The Gate Always Closes: On Injecting Auxiliary Signals into Frozen Vision-Language Models

• URL: http://arxiv.org/abs/2607.23335v1
• Date: Sat, 25 Jul 2026 19:14:34 GMT
• Title: The Gate Always Closes: On Injecting Auxiliary Signals into Frozen Vision-Language Models
• Abstract summary: We find that across five injection designs, every gated pathway becomes behaviourally closed, with accuracy invariant to ablating the pathway at inference even when the gate parameter would nominally pass 30-45% of the signal.<n>We attribute this suppression phenomenon to two regimes, a dead-gradient regime formalised through the caption-invariance of image-derived signals, and a negative-utility regime in which the auxiliary signal actively hurts the loss.
• Score: 18.625239232550932
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Auxiliary signal pathways in VLMs are routinely fitted with learnable gates so the optimiser can decide how much of the signal to admit. We find that the optimiser almost always decides on zero: across five injection designs, every gated pathway becomes behaviourally closed, with accuracy invariant to ablating the pathway at inference even when the gate parameter would nominally pass 30-45% of the signal. We attribute this suppression phenomenon to two regimes, a dead-gradient regime formalised through the caption-invariance of image-derived signals, and a negative-utility regime in which the auxiliary signal actively hurts the loss. Rather than fight suppression, we exploit it: we regularise LoRA fine-tuning with geometric auxiliary losses from hyperbolic visual relational graphs (IoA-driven entailment cones and angular repulsion on the Lorentz manifold), coupled only through the forward pass at training time and dropped at inference. Disaggregating GQA by question type exposes a clean dissociation. Three configurations without geometric losses at inference lose 2.85-3.39pp on relational questions while gaining ~1pp on attribute questions; a fourth that trains with the losses but infers through a soft prompt loses 5.14pp on rel for only +0.23pp on attr, so training-time regularisation alone does not protect relational accuracy without a geometric inference pathway. Configurations that keep the geometric pathway at inference preserve vanilla-level relational accuracy and match the attribute gain. Out of distribution on VSR, the RMS-prefix recipe preserves the spatial signal; stripping the geometric losses (G2) collapses VSR by 4.6pp, isolating them as the OOD source. A secondary result: embedding-norm alignment is necessary for generation-safe prefix injection, and learnable gates should be replaced with fixed, non-optional injection at matched scales.

