Fixed-Reference Sign Advantage

• Fixed-Reference Sign Advantage is achieved by anchoring updates to static references, isolating intrinsic target signals from benchmark-induced fluctuations.
• It strictly probes phase transitions in quantum Monte Carlo by ensuring that free energy changes are reflected without contamination from a varying reference.
• This approach is broadly applicable, improving robustness in statistical inference, reinforcement learning, lattice field theory, and hardware compression designs.

Fixed-Reference Sign Advantage denotes a recurring methodological gain obtained when a sign, signed decomposition, or sign-based update is anchored to a parameter-independent or structurally fixed reference rather than to a moving, endogenous baseline. In the literature considered here, the phrase is used most explicitly for quantum Monte Carlo phase-transition probes, but closely related constructions also appear in decomposition-based inference, RLHF and RLVR, canonical formulations of lattice field theory, arithmetic design, and sign-preserving compression. A plausible cross-domain characterization is that fixing the reference suppresses benchmark-induced contamination and makes the observed sign more directly attributable to the target system or objective (Ma et al., 2023, Quintero et al., 31 Mar 2026, Nie et al., 8 May 2026, Casadei, 2014, Arnold et al., 24 Jul 2025).

1. Conceptual form and recurrent mechanism

Across these works, the central structural distinction is between a varying-reference sign and a fixed-reference sign. In the varying-reference case, the sign depends on both the target object and a reference object that co-varies with the control parameters. In the fixed-reference case, the denominator, benchmark, or anchoring geometry is frozen, so the remaining parameter dependence is either entirely or much more directly attributable to the target quantity itself.

This pattern is explicit in several settings. In quantum Monte Carlo, the ordinary average sign is $Z/Z_+$, so its behavior mixes the original partition function with that of a sign-free auxiliary system. In the Oaxaca–Blinder decomposition, the sign of the explained and unexplained components depends on which group is fixed as the reference. In binary-reward GRPO, subtracting the group mean creates degenerate all-pass and all-fail groups, whereas a fixed baseline $b=0.5$ yields $A=2r-1$. In the signal-plus-background counting model, the limiting prior fixes the background at $b_0$ and thereby yields a closed-form reference prior. In temporally switching networks, fixed topology across snapshots provides a stable structural scaffold on which sign-matching can be redistributed over time (Ma et al., 2023, Quintero et al., 31 Mar 2026, Nie et al., 8 May 2026, Casadei, 2014, M et al., 11 Dec 2025).

This suggests a common mechanism. When the reference co-moves with the system, the sign can inherit thermodynamics, benchmark weights, group-composition effects, or baseline-induced degeneracies that are not intrinsic to the target object. Fixing the reference does not automatically make the sign universally optimal, but it often converts a context-sensitive indicator into a stricter probe.

2. Quantum Monte Carlo: from contaminated average sign to strict phase-transition probe

The most explicit formulation appears in the quantum Monte Carlo analysis of phase-transition detection. The conventional average sign is defined as

$$\langle \mathrm{sign} \rangle = \frac{Z}{Z_+} = \frac{\sum_C W(C)}{\sum_C |W(C)|},$$

with $Z=\sum_C W(C)$ the original partition function and $Z_+=\sum_C |W(C)|$ the partition function of a sign-free auxiliary or “reference system.” Using $F=-T\ln Z$ and $F_+=-T\ln Z_+$, the sign becomes

$$\langle \mathrm{sign} \rangle = e^{-\beta \Delta F}, \qquad \Delta F = F-F_+.$$

The key point is that a phase transition in the original system is encoded in $b=0.5$0, while the measured sign depends on $b=0.5$1. The sign can therefore be shifted, imitated, or erased by structure in the reference-system free energy $b=0.5$2 (Ma et al., 2023).

The paper’s strict criterion follows immediately: $b=0.5$3 up to the additional $b=0.5$4-dependence of $b=0.5$5 when $b=0.5$6. Hence, “The sign can exactly probe phase transition only if the free energy in the reference system is flat under variable parameters.” The same paper calls this condition “almost impossible to design,” because $b=0.5$7 is not an arbitrary constant but the partition function of a genuine sign-free model with its own thermodynamics. Earlier successes are therefore described as “survivorship bias without universality” (Ma et al., 2023).

