the replaced content, while restoring a 4096-byte protected anchor recovers $38.9$ points of Swap success lost to injected bank faults. On a dual-arm physical platform, memory changes saved actions, yet five of nine completed Put Back manipulations reach the wrong target. These results show that a robot can remember and react without reliably using memory to choose the behavior its past warrants. CMA provides a decision-level audit for distinguishing these cases.

I. INTRODUCTION

Task-level outcomes and action changes do not establish whether memory guides an individual decision. The Counterfactual Memory Audit tests that question by comparing memories at an identical present and evaluating the resulting actions across both worlds.

I. INTRODUCTION

  • Two lawful histories can reconverge to the same input while warranting different return actions, leaving the correct branch absent from current observation.The example is perceptual aliasing in a partially observable control problem.Source: §I · p. 1, Figure 1 · p. 1
  • Episode success, retrieval accuracy, ablations, and action distance each leave unresolved whether memory caused the warranted decision at a fixed present.A task can succeed after recovery from a local error, and changed actions need not represent a changed task branch.Source: §I · p. 1
  • CMA crosses two task-consistent histories at a verified-identical present, queries the frozen policy under common randomness, and scores each action in both evaluator-only worlds.Its four audit levels are sensitivity, warranted choice, matched-world physical value, and per-pair reliability.Source: §I · pp. 1–2
  • Memory can exert closed-loop influence, but task success and action change alone do not show that it selects the warranted behavior.Replacing the history bank redirects behavior toward its replaced content; restoring a protected component recovers success lost to injected bank faults.Source: §I · p. 2
  • Across frozen policies and a dual-arm platform, memory can change actions without reliably selecting the warranted branch; restoring a 4096-byte component recovers 38.9 points of success lost to injected bank faults.The audit covers separately analyzed policies, tasks, decision phases, and four intervention classes, with measured resets on hardware.Source: §I · p. 2

II. RELATED WORK

Prior work develops memory-bearing manipulation policies and episode-level benchmarks, alongside counterfactual and causal interventions. The paper distinguishes CMA by auditing decisions at an identical present and scoring saved actions across evaluator worlds.

A. Memory Policies and Benchmarks

  • Memory policies use recurrent states, adaptive recoding, retrieval, and episodic or multi-scale stores, while benchmarks score memory-intensive tasks episode by episode.The cited benchmarks include MemoryBench [7], MIKASA [8], RMBench, RoboMME, and RoboMemArena [9]–[11].Source: A. Memory Policies and Benchmarks · p. 2

B. Counterfactual Tests of Memory Use

  • Counterfactual evaluations range from action variation to counterfactual choice accuracy, but do not combine lawful native histories, an exactly equal current packet, crossed physical value, and bidirectional pair reliability.TFP tests action variation [16], while Chameleon reports counterfactual choice accuracy [17].Source: B. Counterfactual Tests of Memory Use · p. 2

C. Causal Attribution and Memory Intervention

  • Related causal methods probe representations or memory interventions, whereas CMA diagnoses frozen policies through closed-loop behavior and restores a declared memory tensor at a fixed boundary.CMA uses pair-cluster inference and common randomness to address repeated-measure unit errors.Source: C. Causal Attribution and Memory… · p. 2

III. COUNTERFACTUAL MEMORY AUDIT

CMA isolates memory’s role at an audited decision by holding the policy-visible present fixed and changing the native memory. It evaluates each saved action in both worlds without exposing those worlds to the policy.

A. Audit Protocol

  • CMA creates two task-consistent histories that warrant different semantic branches, updates their native memories, and reconverges them to a verified-identical current packet.The packet includes all non-memory inputs available to the policy, while the underlying physical state is verified equal but kept hidden.Source: A. Audit Protocol · pp. 2–3
  • The frozen policy is queried once per memory under common randomness, and eligible pairs are determined without using action outcomes or task success.A failed current-state equality check excludes a pair, and excluded pairs are not replaced after observing the action.Source: A. Audit Protocol · p. 3, Figure 2 · p. 3
  • Each saved action is evaluated in both evaluator-only worlds, isolating memory’s local role without estimating how often perceptual aliasing occurs naturally.World labels affect neither the policy query nor pair eligibility.Source: A. Audit Protocol · p. 3, Figure 2 · p. 3

B. Decision-Level Audit Metrics

The audit separates whether memory changes an action from whether it selects a warranted branch, produces physical value, and does so reliably. It measures these dimensions with matched-world oracle outcomes and continuous action values.