Related papers

• Sensitivity, Causality, and Repair Dissociate: A Layer-Wise Analysis of Perturbation Robustness and Its Scaling [0.0]
  We show that when a language model fails on surface-perturbed input, "which layer is responsible" has three natural operationalizations.<n>We identify two propagation regimes - spike-and-suppress and late-accumulation.<n>A fixed-harness layer sweep across four models (3.8-8B) confirms the core prediction on chain-of-thought GSM8K.
  arXiv  Detail & Related papers  (2026-08-04T15:48:55Z)
• GEMS: Geometric Constraints Enable Multi-Semantic Superposition in LLMs [5.018957704056544]
  Activation steering controls model behavior by modifying intermediate hidden states at inference time without retraining.<n>We show that when multiple semantic directions are superposed without constraints, the model collapses.<n>We propose GEMS, a training-free method that maps each source to a corresponding geometric constraint.
  arXiv  Detail & Related papers  (2026-06-18T08:43:33Z)
• R-APS: Compositional Reasoning and In-Context Meta-Learning for Constrained Design via Reflective Adversarial Pareto Search [18.8323743697237]
  Large language models (LLMs) are fluent on open-ended tasks, yet in agentic settings, fluency does not ensure reliable delivery.<n>We trace this gap to three coupled structural failures: errors propagate without localization, worst-case perturbations go unevaluated, and accumulated knowledge is never invalidated.<n>We argue these share a root cause: abductive, counterfactual, meta-inductive, corrective, and inductive reasoning pull a shared context in incompatible directions.
  arXiv  Detail & Related papers  (2026-06-03T12:45:39Z)
• Coverage Analysis of Rydberg Atom Quantum Receiver Arrays: A Stochastic Geometry Approach [55.47164192369501]
  Rydberg atomic quantum receivers (RAQRs) offer quantum-limited sensitivity and broadband tunability.<n>In dense deployments, aggregate interference can push atomic transducer out of its small-signal regime.<n>This paper addresses the question by embedding the RAQR front end into a geometry coverage analysis.
  arXiv  Detail & Related papers  (2026-05-22T04:05:13Z)
• The Geometric Inductive Bias of Grokking: Bypassing Phase Transitions via Architectural Topology [0.0]
  We study grokking - delayed generalization in Transformers trained on cyclic modular addition (Zp)<n>We identify two independent structural factors in standard Transformers: representational magnitude and data-dependent attention routing.
  arXiv  Detail & Related papers  (2026-03-05T14:41:01Z)
• Routing Absorption in Sparse Attention: Why Random Gates Are Hard to Beat [0.0]
  When sparse attention is trained end-to-end, the model's Q/K/V projections co-adapt to whatever mask is imposed.<n>Differentiable soft gating converges to nearly the same perplexity whether the gate is learned or random.<n>Experts co-adapt to any router, but show that attention exhibits a structurally more severe form.
  arXiv  Detail & Related papers  (2026-02-11T15:06:44Z)
• Generalizing GNNs with Tokenized Mixture of Experts [75.8310720413187]
  We show that improving stability requires reducing reliance on shift-sensitive features, leaving an irreducible worst-case generalization floor.<n>We propose STEM-GNN, a pretrain-then-finetune framework with a mixture-of-experts encoder for diverse computation paths.<n>Across nine node, link, and graph benchmarks, STEM-GNN achieves a stronger three-way balance, improving robustness to degree/homophily shifts and to feature/edge corruptions while remaining competitive on clean graphs.
  arXiv  Detail & Related papers  (2026-02-09T22:48:30Z)
• Geometric-Disentangelment Unlearning [106.99160454669902]
  gradient ascent on forget samples often harms retained knowledge.<n>We propose the Geometric-disment Unlearning (GU) that decomposes any candidate forget gradient update into tangential and normal components to retain space and executes only the normal component.<n>Our method is plug-and-play and can be attached to existing gradient-based unlearning procedures to mitigate side effects.
  arXiv  Detail & Related papers  (2025-11-21T09:58:25Z)
• What Layers When: Learning to Skip Compute in LLMs with Residual Gates [66.23658560048241]
  GateSkip is a residual-stream gating mechanism that enables token-wise layer skipping in decoder-only LMs.<n>Each Attention/MLP branch is equipped with a sigmoid-linear gate that condenses the branch's output before it re-enters the residual stream.
  arXiv  Detail & Related papers  (2025-10-13T16:31:50Z)
• Two failure modes of deep transformers and how to avoid them: a unified theory of signal propagation at initialisation [8.973965016201822]
  Finding the right initialisation for neural networks is crucial to ensure smooth training and good performance.<n>In transformers, the wrong initialisation can lead to one of two failure modes of self-attention layers: rank collapse, where all tokens collapse into similar representations, and entropy collapse, where highly concentrated attention scores lead to instability.<n>Here, we provide an analytical theory of signal propagation through deep transformers with self-attention, layer normalisation, skip connections and gradients.
  arXiv  Detail & Related papers  (2025-05-30T08:18:23Z)
• Robust Self-Supervised LiDAR Odometry via Representative Structure Discovery and 3D Inherent Error Modeling [67.75095378830694]
  We develop a two-stage odometry estimation network, where we obtain the ego-motion by estimating a set of sub-region transformations. In this paper, we aim to alleviate the influence of unreliable structures in training, inference and mapping phases. Our two-frame odometry outperforms the previous state of the arts by 16%/12% in terms of translational/rotational errors.
  arXiv  Detail & Related papers  (2022-02-27T12:52:27Z)
• The KFIoU Loss for Rotated Object Detection [115.334070064346]
  In this paper, we argue that one effective alternative is to devise an approximate loss who can achieve trend-level alignment with SkewIoU loss. Specifically, we model the objects as Gaussian distribution and adopt Kalman filter to inherently mimic the mechanism of SkewIoU. The resulting new loss called KFIoU is easier to implement and works better compared with exact SkewIoU.
  arXiv  Detail & Related papers  (2022-01-29T10:54:57Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.