"""Claims 4 & 5: internal-consistency audit of the paper's reported tables. Paper: VLA-MBPO, arXiv:2603.20607 (OpenReview yKQ8GrwEhr). Reproducing Claim 4 end-to-end (LIBERO +9.1) needs pi_0.5 + a finetuned BAGEL-7B world model + PPO across 40 tasks -- far outside this budget, and no code or checkpoint was released. What CAN be checked without any of that is whether the reported numbers are internally consistent: the deltas, the averages, and the evaluation-protocol granularity all have to line up. Claim 5 ("a single universal set of hyperparameters across all tasks") is a statement *about* Table 5, so it can be checked directly against Table 5. """ import json import os from fractions import Fraction # --------------------------------------------------------------------------- # Table 2 (LIBERO), transcribed from the paper # --------------------------------------------------------------------------- TABLE2 = { "pi_0.5 (SFT)": {"Spatial": 78.2, "Object": 88.6, "Goal": 85.8, "Long": 54.6, "Avg": 76.8}, "BC (WM)": {"Spatial": 80.6, "Object": 89.8, "Goal": 85.0, "Long": 48.6, "Avg": 76.0}, "pi_RL": {"Spatial": 86.0, "Object": 92.4, "Goal": 90.8, "Long": 61.2, "Avg": 82.6}, "IDQL": {"Spatial": 79.0, "Object": 92.4, "Goal": 86.4, "Long": 52.2, "Avg": 77.5}, "VLA-MBPO": {"Spatial": 87.8, "Object": 96.6, "Goal": 92.8, "Long": 66.8, "Avg": 85.9}, } DELTA_ROW = {"Spatial": 9.6, "Object": 8.0, "Goal": 6.8, "Long": 12.2, "Avg": 9.1} SUITES = ["Spatial", "Object", "Goal", "Long"] # Table 3: rollout-scheme ablation on LIBERO-Long TABLE3 = {"branched n=1": 63.9, "branched n=2": 66.8, "branched n=4": 62.9, "full horizon": 52.8} # Table 5: RL hyperparameters per task suite TABLE5 = { "Sample size": {"Spatial": 512, "Object": 512, "Goal": 512, "Long": 1280, "Real-World": 512}, "Batch size": {"Spatial": 512, "Object": 512, "Goal": 512, "Long": 512, "Real-World": 512}, "Rollout chunk": {"Spatial": 2, "Object": 2, "Goal": 2, "Long": 2, "Real-World": 2}, "Actor lr": {"Spatial": 5e-6, "Object": 5e-6, "Goal": 5e-6, "Long": 5e-6, "Real-World": 5e-6}, "Critic lr": {"Spatial": 1e-4, "Object": 1e-4, "Goal": 1e-4, "Long": 1e-4, "Real-World": 1e-4}, "Reward discount gamma": {"Spatial": 0.99, "Object": 0.99, "Goal": 0.99, "Long": 0.99, "Real-World": 0.99}, "GAE lambda": {"Spatial": 0.95, "Object": 0.95, "Goal": 0.95, "Long": 0.95, "Real-World": 0.95}, "Clip ratio eps": {"Spatial": 0.1, "Object": 0.1, "Goal": 0.1, "Long": 0.1, "Real-World": 0.1}, "Action chunk H": {"Spatial": 10, "Object": 10, "Goal": 10, "Long": 10, "Real-World": 10}, "Denoise steps": {"Spatial": 3, "Object": 3, "Goal": 3, "Long": 3, "Real-World": 3}, "Noise level": {"Spatial": 0.5, "Object": 0.5, "Goal": 0.5, "Long": 0.5, "Real-World": 0.5}, "Update to data": {"Spatial": 20, "Object": 20, "Goal": 20, "Long": 50, "Real-World": 20}, } # Evaluation protocol: "average success rate over 50 evaluation episodes per task # across all 10 tasks in each suite" => 500 episodes => granularity 0.2%. EPISODES_PER_SUITE = 50 * 10 GRAIN = Fraction(100, EPISODES_PER_SUITE) # = 0.2 def approx(a, b, tol=1e-9): return abs(a - b) < tol def is_on_grid(v, grain=GRAIN): """Is v attainable as a success rate over EPISODES_PER_SUITE episodes?""" return (Fraction(str(v)) / grain).denominator == 1 def audit_claim4(): out = {"delta_row": [], "averages": [], "granularity": [], "table3": []} print("=" * 