Direction Binding Gap

• Direction binding gap is a unifying concept where directional variables control the existence or magnitude of gaps and binding relations across diverse systems.
• In graphene electronics and plasmonic binding, directional control governs phenomena such as band gaps, conduction gaps, and stable nanoparticle configurations.
• In language models and translation surfaces, directional cues—from activation-space geometry to angular arrangements—critically determine binding accuracy and gap scaling.

Searching arXiv for the cited papers to ground the article in the current literature. “Direction binding gap” (Editor’s term) can be used as a compact umbrella for several arXiv-described phenomena in which the existence, magnitude, or repair of a gap or binding relation is controlled by direction, orientation, or a structured association. In the cited literature, the operative “direction” ranges from real-space strain orientation and momentum-space Dirac-cone displacement in graphene, to the preferred axis of plasmon-mediated optical binding, to a low-rank activation-space direction governing entity–attribute association in LLMs, to modality-specific failures of entity tracking in speech reasoning, and finally to the angular spacing of saddle-connection directions on translation surfaces (Naumov et al., 2011, Nguyen et al., 2014, Nguyen et al., 2015, Kostina et al., 2017, Dai et al., 2024, Hsu et al., 3 Jun 2026, Athreya et al., 2010).

1. Cross-domain structure of the concept

Across these works, the common structure is not a single physical mechanism but a recurring dependency: a system remains gapless, weakly bound, or binding-imprecise unless a directional variable takes the appropriate value. In graphene, the relevant variable can be the corrugation direction, the strain angle $\theta$, the transport angle $\phi$, or the transverse momentum offset between Dirac cones. In plasmonic optical binding, the decisive variable is the direction of dipole polarization relative to the substrate-mediated surface-wave channel. In LLMs, the decisive variable is an activation-space direction that encodes Ordering ID (OI). In speech reasoning, the relevant failure is localized to entity tracking rather than uniformly distributed across tasks. In translation-surface dynamics, the quantity of interest is the smallest angular gap among saddle-connection directions.

Domain	Directional variable	Observable
Graphene electronics	corrugation direction, $\theta$, $\phi$, or $k_y$ misalignment	global band gap, conduction gap, or transmission gap
Plasmonic dimers	polarization direction and SPP propagation axis	stable optical binding geometry
LM and SLLM reasoning	OI-PC direction or explicit entity enumeration	correct entity–attribute binding
Translation surfaces	saddle-connection directions	smallest angular gap $\gamma^\omega(R)$

This suggests that “direction binding gap” is best understood as a family resemblance across disciplines: direction can mean spatial direction, momentum-space displacement, hidden-state geometry, or literal angular direction, but in each case directional structure governs whether a forbidden window, a stable channel, or a correct binding relation exists.

2. Direction-controlled electronic gaps in graphene

In strained graphene, the strongest directional dependence appears in the contrast between homogeneous deformation, periodic corrugation, and heterojunction transport. Ab-initio calculations show that uniform deformation either leaves graphene (semi)metallic or opens up a small gap yet only beyond the mechanical breaking point of the graphene, contrary to claims based on tight-binding calculations. By contrast, a sine-like one-dimensional inhomogeneous deformation,

$$u_z(\mathbf r)=A\sin(\mathbf k\cdot \mathbf r+\varphi),$$

can open a global gap when applied along any direction but the armchair one, with the largest gap for corrugation along the zigzag direction, reaching about $0.5$ eV without electrostatic gating. When inversion symmetry is preserved, the gap opening has a threshold character with $(A/\lambda)_c \sim 0.1,$ and rises very sharply once this ratio is exceeded; when inversion symmetry is broken, the opening is threshold-less, although initially tiny for small $A/\lambda$ (Naumov et al., 2011).

