--- title: Default-Direction Asymmetry Overview url: https://www.emergentmind.com/topics/default-direction-asymmetry type: topic --- # Default-Direction Asymmetry Overview Searching arXiv for the provided topic and ids to ground the synthesis in current literature. arxiv_search(query="\"default-direction asymmetry\" OR \"entropy polarity\" OR \"boundary suppression asymmetry\" OR \"causal computational asymmetry\"", max_results=10) Default-Direction Asymmetry denotes a structurally induced bias by which one direction of update, inference, transport, steering, or observation is easier to trigger, harder to suppress, or more stably realized than its nominal opposite. Across current literature, the term does not refer to a single invariant formalism; rather, it recurs in several technically distinct settings, including reinforcement fine-tuning for large language models, post-training controllability of assistants, asymmetric routing, causal-direction identification, nonequilibrium first-passage phenomena, engineered nonreciprocity, and preferred-direction signals in astrophysical data. The common pattern is that an apparently symmetric task acquires a privileged orientation through sampling, architecture, optimization, dissipation, or observational geometry [2605.11775] [2605.27969] [2606.02136] [2602.22254]. ## 1. Conceptual structure A recurring formulation has three ingredients. First, there is a **privileged baseline direction**: for example, positive-advantage on-policy reinforcement of likely tokens, anti-underanswering assistant behavior, node-centric routing decisions, the forward causal direction, or a sky axis singled out by anisotropy. Second, there is a **mechanism that breaks effective symmetry**: sampled-token entropy mechanics, content-budget overshoot plus continuation persistence, representation–decision mismatch, residual dependence, broken detailed balance, or dipole modulation. Third, there is an **observable asymmetry**: different entropy trajectories, different suppressibility costs, unequal forward/backward hitting times, unequal transmission amplitudes, one-way steering, or direction-dependent parameter estimates [2605.11775] [2605.27969] [2007.07335] [2212.12777] [1510.00154]. This suggests a useful cross-domain distinction between mere anisotropy and default-direction asymmetry. The latter is not only a difference between two directions; it is a difference that is preferentially activated by the system’s native dynamics, objective, or measurement protocol. In RLVR, the default direction under positive advantage and on-policy sampling is entropy contraction. In assistant post-training, the default direction is anti-underanswering over-expansion. In asymmetric routing, the default direction is a node-centric scoring bias that under-expresses the currently chosen directed transition. In causal inference from optimization time, the default direction is faster convergence in the true causal orientation [2605.11775] [2605.27969] [2606.02136] [2602.22254]. The literature also makes clear that the sign of the asymmetry need not be universal. In a two-lane lattice, either forward or backward transition can be faster depending on the sign of the nonequilibrium current; in biased run-and-tumble dynamics, first-passage duality can fail at finite distance yet be restored asymptotically when a Gallavotti–Cohen symmetry holds; in driven-dissipative resonators, the preferred transport direction is set by a relative phase and can be reversed by changing that phase [2007.07335] [2505.24483] [2212.12777]. | Domain | Default direction | Structural source | |---|---|---| | RLVR entropy control | Contraction under typical rewarded on-policy updates | Sampled-token term dominates correction term | | Assistant controllability | Anti-underanswering over-expansion | Planning overshoot and continuation persistence | | Asymmetric routing | Node-centric scoring | Representation–decision mismatch | | Causal inference | Faster forward training | Reverse residual dependence and gradient-noise structure | | Nonequilibrium transport | Direction selected by current or bias | Broken detailed balance or hidden internal dynamics | | Quantum/optical transport | Phase-selected one-way flow | Interference of coherent and dissipative couplings | ## 2. Learning and control: reinforcement fine-tuning, assistant post-training, and learned defaults In RLVR for LLMs, default-direction asymmetry is formalized at the token level through a first-order entropy expansion. For a local policy entropy $\mathcal{H}_t$ and a sampled token $y_t$ with probability $p_t$, one gradient-ascent step on $A \log \pi_\theta(y_t \mid s_t)$ yields \[ \Delta\mathcal{H}_t = -\eta A\,t_1(s_t,y_t) + \eta A\,t_2(s_t) + O(\eta^2), \] with \[ t_1(s_t,y_t):=p_t(\mathcal{H}_t+\log p_t), \qquad t_2(s_t):=\sum_{v\in\mathcal{V}} p_v^2(\mathcal{H}_t+\log p_v). \] The paper defines the intrinsic entropy tendency \[ \mathcal{T}(s_t,y_t):=-t_1(s_t,y_t)+t_2(s_t), \] and the realized entropy polarity \[ \mathcal{P}(s_t,y_t,A):=A\,\mathcal{T}(s_t,y_t), \qquad \Delta\mathcal{H}_t=\eta\,\mathcal{P}(s_t,y_t,A)+O(\eta^2). \] The sampled-token term changes sign at the entropy-induced threshold $p_t=\exp(-\mathcal{H}_t)$, while $t_2(s_t)\ge 0$ is always expansive. Because on-policy sampling draws high-probability tokens more often, positive-advantage updates are structurally biased toward contraction; expansion typically requires low-probability samples or a sufficiently strong state-wise correction. This is the paper’s “default direction under $A>0$ and on-policy sampling.” Empirically, polarity tracks measured entropy change, the polarity magnitude distribution is heavy-tailed, and single-polarity ablations show that the negative branch drives exploitation whereas the positive branch preserves exploration. The proposed Polarity-Aware Policy Optimization reweights advantages by polarity sign and uses an entropy-slope phase signal; on math reasoning it improves over strong baselines, including +3.3% on AIME24 and +2.1% on AIME25 for Qwen2.5-14B, and +2.4% on Minerva and +2.2% on AMC for Qwen2.5-7B [2605.11775]. In post-trained assistants, the same idea appears as **boundary-suppression asymmetry**. The default response direction induced by anti-underanswering optimization is more complete, cautious, and proactively helpful, but that direction is harder to suppress when prompts explicitly request narrower answers. The paper operationalizes this through residual expansion under boundary-control prompts, using length-based costs such as \[ \Delta_b(\text{anti},\text{base}) = \mathbb{E}[L \mid \pi=\text{anti\_underanswer}, b] - \mathbb{E}[L \mid \pi=\text{baseline}, b]. \] On the main controlled family, the anti-underanswering policy remains substantially more expansive than baseline under matched controls: for `scope_minimal_sufficient`, mean lengths are 22.16 for baseline, 40.38 for anti, and 27.28 for minimal; for `avoid_underanswer`, 23.44, 33.78, and 28.74; and in forced-prefix continuation, 2.54, 4.08, and 2.51, with immediate-stop rates 62%, 48%, and 63%, respectively. Mechanism probes reject pure EOS failure, pure uncertainty compensation, and pure local continuation bias as complete explanations. The paper instead supports a mixed planning/stopping account in which **content-budget overshoot** and **continuation persistence** jointly make the anti-underanswering direction harder to pull back [2605.27969]. A related but distinct asymmetry appears in KL-regularized RL with a learned default policy. The objective \[ L(\pi,\pi_0)= \mathbb{E}_{\tau\sim\pi}\Big[\sum_t \gamma^t r(s_t,a_t)-\alpha \gamma^t \mathrm{KL}\big(\pi(\cdot\mid x_t)\,\|\,\pi_0(\cdot\mid D_t)\big)\Big] \] combines an **information asymmetry**, because $\pi_0$ is restricted to $D_t$, with a **divergence-direction asymmetry**, because the forward KL pulls the full policy toward the default. The default learns broad, reusable behaviors precisely because it sees less information, while the main policy deviates only when reward justifies it. The paper links this to an information-bottleneck view and reports faster learning and improved performance in tasks with repeated structure [1905.01240]. ## 3. Inference and optimization: routing decoders, convergence-time causality, and topological residual geometry In neural asymmetric routing, default-direction asymmetry is formulated as a decoder-side failure to expose the decision-critical directed transition. For