78) print("CLAIM 4 -- Table 2 internal consistency") print("=" * 78) # (a) Delta row = VLA-MBPO - pi_0.5(SFT)? print("\n(a) Delta row should equal VLA-MBPO minus pi_0.5 (SFT):") base, best = TABLE2["pi_0.5 (SFT)"], TABLE2["VLA-MBPO"] for c in SUITES + ["Avg"]: computed = round(best[c] - base[c], 10) ok = approx(computed, DELTA_ROW[c]) out["delta_row"].append({"column": c, "reported_delta": DELTA_ROW[c], "computed_delta": computed, "match": ok}) print(f" {c:8s} reported {DELTA_ROW[c]:+5.1f} computed {computed:+5.1f} " f"{'OK' if ok else '<-- MISMATCH'}") # (b) Avg column = mean of the four suites? (each suite is 10 tasks => equal weight) print("\n(b) Avg column should equal the mean of the four suites:") for model, row in TABLE2.items(): mean = sum(row[c] for c in SUITES) / 4 ok = approx(mean, row["Avg"]) trunc_ok = approx(int(mean * 10) / 10, row["Avg"]) out["averages"].append({"model": model, "reported_avg": row["Avg"], "computed_mean": mean, "exact_match": ok, "match_if_truncated": trunc_ok}) tag = "OK" if ok else ("OK (only if truncated, not rounded)" if trunc_ok else "<-- MISMATCH") print(f" {model:14s} reported {row['Avg']:5.2f} computed {mean:6.3f} {tag}") # (c) Are all reported values attainable on a 500-episode grid? print(f"\n(c) All values must be multiples of {float(GRAIN)} " f"({EPISODES_PER_SUITE} eval episodes per suite):") for model, row in TABLE2.items(): for c in SUITES: ok = is_on_grid(row[c]) out["granularity"].append({"table": "2", "model": model, "column": c, "value": row[c], "on_grid": ok}) if not ok: print(f" {model:14s} {c:8s} {row[c]:5.1f} <-- NOT on the 0.2 grid") bad2 = [g for g in out["granularity"] if not g["on_grid"]] print(f" {len(out['granularity']) - len(bad2)}/{len(out['granularity'])} on-grid") # (d) Table 3 ablation print("\n(d) Table 3 (rollout scheme, LIBERO-Long):") for k, v in TABLE3.items(): ok = is_on_grid(v) out["table3"].append({"scheme": k, "value": v, "on_grid": ok}) print(f" {k:14s} {v:5.1f} {'on-grid' if ok else '<-- NOT on the 0.2 grid'}") consistent = approx(TABLE3["branched n=2"], TABLE2["VLA-MBPO"]["Long"]) out["table3_matches_table2"] = consistent print(f" Table 3 'branched n=2' ({TABLE3['branched n=2']}) == Table 2 Long " f"({TABLE2['VLA-MBPO']['Long']})? {'YES' if consistent else 'NO'}") # (e) Can one typo explain everything? print("\n(e) Hypothesis: the Goal cell is a typo. Solve for the value that makes") print(" BOTH the delta row and the Avg column self-consistent:") goal_from_delta = base["Goal"] + DELTA_ROW["Goal"] goal_from_avg = TABLE2["VLA-MBPO"]["Avg"] * 4 - sum( TABLE2["VLA-MBPO"][c] for c in ["Spatial", "Object", "Long"]) print(f" implied by delta row (+6.8): Goal = {goal_from_delta:.2f}") print(f" implied by Avg (85.9, exact): Goal = {goal_from_avg:.2f}") cand = goal_from_delta mean_c = (TABLE2["VLA-MBPO"]["Spatial"] + TABLE2["VLA-MBPO"]["Object"] + cand + TABLE2["VLA-MBPO"]["Long"]) / 4 print(f" If Goal = {cand:.1f} (on-grid: {is_on_grid(round(cand,1))}):") print(f" Avg = {mean_c:.3f} -> truncates to {int(mean_c*10)/10:.1f} " f"(paper prints {TABLE2['VLA-MBPO']['Avg']}) " f"{'MATCH' if approx(int(mean_c*10)/10, TABLE2['VLA-MBPO']['Avg']) else 'no'}") d_avg = mean_c - base["Avg"] print(f" Delta_avg = {d_avg:.3f} -> truncates to {int(d_avg*10)/10:.1f} " f"(paper prints {DELTA_ROW['Avg']}) " f"{'MATCH' if approx(int(d_avg*10)/10, DELTA_ROW['Avg']) else 