The proposed remedy is the modified, fixed-reference sign

$b=0.5$8

where $b=0.5$9 are fixed constants. If $A=2r-1$0 is fixed, then $A=2r-1$1, and

$A=2r-1$2

The decisive consequence is

$A=2r-1$3

so second derivatives of $A=2r-1$4 reproduce second derivatives of the original free energy with no parameter-dependent contamination from the reference system. This is the precise sense in which the modified sign probes phase transitions “strictly” (Ma et al., 2023).

The same work gives both positive and negative examples for the ordinary sign. In a flat-band interacting model with long-ranged screened Coulomb interaction, the conventional $A=2r-1$5 has a minimum close to the thermodynamic transition temperature $A=2r-1$6, and the minimum approaches $A=2r-1$7 as system size grows. By contrast, in the anisotropic $A=2r-1$8-$A=2r-1$9 Heisenberg model on the square lattice, the reference free energy $b_0$0 shows a strong feature even though the original free energy $b_0$1 does not. The conventional sign and its derivative then suggest an apparent transition around $b_0$2, despite the absence of any finite-temperature phase transition by the Mermin–Wagner theorem (Ma et al., 2023).

The modified sign is then implemented in stochastic series expansion. Starting from

$b_0$3

the estimator becomes

$b_0$4

For wide parameter sweeps, the paper recommends a chain of intermediate reference points,

$b_0$5

to maintain overlap. In the frustrated spin model with $b_0$6 and $b_0$7, using $b_0$8 and $b_0$9, the computed quantity

$$\langle \mathrm{sign} \rangle = \frac{Z}{Z_+} = \frac{\sum_C W(C)}{\sum_C |W(C)|},$$0

per site shows a system-size-growing peak, consistent with the expected diverging behavior of a second-order transition. The fixed-reference advantage is therefore conceptual, mathematical, and numerical at once (Ma et al., 2023).

3. Benchmark dependence in inference and objective Bayes

A closely related issue arises in the Oaxaca–Blinder decomposition. With group-specific linear models

$$\langle \mathrm{sign} \rangle = \frac{Z}{Z_+} = \frac{\sum_C W(C)}{\sum_C |W(C)|},$$1

the mean gap $$\langle \mathrm{sign} \rangle = \frac{Z}{Z_+} = \frac{\sum_C W(C)}{\sum_C |W(C)|},$$2 can be written as

$$\langle \mathrm{sign} \rangle = \frac{Z}{Z_+} = \frac{\sum_C W(C)}{\sum_C |W(C)|},$$3

or instead as

$$\langle \mathrm{sign} \rangle = \frac{Z}{Z_+} = \frac{\sum_C W(C)}{\sum_C |W(C)|},$$4

The total gap is invariant, but the explained and unexplained components are not. Sign reversals occur when the same covariate mean difference $$\langle \mathrm{sign} \rangle = \frac{Z}{Z_+} = \frac{\sum_C W(C)}{\sum_C |W(C)|},$$5 projects with opposite signs onto $$\langle \mathrm{sign} \rangle = \frac{Z}{Z_+} = \frac{\sum_C W(C)}{\sum_C |W(C)|},$$6 and $$\langle \mathrm{sign} \rangle = \frac{Z}{Z_+} = \frac{\sum_C W(C)}{\sum_C |W(C)|},$$7, or when $$\langle \mathrm{sign} \rangle = \frac{Z}{Z_+} = \frac{\sum_C W(C)}{\sum_C |W(C)|},$$8 lies between $$\langle \mathrm{sign} \rangle = \frac{Z}{Z_+} = \frac{\sum_C W(C)}{\sum_C |W(C)|},$$9 and $Z=\sum_C W(C)$0. In the normalized bounded parameter space studied in the paper, explained-component sign flips occupy exactly half of the parameter space, while unexplained-component flips approach one half as dimension grows and exceed $Z=\sum_C W(C)$1 for even modest $Z=\sum_C W(C)$2 (Quintero et al., 31 Mar 2026).

