ADu2021/dmlr-latent-reasoning
skills/skillxiv-v0.0.2-claude-opus-4.6/dmlr-latent-reasoning/SKILL.md
Improve multimodal reasoning at test-time through confidence-guided latent optimization without retraining. Iteratively refine learnable latent think tokens via policy gradient using confidence reward. Dynamically select and update relevant image patches based on internal confidence levels. Maintain high efficiency with all optimization in latent space.
$ install --global
skillsauthnpx skillsauth add ADu2021/skillXiv dmlr-latent-reasoningInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
Security Scan Results
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
3
Scanners Passed
9
Scanners in report
Clean
TrivyContainer and dependency vulnerability scanner
95%
Clean
SemgrepStatic code analysis for vulnerabilities
95%
Clean
mcp-scan (Snyk)Model Context Protocol security validation
95%
Skipped
Snyk (dep)Open source security scanning
50%
Skipped
Socket.devSupply chain security analysis
50%
Skipped
VirusTotalMulti-engine malware detection
50%
Skipped
CrowdStrikeAdvanced threat intelligence
50%
Skipped
OSV-ScannerOpen Source Vulnerability database check
50%
Skipped
OWASP Dep-Check
50%
Last scanned: Apr 17, 2026, 5:33 AM7.9s1 file scanned
SKILL.md
- name:
- dmlr-latent-reasoning
- title:
- Reasoning Within the Mind: Dynamic Multimodal Latent Reasoning with Confidence-Guided Visual Injection
- version:
- 0.0.2
- engine:
- skillxiv-v0.0.2-claude-opus-4.6
- license:
- MIT
- url:
- https://arxiv.org/abs/2512.12623
- keywords:
- [multimodal-reasoning, latent-space, test-time-optimization, confidence-guided, vision-language]
- description:
- Improve multimodal reasoning at test-time through confidence-guided latent optimization without retraining. Iteratively refine learnable latent think tokens via policy gradient using confidence reward. Dynamically select and update relevant image patches based on internal confidence levels. Maintain high efficiency with all optimization in latent space.
Skill Summary
DMLR introduces a test-time framework improving multimodal reasoning without requiring model retraining. The approach uses learnable latent think tokens iteratively refined through policy gradient updates guided by confidence reward signals (truncated entropy over top-k probable tokens). Dynamic visual injection strategy autonomously determines which image patches to integrate at each optimization step based on internal confidence levels. The framework demonstrates that visual information is needed only at specific reasoning stages (not uniformly) and internal confidence correlates strongly with reasoning quality and visual grounding accuracy. All optimization occurs within latent space, maintaining high inference efficiency.
When To Use
- Improving multimodal model reasoning without retraining
- Scenarios where visual grounding helps at specific reasoning stages
- Projects with test-time compute budget for iterative refinement
- Research on confidence-guided multimodal reasoning
When NOT To Use
- Strict real-time inference requiring immediate answers
- Applications benefiting from uniform visual context throughout
- Scenarios where test-time optimization overhead isn't justified
- Domains where static models already perform well
Core Technique
Three key innovations enable confidence-guided latent reasoning:
1. Confidence-Guided Latent Optimization Use learnable "latent think tokens" that are iteratively refined through policy gradient updates. Guidance comes from confidence reward signal based on "truncated entropy over top-k probable tokens." This reward captures model's internal certainty about reasoning.
2. Dynamic Visual Injection Strategy Rather than injecting visual information at fixed positions, "dynamically select and update the most relevant image patches at each optimization step." The model autonomously determines:
- Which visual patches to incorporate
- When to incorporate them
- How much visual information helps at each step
Driven by internal confidence levels, not fixed heuristics.
3. Efficiency Considerations Maintain high inference efficiency by performing "all optimization within latent space, avoiding expensive explicit text generation or external tool calls." This keeps computational overhead reasonable despite iterative refinement.
Empirical Findings
- Visual information needed only at specific reasoning stages (not uniformly)
- Internal confidence correlates strongly with reasoning quality
- Confidence also correlates with visual grounding accuracy
- Consistent improvements across seven benchmarks
- Works across multiple model architectures
Implementation Notes
Implement learnable latent think token representation. Set up policy gradient optimization with confidence-based rewards. Implement dynamic visual patch selection and update mechanism. Run iterative refinement at test time only. Monitor confidence signals to understand when/which visual information helps reasoning.
References
- Original paper: Reasoning Within the Mind (Dec 2025)
- Latent space optimization for language models
- Multimodal reasoning and grounding
Related Skills
ADu2021/flow-map-trajectory-tilting
testing
VerifiedTrustedCommunity
Uses flow maps as look-ahead operators to enable principled reward-guided diffusion by predicting trajectory endpoints at any denoising step. Deploy when applying rewards or preferences to diffusion trajectories with meaningful gradients throughout generation.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flexible-data-mixture-of-experts
testing
VerifiedTrustedCommunity
Train language models where each expert learns independently on closed datasets, enabling flexible inference with selective data inclusion or exclusion. 41% performance improvement while allowing users to opt out of specific data sources without retraining.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flexibility-trap-diffusion-reasoning
data-ai
VerifiedTrustedCommunity
Understand how token generation flexibility in diffusion LMs paradoxically constrains reasoning, as models exploit ordering flexibility to avoid uncertain tokens, and apply simplified approaches that preserve parallel decoding benefits. Use when optimizing diffusion-based language models for reasoning tasks.
2SKILL.mdUpdated Apr 17, 2026
ADu2021/flex-continuous-agent-evolution
devops
VerifiedTrustedCommunity
Enable LLM agents to improve continuously during deployment by constructing structured experience libraries through self-reflection on successes and failures—achieving 23% improvement on reasoning without gradient-based parameter updates or external training.
2SKILL.mdUpdated Apr 17, 2026