Latent-Implicit Thinking with Proof-Carrying Neuro-Symbolic Outputs for Biomedical Discovery

David Scott Lewis, Enrique Zueco

Track: long paper (up to 10 pages)

Keywords: latent reasoning, implicit thinking, latent-to-symbolic compilation, proof-carrying outputs, neuro-symbolic verification, structured causal graphs, mechanistic constraints, warrant emission, constraint satisfaction, token efficiency, chain-of-thought alternatives, continuous thought evolution, symbolic extractor, verification certificates, faithfulness probing, linear probes, causal direction decoding, scientific discovery, multi-domain benchmarks, NAD+/AD case study, compressed interpretability, neural ODE framing

TL;DR: LaSy compiles latent reasoning into proof-carrying outputs: causal graphs, constraints, warrants. Keeps efficiency (45 tokens vs 319 CoT) and raises constraint satisfaction to 93% (vs 63%), enabling falsifiable NAD+/AD hypotheses.

Abstract: Recent work on latent reasoning—where large language models (LLMs) perform intermediate computation in continuous representation spaces rather than generating explicit token chains—achieves dramatic efficiency gains (80–90% token reduction) but sacrifices the transparency that makes chain-of-thought (CoT) reasoning auditable. We propose Latent-to-Symbolic Compilation (LaSy), a four-component pipeline that enables models to reason efficiently in latent space while emitting proof-carrying structured outputs: causal graphs with typed edges, mechanistic constraints, and minimal verification warrants. The pipeline comprises a latent reasoner for continuous thought evolution, a symbolic extractor that decodes latent states into formal graph structures, a constraint verifier that checks domain axioms, and a warrant emitter that produces sparse evidence certificates. We evaluate LaSy on 50 reasoning tasks across three scientific domains and demonstrate that it matches latent-only efficiency (45 tokens vs. 319 for explicit CoT) while achieving 93% constraint satisfaction—compared to 63% for unverified latent reasoning. In a case study on NAD+-centered Alzheimer’s disease reversal, LaSy discriminates between three competing mechanistic hypotheses and generates falsifiable experimental proposals. Faithfulness probing reveals that latent states encode semantically meaningful structure (89.6% linear probe accuracy for causal direction), providing evidence that implicit reasoning is not opaque but rather compressed.

Presenter: ~David_Scott_Lewis1

Format: Yes, the presenting author will definitely attend in person because they attending ICLR for other complementary reasons.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Funding: No, the presenting author of this submission does *not* fall under ICLR’s funding aims, or has sufficient alternate funding.

Submission Number: 144