The paper stresses that this is not merely a small-sample or misspecification artifact. In a simulated healthcare example with one covariate and correctly specified linear models, the unexplained component changes sign between references. In real ICU data on a gender mortality gap with $Z=\sum_C W(C)$3, the explained component is $Z=\sum_C W(C)$4 with women as the reference but $Z=\sum_C W(C)$5 with men as the reference. The methodological recommendation is correspondingly practical rather than doctrinal: “We recommend that data analysts using the OBD always report results using both reference groups.” The fixed-reference issue here is thus an algebraic dependence on benchmark choice, not evidence that one benchmark is normatively privileged (Quintero et al., 31 Mar 2026).

Objective Bayesian analysis of the signal-plus-background Poisson counting model exhibits the same structural contrast. With

$Z=\sum_C W(C)$6

and a Gamma prior on the nuisance background

$Z=\sum_C W(C)$7

the exact reference prior for $Z=\sum_C W(C)$8 is defined through the Fisher information of the marginal model. In the limit of perfect prior background knowledge, $Z=\sum_C W(C)$9, the reference prior becomes

$Z_+=\sum_C |W(C)|$0

which is precisely Jeffreys’ prior for the shifted Poisson mean $Z_+=\sum_C |W(C)|$1. The resulting posterior is a shifted Gamma law,

$Z_+=\sum_C |W(C)|$2

The paper’s substantive claim is that this limiting prior is both conceptually correct and practically preferable to the flat prior $Z_+=\sum_C |W(C)|$3. The flat prior is said to approximate the reference prior well only when the expected background is very high, whereas $Z_+=\sum_C |W(C)|$4 differs from the full reference prior by less than $Z_+=\sum_C |W(C)|$5 whenever $Z_+=\sum_C |W(C)|$6 is larger than about $Z_+=\sum_C |W(C)|$7, and a two-parameter fit can reproduce the reference prior with a speed-up of roughly $Z_+=\sum_C |W(C)|$8 to $Z_+=\sum_C |W(C)|$9 (Casadei, 2014).

These two cases illustrate different consequences of fixing a reference. In OBD, a fixed reference can mechanically tilt signs and must therefore be reported symmetrically. In the Poisson counting model, fixing the background at $F=-T\ln Z$0 isolates the relevant reference geometry and yields the correct limiting prior. The common thread is not that fixing a reference is always sufficient, but that the source of sign or posterior variation becomes explicit once the benchmark is made nonmoving.

4. Sign-based learning algorithms: fixed baselines, certified signs, and softened geometry

In binary-reward RLVR, the fixed-reference idea is operationalized as a replacement for group-mean centering. Standard GRPO uses

$F=-T\ln Z$1

so all-correct groups and all-wrong groups are degenerate: every advantage is exactly zero when $F=-T\ln Z$2 or $F=-T\ln Z$3. For a prompt with success probability $F=-T\ln Z$4, the degeneracy rate is

$F=-T\ln Z$5

and Jensen’s inequality implies $F=-T\ln Z$6. On logged Qwen3.5-9B GSM8K training at group size four, the paper observes a $F=-T\ln Z$7 degeneracy rate, with $F=-T\ln Z$8 all-fail and $F=-T\ln Z$9 all-pass groups. The proposed fixed-reference Sign advantage is

$F_+=-T\ln Z_+$0

equivalently a fixed baseline $F_+=-T\ln Z_+$1. This avoids gradient starvation on homogeneous groups and yields $F_+=-T\ln Z_+$2 accuracy versus $F_+=-T\ln Z_+$3 for standard normalized group-mean DrGRPO at group size four, a $F_+=-T\ln Z_+$4-point gain with $F_+=-T\ln Z_+$5. The paper further proves that the all-fail contribution performs “pass@$F_+=-T\ln Z_+$6 failure descent,” directly reducing $F_+=-T\ln Z_+$7 and increasing $F_+=-T\ln Z_+$8 (Nie et al., 8 May 2026).