The same directional logic reappears in graphene strain junctions, where the central observable is not an intrinsic bulk bandgap but a conduction gap defined from transport. In the Green’s-function formulation,

$\phi$0

and the conduction gap is the energy window around the Fermi level where the conductance vanishes or is strongly suppressed. The mechanism is the shift of Dirac points in $\phi$1-space between unstrained and strained regions, not a true bandgap opening in uniformly strained graphene. For transport along the armchair direction $\phi$2 at $\phi$3, the reported conduction gaps are about $\phi$4 meV at $\phi$5, about $\phi$6 meV at $\phi$7, and about $\phi$8 meV at $\phi$9; at $\theta$0, the conduction gap can reach about $\theta$1 meV, while vanishing near specific intermediate angles where the strained and unstrained Dirac points align in the same transverse momentum channel (Nguyen et al., 2014).

A closely related momentum-space filtering effect occurs in vertical devices made of two incommensurately misoriented graphene layers. There, the relevant classification is whether the two rotated lattices belong to the same or different orientation classes for a fixed transport axis. When the layers are of different lattice classes, the Dirac cones lie at different $\theta$2 points, low-energy states in one layer cannot efficiently couple to states in the other, and a significant transmission gap of a few hundreds of meV appears. Near the Dirac point, the paper gives the estimate

$\theta$3

with reported examples including $\theta$4 meV, $\theta$5 meV, and $\theta$6 meV; strain engineering can enlarge the gap further (Nguyen et al., 2015).

A persistent misconception in this literature is that any sufficiently large strain should open a practically useful graphene gap. The ab-initio result is more restrictive: under realistic tensile strain the $\theta$7 band drops toward the Fermi level and suppresses the gap that Dirac-point merging would otherwise produce, so practical homogeneous deformation does not open a useful gap (Naumov et al., 2011). The junction and vertical-device results therefore do not contradict the gapless nature of bulk graphene; they exploit direction-dependent mismatch and state filtering rather than uniform semiconducting conversion.

3. Direction-selective optical binding near plasmonic substrates

In optical binding, the directional effect concerns the geometry of mechanically stable nanoparticle configurations. For a dimer of dielectric nanoparticles next to a flat metallic substrate, the interaction is governed not mainly by free-space photon exchange but by interference of surface plasmon polariton (SPP) waves launched along the metal surface. The effective SPP wavelength,

$\theta$8

rather than the free-space wavelength, sets the spatial period of the lateral force landscape, allowing stable separations significantly below the diffraction limit (Kostina et al., 2017).

The decisive directional result is that stable binding occurs in the direction of the dipole polarization, in contrast to vacuum optical binding, where the stable configuration is formed perpendicularly to the polarization of the dipole moments. The paper attributes this reversal to the vectorial structure and anisotropic scattering diagram of the SPP channel. For a silver/vacuum interface with $\theta$9 nm dielectric nanoparticles at $\phi$0–$\phi$1 nm, the SPP resonance occurs around $\phi$2 nm, and the dynamics figure reports equilibrium attraction toward positions such as $\phi$3 nm at $\phi$4 (Kostina et al., 2017).

The substrate-mediated channel also changes force scale. When the SPP channel is removed and only free-space photons are kept, the interaction force becomes about one order of magnitude weaker, and the trap stiffness is likewise enhanced by roughly one order of magnitude in the SPP-mediated case. In this literature, the relevant “gap” is not a bandgap but what the paper summary describes as a directional anisotropy gap between two binding channels: vacuum photon binding favors the transverse direction, whereas the plasmonic interface opens a stable binding channel along the polarization direction (Kostina et al., 2017).

This directional reversal is significant because it shows that a substrate can qualitatively reconfigure the allowable binding geometry rather than merely perturbing an existing potential minimum. The result is an anisotropic organization principle with different periods along and across the field polarization.

4. Activation-space directions and binding in LLMs

In language-model interpretability, the relevant direction is neither spatial nor momentum-space, but a low-rank direction in hidden-state geometry. The central claim of “Representational Analysis of Binding in LLMs” is that the previously hypothesized Binding ID (BI), introduced in earlier work by Feng & Steinhardt, is better understood as a manifestation of Ordering ID (OI). For a relation-specific activation matrix

$\phi$5

the first principal component, or a small number of leading components, defines a low-rank OI subspace with basis

$\phi$6

The paper often treats the first principal component as the main OI encoding direction, denoted OI-PC (Dai et al., 2024).