ATSP, the Bellman-style action value is \[ Q^*(j \mid s_t) = D_{c_t,j} + V^*(j,\,U_t\setminus\{j\}), \qquad V^*(i,\emptyset) = D_{i,s_{\text{start}}}, \] yet the RADAR-style baseline decoder scores candidates primarily through context–node compatibility, \[ s_t^{\mathrm{base}}(j)=\frac{z_t^\top k_j}{\sqrt d}. \] The paper identifies a representation–decision mismatch: pairwise directed information may be encoded upstream, but the final score remains node-centric. Its edge-aware decoder adds explicit local edge, closure, and static lookahead terms through a candidate descriptor $\phi_t(j)$ and a learned bias $b_t(j)=\mathrm{MLP}_{\mathrm{edge}}(\phi_t(j))$, yielding $\tilde s_t(j)=s_t^{\mathrm{base}}(j)+b_t(j)$. On a controlled SVD/Sinkhorn asymmetric backbone, this reduces the ATSP-1000 gap from 4.13% to 2.73%; removing the local-edge term degrades the ATSP-1000 gap to 5.04%, worse than the RADAR reference, which sharpens the claim that explicit exposure of the current directed edge is the decisive corrective signal [2606.02136]. In bivariate causal discovery, the paper on Causal Computational Asymmetry posits an optimization-time default direction. Under the additive-noise model $Y=f(X)+\varepsilon$ with $\varepsilon \perp X$ and $f$ nonlinear and injective, one trains parity-matched networks in both directions and compares the hitting times \[ T_{\mathrm{fwd}}(L_{\mathrm{thr}})= \inf\{t:\mathcal{L}_{\mathrm{fwd}}(\theta_t)\le L_{\mathrm{thr}}\}, \qquad T_{\mathrm{rev}}(L_{\mathrm{thr}})= \inf\{t:\mathcal{L}_{\mathrm{rev}}(\phi_t)\le L_{\mathrm{thr}}\}. \] The CCA score is \[ \mathrm{CCA}(X\to Y)= T_{\mathrm{fwd}}(L_{\mathrm{thr}}) - T_{\mathrm{rev}}(L_{\mathrm{thr}}). \] The theoretical asymmetry arises because forward residuals converge to independent noise, whereas reverse residuals remain statistically dependent on the input, which induces a higher irreducible loss floor and non-separable gradient noise. Under local PL conditions, the paper states that \[ \mathbb{E}[T_{\mathrm{rev}}(L_{\mathrm{thr}})] > \mathbb{E}[T_{\mathrm{fwd}}(L_{\mathrm{thr}})]. \] Z-scoring of both variables is mandatory, since otherwise scale alone can invert the convergence ordering. Empirically, the method achieves 26/30 correct synthetic causal identifications across six architectures, including 30/30 on sine and exponential data-generating processes [2602.22254]. A second causal-direction method, Topological Residual Asymmetry, makes the asymmetry geometric rather than optimization-dynamic. After cross-fitted regression in both directions and rank-based copula standardization, the correct-direction regressor–residual cloud is approximately two-dimensional, whereas the wrong-direction cloud concentrates near a one-dimensional tube, especially in a small-noise regime. TRA measures this bulk–tube contrast with a 0D persistent-homology functional computed from Euclidean MST edge-length profiles. The signed score \[ \Delta_n = \overline{TP}^{[\alpha_n,\beta_n]}_0(\widetilde R(Y\mid X)) - \overline{TP}^{[\alpha_n,\beta_n]}_0(\widetilde R(X\mid Y)) \] is positive in the forward direction under the paper’s consistency regime, while TRA-s extends the method to fixed noise by binning reverse residuals, and TRA-C adds a confounding-aware abstention rule calibrated by a Gaussian-copula plug-in bootstrap [2602.00427]. ## 4. Nonequilibrium transport and first-passage asymmetry In stochastic transport, default-direction asymmetry is often literal: forward and backward transition or first-passage times cease to be interchangeable once hidden structure or nonequilibrium currents are present. In a two-lane lattice random walk with rates $u,w,u',w',a,b$, microscopic reversibility no longer forces equal forward and backward transition times, because parallel pathways and inter-lane switching create state-space cycles with thermodynamic affinity \[ A = k_B T \ln\!\left(\frac{u w'}{u' w}\right). \] The mean forward and backward transition times, $T_f$ and $T_b$, coincide only at equilibrium, \[ u w' = u' w \iff A=0 \iff T_f=T_b. \] For the illustrative parameter set $u=1$, $u'=4$, $w'=1$, $a=b=1$, varying $w$ changes the sign of the asymmetry: $w<1/4$ gives $T_f>T_b$, $w=1/4$ restores symmetry, and $w>1/4$ gives $T_f