In RLHF, the same paper-independent motif is reframed as advantage sign robustness. For Dr.GRPO,

$F_+=-T\ln Z_+$9

and the sign of $$\langle \mathrm{sign} \rangle = e^{-\beta \Delta F}, \qquad \Delta F = F-F_+.$$0 determines whether the policy increases or decreases the probability of completion $$\langle \mathrm{sign} \rangle = e^{-\beta \Delta F}, \qquad \Delta F = F-F_+.$$1. Reward hacking is interpreted as often arising from flipped advantage signs. Around a fixed nominal reward head $$\langle \mathrm{sign} \rangle = e^{-\beta \Delta F}, \qquad \Delta F = F-F_+.$$2, the certified sign-preservation radius is

$$\langle \mathrm{sign} \rangle = e^{-\beta \Delta F}, \qquad \Delta F = F-F_+.$$3

Sign-Certified Policy Optimization then down-weights non-robust completions through

$$\langle \mathrm{sign} \rangle = e^{-\beta \Delta F}, \qquad \Delta F = F-F_+.$$4

Unlike ensemble-style remedies, the method operates purely at policy-optimization time using the fixed RM parameters and on-policy completions (Ono et al., 3 Apr 2026).

Optimizer design uses the same vocabulary more geometrically than literally. LionMuon alternates between Lion’s sign update and Muon’s spectral matrix-sign update with a fixed period $$\langle \mathrm{sign} \rangle = e^{-\beta \Delta F}, \qquad \Delta F = F-F_+.$$5, sharing a single dual-EMA momentum buffer. In this setting, the sign step is not presented as intrinsically stronger than Muon. Rather, if Muon is taken as the stronger reference update, sign steps are cheaper tracking steps between periodic Muon refreshes. At $$\langle \mathrm{sign} \rangle = e^{-\beta \Delta F}, \qquad \Delta F = F-F_+.$$6, LionMuon is reported to Pareto-dominate Muon, Lion, Signum, and AdamW at $$\langle \mathrm{sign} \rangle = e^{-\beta \Delta F}, \qquad \Delta F = F-F_+.$$7M model size, and the same alternation effect persists at $$\langle \mathrm{sign} \rangle = e^{-\beta \Delta F}, \qquad \Delta F = F-F_+.$$8M and $$\langle \mathrm{sign} \rangle = e^{-\beta \Delta F}, \qquad \Delta F = F-F_+.$$9M scale. SignMuon, the single-EMA variant, already outperforms pure Muon (Bolatov et al., 19 May 2026).

SoftSignum and SoftMuon turn the same observation into a critique of rigid fixed-magnitude geometry. The paper describes hard sign and Muon as LMO-based methods that map every update to a fixed geometric boundary. That gives robustness early in training, but “fixed-magnitude updates can hurt terminal convergence” because they ignore parameter heterogeneity and can induce oscillation near minima. SoftSignum replaces the hard sign by

$b=0.5$00

with update

$b=0.5$01

and a quantile-based temperature schedule. Large coordinates remain in the sign-like regime, while small coordinates transition to magnitude-sensitive SGD-like behavior. SoftMuon applies the same principle to singular values,

$b=0.5$02

A plausible implication is that the fixed-reference sign advantage in optimization is regime-dependent: rigid sign geometry is useful when magnitude information is unreliable, but strict saturation can become too coarse once coordinates become heterogeneous or nearly converged (Feoktistov et al., 29 May 2026).

5. Structural fixed references in combinatorics, operator theory, and network control

In algebraic combinatorics, fixing boundary data can expose linear structure that is invisible in aggregate counts. For alternating sign matrices, the mixed row/column refinement

$b=0.5$03

counts $b=0.5$04 ASMs with $b=0.5$05s at $b=0.5$06, $b=0.5$07, and $b=0.5$08. Fischer proves the linear relation

$b=0.5$09

showing that fixing left and right columns does not create a wholly new irreducible class of refinements but can be reduced to the previously developed row-based theory. The paper presents this as “a first indication” that general fixed-boundary data may reduce to fixed top/bottom row data (Fischer, 2010).

In operator theory, fixed block decompositions play the same role. For a self-adjoint block operator matrix $b=0.5$10, positive definiteness is characterized by recursive first-kind Schur operators,

$b=0.5$11

and their iterates. The paper proves that $b=0.5$12 if and only if all recursively generated Schur operators are positive definite, while nonnegativity admits an analogous criterion under additional invertibility assumptions. Here the “advantage” of the fixed reference is that the global sign of $b=0.5$13 is replaced by a finite family of lower-dimensional sign conditions attached to a fixed recursive block structure (Orlov et al., 2010).