Layer-wise PCA shows that middle layers contain the clearest OI structure, especially around layer $\phi$7 in Llama2-7B. Early layers are described as more tangled, middle layers as the point where OI becomes linearly visible, and later layers as less directly interpretable for this feature. The authors then move from correlation to causality by editing hidden states along the OI direction: $\phi$8 Under this intervention, the model’s predicted binding changes in a controlled manner. In the illustrative example, “The coffee is in Box Z, the stone is in Box M, the map is in Box H,” the default answer to “Box Z contains the …” is coffee; after one step along OI-PC, the answer tends to become stone; after two steps, map (Dai et al., 2024).

The paper further argues that this is not merely position encoding. Using a pseudo relation dataset, a filler-word dataset, and an interjection dataset, it reports that the OI-PC remains strongly correlated with OI while its correlation with Position Information (PI) is near zero or much smaller. A distance-based classifier using the first OI PC also separates related entity–attribute pairs better than other principal components. The phenomenon is reported across Llama2-7B, Llama3-8B, Qwen1.5-7B, Pythia-6.9B, and Float-7B, with even stronger intervention effects in the code-fine-tuned Float-7B model (Dai et al., 2024).

Within this framework, a binding failure is interpretable as a failure to recover or correctly use the OI-coded hidden-state direction. The “direction binding gap” in this setting is therefore a representational one: the model’s success at entity–attribute association depends on access to, and manipulation of, a specific low-rank direction in activation space.

5. Modality-specific entity binding failure in speech reasoning

The speech-LLM literature sharpens the notion of a binding gap by showing that a modality gap need not be uniform across reasoning tasks. “Entity Binding Failures in Speech LLM Reasoning: Diagnosis and Chain-of-Thought Intervention” compares speech-to-text (S2T) and text-to-text (T2T) performance on four BBH/VoiceBench categories with $\phi$9 items each, for a total of $k_y$0 items. The evaluated models are Qwen2.5-Omni-7B and Phi-4-Multimodal. On hyperbaton (HYP), navigate (NAV), and sports understanding (SPO), S2T is comparable to T2T and sometimes better. On web of lies (WOL), however, S2T collapses to chance while T2T remains high: Qwen2.5-Omni reports $k_y$1 on WOL, and Phi-4-MM reports $k_y$2 (Hsu et al., 3 Jun 2026).

The paper diagnoses this as an entity binding failure rather than a general reasoning deficit. Its explanation is that speech encoders rely on temporal pooling and downsampling, which preserve broad semantic content but blur fine-grained acoustic or token boundaries. The model may therefore retain the overall story while losing precise entity–property associations during implicit reasoning. The authors explicitly distinguish this from a pure ASR-noise account: in a name-corruption test on T2T web-of-lies, even $k_y$3 name corruption reduces accuracy by only $k_y$4 pp, far less than the S2T gap (Hsu et al., 3 Jun 2026).

To repair the failure, the paper proposes Entity-Aware Chain-of-Thought (EA-CoT), which for WOL instructs the model to perform entity enumeration, claim recording, step-by-step reasoning, and answer extraction. The generation budget is expanded from $k_y$5 tokens to $k_y$6 tokens, but the token-budget control shows that increasing tokens alone yields no meaningful speech improvement, with $k_y$7 pp. The gains come from the instruction itself. Task-specific CoT improves speech accuracy by $k_y$8 pp overall for Qwen and $k_y$9 pp overall for Phi-4; on WOL, Qwen S2T improves from $\gamma^\omega(R)$0 to $\gamma^\omega(R)$1 ($\gamma^\omega(R)$2 pp), and Phi-4 S2T improves from $\gamma^\omega(R)$3 to $\gamma^\omega(R)$4 ($\gamma^\omega(R)$5 pp). The paper states that these gains are significant by McNemar’s test (Hsu et al., 3 Jun 2026).