B. Decision-Level Audit Metrics

  • Action sensitivity is unsigned: different actions may approach the same tray, choose neither branch, or follow the same branch.S averages a preregistered executed-action distance across eligible pairs and action-seed repeats.Source: B. Decision-Level Audit Metrics · p. 3
  • The oracle scores actions as warranted, alternative, or ambiguous; correct-choice and ambiguity rates retain the underlying matched-world outcomes summarized by signed faithfulness.The oracle uses registered branch criteria; in the primary Put Back audit, the progress margin is δ = 0.05.Source: B. Decision-Level Audit Metrics · p. 3
  • Matched-world value measures the World × Memory contrast, while U1 reads one executed chunk and absolute regret measures shortfall from normalized reference value 1.Unlike branch choice, U1 and regret retain graded physical progress within a branch.Source: B. Decision-Level Audit Metrics · pp. 3–4
  • Reliability requires both memories to select their warranted branches across every registered action seed and both directions.The protocol reports direction, seed, and bidirectional results separately.Source: B. Decision-Level Audit Metrics · p. 4

C. Pair-Level Inference and Controls

Inference treats constructed audit pairs as clusters and uses ordinary episodes for natural-rollout tests. Irrelevant-history controls test nuisance effects and route consistency using preregistered equivalence margins.

1) Statistical units:

  • Pair clusters, not their repeated memories, queries, evaluator worlds, or replays, are the independent units of constructed-audit inference.Natural-rollout tests instead resample ordinary episodes within layout strata; models, tasks, and local-audit cohorts are not pooled.Source: 1) Statistical units · p. 4

2) Irrelevant-history control:

  • Semantic effects must satisfy the positive rule on both detour routes, while nuisance and interaction effects require their full 90% pair-bootstrap intervals to lie within preregistered margins.Merely failing to reject a difference is insufficient evidence of equivalence.Source: 2) Irrelevant-history control · p. 4

D. Native-Rollout Memory Interventions

Native-rollout interventions replace or restore a declared history carrier while fixing the recipient’s current input and non-memory state at intervention. The experiments examine closed-loop influence, calibration and controls, natural-rollout decision structure, and transfer to physical hardware.

D. Native-Rollout Memory Interventions

  • Native-event interventions replace only the declared history carrier at an outcome-blind event, holding the recipient’s current input and non-memory state fixed.Replacement content comes from the recipient’s own history or a donor episode; measured splice residuals bound interpretation to the declared carrier.Source: D. Native-Rollout Memory Interventions · p. 4
  • Matched-prefix restoration clears the bank for a fixed fault period, then restores a declared component at the next query under a shared intervention budget.Its endpoint is official full-episode success, distinct from one-chunk physical value U1.Source: D. Native-Rollout Memory Interventions · p. 4
  • Native-event interventions read out commitment as a sustained approach toward one target over the remaining continuation, with transient earlier direction reported separately.The intervention stage and the readout window are distinct; afterward, native rollout continues, so a cleared bank refills from post-clear observations, including a fresh Mem-0 anchor.Source: D. Native-Rollout Memory Interventions · p. 4

IV. EXPERIMENTS

  • The experiments ask whether success and action sensitivity establish memory-guided choice, whether history replacement and restoration exert closed-loop influence, whether controls and natural rollouts match the audit, and whether the distinction persists on hardware.Source: §IV · p. 4

A. Experiment Setup

The experiments audit frozen Mem-0 and BridgeVLA++ policies, with policies analyzed separately and simulated checkpoints left unmodified. On the same frozen Mem-0 checkpoint, Put Back is correct on average but unreliable pair by pair, while later Swap changes every action without changing its branch.

A. Experiment Setup

  • The audit separately examines Mem-0 on Put Back, Rearrange, and Swap, and BridgeVLA++ on Put Back; simulated evaluations use frozen checkpoints.The physical Mem-0 modules are retrained for the dual-arm PiPER embodiment and frozen before auditing.Source: A. Experiment Setup · p. 4
  • Both policies pass native qualification before audit cohorts, and confirmatory intervals use 10,000 preregistered bootstrap draws.Mem-0 scores 20/20, 18/20, and 19/20 on Put Back, Rearrange, and Swap; qualification episodes are excluded from later analyses.Source: A. Experiment Setup · p. 4
  • The primary constructed audit places the later-Swap decision after the common grasp-and-lift phase, where histories warrant different branches despite identical verified current inputs.Official Swap success requires the complete swapped configuration.Source: A. Experiment Setup · pp. 4–5

B. When Action Change Misleads (RQ1)

  • Put Back is correct on average but unreliable pair by pair, while later Swap changes every action without changing its branch.Put Back has a 0.729 correct-choice rate but only 20/64 fully reliable pairs; later Swap has 24/24 action changes but no branch changes.Source: B. When Action Change Misleads (RQ1) · p. 5, Table 2 · p. 5
  • At Put Back target 3, correct-choice rate is 0.115 versus 0.865–0.979 at targets 0–2, a target-associated pattern rather than an identified geometric cause.The pattern also appears in the held-out population; target location was not intervened on.Source: B. When Action Change Misleads (RQ1) · p. 5

C. When Replacement Redirects and Restoration Recovers (RQ2)

Native history replacement redirects behavior toward the replacement content, including when that content encodes the opposite target. Early readouts can remain ambiguous even when later commitment follows the encoded history.