Temporal network herdability yields a more literal fixed-reference advantage. For temporally switching signed networks whose snapshots share the same underlying topology, the repeated digraph acts as a stable structural reference. The paper shows that when all snapshots share the same topology, complete $b=0.5$14 herdability can be achieved within two snapshots even in the presence of signed or layer dilations, provided the snapshots use different parametric realizations. By contrast, if the same topology is repeated with identical edge weights, the temporal network inherits the static non-herdability obstruction. The result therefore depends on fixed topology plus changing realization, not on temporality alone (M et al., 11 Dec 2025).

These cases are structurally disparate, but they support the same methodological point. Fixing rows, columns, blocks, or topology does not merely constrain a problem; it can reorganize sign information so that hidden linear relations, Schur-complement tests, or cross-snapshot matching become available.

6. Physical, hardware, and compression realizations, and the scope of the concept

In heavy-dense QCD, fixing baryon number rather than chemical potential yields a canonical ensemble in which the sign problem is reduced by orders of magnitude at low temperature relative to the corresponding grand-canonical ensemble. In the strong-coupling canonical formulation, integrating out the Polyakov loops leaves a partition function over baryon occupation numbers with positive weights, so the sign problem is absent. Away from strong coupling, an effective gauge action leads to a positive occupation-plus-flux representation. The fixed reference here is the conserved-charge sector itself: by projecting onto a definite baryon number, the formulation removes cancellations among irrelevant triality sectors that plague the grand-canonical trace (Bühlmann et al., 2021).

A different but related engineering realization appears in multiplier design. The arithmetic paper argues that when signed fixed-point values are concentrated around a known reference value, especially zero, explicit sign-magnitude encoding can reduce switching activity relative to two’s complement because negative near-zero values no longer carry sign-extension activity. Using a decomposed TC-input/TC-output multiplier with TC$b=0.5$15SME encoding, post-synthesis simulations of the 4-bit design show $b=0.5$16 lower switching activity at $b=0.5$17, and with reduced range $b=0.5$18 to $b=0.5$19 the reduction reaches $b=0.5$20. The benefit is therefore strongest when the data are zero-centered, while area and depth can worsen in the strictly logic-equivalent configuration (Arnold et al., 24 Jul 2025).

The sign-lock-in study reframes the reference as initialization. Across Transformers, CNNs, and MLPs, trained sign matrices resist low-rank approximation and are spectrally close to an i.i.d. Rademacher baseline, yet most signs remain inherited from initialization. The paper’s stopping-time theory shows that effective outer-to-outer sign flips have a geometric tail under bounded updates and a rare re-entry condition into a neighborhood of zero. This motivates gap-based initialization and an outward-drift regularizer, which reduce the effective flip rate to approximately $b=0.5$21 with only about a one-point increase in perplexity. A plausible implication is that initialization itself can act as a quasi-static sign reference, especially in sub-bit compression where sign bits become the storage bottleneck (Sakai et al., 19 Feb 2026).

A neighboring but distinct motif appears in even-dimensional resonance counting for Schrödinger operators with fixed-sign potentials. There the advantage comes not from a fixed reference denominator but from the monotonicity and positivity structure induced by $b=0.5$22 or $b=0.5$23, which allows determinant lower bounds and maximal-order resonance counting on nonphysical sheets. The paper is therefore relevant as a contrast case: fixed sign can play a role analogous to fixed reference, but the mechanism is operator monotonicity rather than baseline removal (Christiansen, 2013).

Taken together, these works also delimit the scope of the concept. Fixed references do not eliminate every pathology. In OBD, fixing one reference can itself reverse substantive conclusions; the remedy is to report both references. In QMC, the fixed-reference sign is strict only if sampling overlap remains adequate, and wide parameter sweeps may require intermediate bridges. In RLVR and RLHF, fixed-reference signs are highly effective in the studied regimes, but the papers do not claim universal superiority across all datasets, models, or optimization settings. The literature therefore supports a narrower encyclopedia-grade statement: fixed-reference constructions are most valuable when the moving reference is the dominant source of spurious sign variation, and least decisive when the principal difficulty lies elsewhere.