The ablations localize the source of improvement. On Qwen speech WOL, the sequence Baseline $\gamma^\omega(R)$6, + Format only $\gamma^\omega(R)$7, + Step-by-step $\gamma^\omega(R)$8, + Entity enum $\gamma^\omega(R)$9, and Full EA-CoT $$u_z(\mathbf r)=A\sin(\mathbf k\cdot \mathbf r+\varphi),$$0 implies that entity enumeration is the single largest contributor and accounts for $$u_z(\mathbf r)=A\sin(\mathbf k\cdot \mathbf r+\varphi),$$1 of the full EA-CoT effect. A generic “let’s think step by step” prompt yields only $$u_z(\mathbf r)=A\sin(\mathbf k\cdot \mathbf r+\varphi),$$2 pp, and EA-CoT does not help on MMSU, where the crucial cues are acoustic and are lost in text (Hsu et al., 3 Jun 2026).

The significance of these results is conceptual as much as empirical. The speech modality gap is reframed as a localized, modality-induced entity-binding bottleneck. Speech preserves enough global meaning for spatial, syntactic, and factual reasoning, but continuous speech representations interfere with stable symbolic anchoring on entity-centric logical tasks. The remedy is explicit text-space binding before downstream inference.

6. Angular gap phenomena on translation surfaces

A more literal directional-gap problem appears in the geometry of translation surfaces. For a normalized holomorphic differential $$u_z(\mathbf r)=A\sin(\mathbf k\cdot \mathbf r+\varphi),$$3 on a compact Riemann surface of genus $$u_z(\mathbf r)=A\sin(\mathbf k\cdot \mathbf r+\varphi),$$4, each saddle connection $$u_z(\mathbf r)=A\sin(\mathbf k\cdot \mathbf r+\varphi),$$5 determines a holonomy vector

$$u_z(\mathbf r)=A\sin(\mathbf k\cdot \mathbf r+\varphi),$$6

and hence a direction $$u_z(\mathbf r)=A\sin(\mathbf k\cdot \mathbf r+\varphi),$$7. Writing the directions of saddle connections of length at most $$u_z(\mathbf r)=A\sin(\mathbf k\cdot \mathbf r+\varphi),$$8 in increasing order as

$$u_z(\mathbf r)=A\sin(\mathbf k\cdot \mathbf r+\varphi),$$9

the smallest angular gap is

$0.5$0

Masur’s counting theorem implies quadratic growth in the number of such directions, so the natural comparison scale is $0.5$1 (Athreya et al., 2010).

The main theorem is that for $0.5$2-almost every $0.5$3 in any stratum,

$0.5$4

Equivalently, the smallest gap is little-$0.5$5 of $0.5$6. The paper also proves that for every $0.5$7, a positive proportion of the gaps are smaller than $0.5$8. The proof proceeds through $0.5$9 equidistribution and a thin-wedge counting problem, replacing the angular-spacing question by counts of saddle-connection vectors in a trapezoidal region after rescaling (Athreya et al., 2010).

The exceptional set is characterized exactly. A translation surface $(A/\lambda)_c \sim 0.1,$0 is a lattice surface if and only if it has no small gaps, meaning there exists $(A/\lambda)_c \sim 0.1,$1 such that

$(A/\lambda)_c \sim 0.1,$2

The converse uses the Smillie–Weiss characterization of lattice surfaces via the absence of small triangles. The paper also explains the billiard motivation: after unfolding a rational polygon, generalized diagonals correspond to saddle connections, so the directional gap problem becomes a statement about how close in angle generalized diagonals can be in rational billiards (Athreya et al., 2010).

In this geometric setting, the phrase “direction binding gap” would refer not to energetic or representational binding, but to the fine-scale clustering of allowed directions themselves. The result is nonetheless structurally parallel to the other literatures: directional organization is not uniform, generic systems exhibit unexpectedly small directional separations, and exceptional highly structured systems retain a robust gap at the natural scaling.