'no'}") out["typo_hypothesis"] = { "goal_implied_by_delta": goal_from_delta, "goal_implied_by_avg": goal_from_avg, "printed_goal": TABLE2["VLA-MBPO"]["Goal"], "candidate": cand, "candidate_on_grid": is_on_grid(round(cand, 1)), "avg_if_candidate": mean_c, "avg_truncated": int(mean_c * 10) / 10, "avg_matches_printed": approx(int(mean_c * 10) / 10, TABLE2["VLA-MBPO"]["Avg"]), "delta_avg_if_candidate": d_avg, "delta_avg_truncated": int(d_avg * 10) / 10, "delta_avg_matches_printed": approx(int(d_avg * 10) / 10, DELTA_ROW["Avg"]), } # (f) headline claims print("\n(f) Headline claims of Claim 4:") largest = max(SUITES, key=lambda c: DELTA_ROW[c]) beats_rl = {c: TABLE2["VLA-MBPO"][c] > TABLE2["pi_RL"][c] for c in SUITES + ["Avg"]} out["headline"] = { "avg_gain_reported": DELTA_ROW["Avg"], "largest_gain_suite": largest, "largest_gain_value": DELTA_ROW[largest], "largest_gain_is_long": largest == "Long", "beats_piRL_everywhere": all(beats_rl.values()), "beats_piRL_per_suite": beats_rl, } print(f" '+9.1 average' -> reported Avg delta = {DELTA_ROW['Avg']} OK") print(f" 'largest gain on Long' -> largest delta is {largest} " f"({DELTA_ROW[largest]:+.1f}) {'OK' if largest == 'Long' else 'MISMATCH'}") print(f" 'outperforms pi_RL' -> beats pi_RL on every suite: " f"{all(beats_rl.values())} {beats_rl}") return out def audit_claim5(): print() print("=" * 78) print("CLAIM 5 -- 'a single universal set of hyperparameters across all tasks'") print("=" * 78) cols = ["Spatial", "Object", "Goal", "Long", "Real-World"] constant, varying = [], [] for p, row in TABLE5.items(): vals = [row[c] for c in cols] if len(set(vals)) == 1: constant.append(p) else: varying.append({"parameter": p, "values": row, "distinct": sorted(set(vals), key=float)}) print(f"\n constant across all task suites : {len(constant)}/{len(TABLE5)}") print(f" VARYING across task suites : {len(varying)}/{len(TABLE5)}") for v in varying: print(f" - {v['parameter']:18s} " + " ".join(f"{c}={v['values'][c]}" for c in cols)) verdict = len(varying) == 0 print(f"\n Claim 5 as literally stated ('single universal set') holds? " f"{'YES' if verdict else 'NO'}") return {"n_params": len(TABLE5), "n_constant": len(constant), "n_varying": len(varying), "constant": constant, "varying": varying, "claim_holds_literally": verdict, "paper_own_text": ("Appendix: 'the majority of hyperparameters remain constant " "across all task suites. The primary adaptation required is " "scaling the Sample Size (and the corresponding Update to data " "steps) for long horizon tasks.'"), "main_text": ("Sec 3.3: 'our method maintains a single set of hyperparameters " "across all tasks (Table 5)'; Abstract: 'maintains a universal set " "of hyperparameters across all tasks'")} def main(): os.makedirs("outputs/claims45", exist_ok=True) c4 = audit_claim4() c5 = audit_claim5() payload = { "meta": {"paper": "arXiv:2603.20607 (OpenReview yKQ8GrwEhr)", "note": "Internal-consistency audit of reported tables; no training involved.", "episodes_per_suite": EPISODES_PER_SUITE, "grid": float(GRAIN)}, "table2": TABLE2, "delta_row": DELTA_ROW, "table3": TABLE3, "table5": TABLE5, "claim4_audit": c4, "claim5_audit": c5, } with open("outputs/claims45/claims45_audit.json", "w", encoding="utf-8") as f: json.dump(payload, f, indent=2) print("\nWrote outputs/claims45/claims45_audit.json") if __name__ == "__main__": main()