1) History replacement at native events:

  • Under opposite-target history, official success is 0/48 for each policy, and all 48 continuations per policy commit to the donor target.Own-history reconstruction preserves success in 48/48 BridgeVLA++ and 47/48 Mem-0 continuations.Source: 1) History replacement at native events · p. 5, Figure 4 · p. 6
  • Opposite-target replacement raises donor-corresponding first sustained directions over bank clearing by +0.583 early and +0.667 late.Self-restore controls reproduce saved actions, query schedules, and tracked block trajectories exactly.Source: 1) History replacement at native events · p. 5
  • At the early event, readouts remain ambiguous, yet all 24 continuations per history condition later commit to its encoded target.At the later center-grasp event, 21/24 readouts per condition are already content-corresponding.Source: 1) History replacement at native events · p. 5, Table 1 · p. 6
  • The native-event tests probe subsequent closed-loop commitment, unlike the constructed exact-current-state test of local decision dependence, where all 96 outcomes remain ambiguous.For BridgeVLA++ in the constructed setting, U1 = −0.0026 [−0.0044, −0.0010].Source: 1) History replacement at native events · p. 6

2) Memory Restoration Recovers Success:

Restoring a protected anchor recovers Swap success lost after injected bank faults, while full rebuild produces the same tested banks and subsequent trajectories as anchor restoration.

2) Memory Restoration Recovers Success:

  • After bank clearing, rebuilding from full observed history recovers all 12 native successes lost under removal in 24 fresh Mem-0 Swap episodes.No losses occurred in the primary comparison.Source: 2) Memory Restoration Recovers Success · p. 6
  • Anchor restoration raises official success to 20/36 from 6/36 without restoration, a 38.9 percentage-point improvement, but remains below the 26/36 native reference.Full rebuild also yields 20/36; the anchor-versus-no-restoration layout-stratified 95% bootstrap interval is [+27.8, +50.0].Source: 2) Memory Restoration Recovers Success · p. 6, Figure 5 · p. 6
  • At the tested recovery boundary, anchor restoration and full rebuild produce identical complete banks, states, actions, and subsequent trajectories.The recent-observation window has been replenished from actual observations, so the disturbed and rebuilt banks differ only in the anchor tensor.Source: 2) Memory Restoration Recovers Success · pp. 6–7

D. Calibration and Controls (RQ3)

CMA is checked against known semantic regimes, irrelevant history, and untreated behavior. The controls show that positive averages can coexist with individual failures, while the natural-behavior contrast remains unresolved and observational.

1) Known-Answer Calibration:

  • Four known-answer calibrator cells returned distinguishable regimes, while controlled-update semantic-negative prevalence was 0.217 [0.127, 0.321] for recurrent and 0.100 [0.061, 0.143] for retrieval policies despite positive crossed averages.Ordered-Route certified the recurrent family; retrieval was excluded because only 2/5 seeds qualified. A branch-aware local controller achieved F = 1 [1] with bidirectional correctness one over 48 pairs, and reusing one genuine action across balanced pseudo-memory labels returned F = U1 = 0.Source: 1) Known-Answer Calibration · p. 7

2) Irrelevant-History Control:

  • Both mirror-detour routes passed the CMA positive rule, with pcorrect = 0.753/0.757, and all three nuisance intervals lay inside preregistered margins.The confirmation used 48 pair clusters with three common seeds; equivalence is limited to these two routes.Source: 2) Irrelevant-History Control · p. 7

3) Natural Occurrence Without Intervention:

  • 86/96 natural episodes contained the audited decision type, but the primary remaining-step contrast was unresolved and the secondary final-success association was noncausal.The 24 constructed later-Swap pairs could not be exactly reconstructed in natural rollouts, so constructed and natural cohorts were not pooled.Source: 3) Natural Occurrence Without… · p. 7

E. The Crossing Runs on Physical Hardware (RQ4)

On dual-arm hardware, the crossing reproduces the decision-level dissociation between memory-consistent choices and task success, alongside substantial physical variability. The discussion positions CMA as a decision-level complement whose conclusions remain bounded to audited checkpoints and tasks.

E. The Crossing Runs on Physical Hardware (RQ4)

  • The hardware crossing used two task-specific Mem-0 modules retrained for the embodiment and frozen before evaluation, with shared current packets and replayed prefixes before fresh-observation closed loops.Each task used eight native history pairs and two action seeds, with controlled resets before execution.Source: E. The Crossing Runs on Physical… · pp. 7–8
  • 6/32 Put Back and 8/32 Repick executions selected the history-consistent target, while full success was 4/32 and 5/32, respectively; no pair was correct in all four executions.Table 3 covers eight pair clusters and 32 executions per task; counts use operator-provided labels, not independent blinded annotation.Source: Table 3 · p. 8, E. The Crossing Runs on Physical… · p. 8

V. DISCUSSION AND CONCLUSION

  • Native-event bank replacement redirected commitment toward replaced content, while restoring one protected 4096-byte component recovered 38.9 points of success lost to injected bank faults.Source: §V · p. 8
  • CMA complements episode-scored evaluation by auditing sensitivity, warranted choice, matched-world value, and reliability at a fixed present; its conclusions are bounded to audited checkpoints and tasks.Two held-out comparisons showed no exclusive predictive gain over strong simple baselines.Source: §V · p. 8
Loading 2609.27247v1…