Large Language Models Are Neurosymbolic Reasoners

Original Article ↗ Hacker News Discussion ↗

Large language models are prompting renewed arguments over whether they genuinely “reason” or merely act as vast compressed lookup tables over finite inputs. Commenters debate parallels between neural networks and humans under physical determinism, the role of world models and embodiment, and whether concepts like consciousness or understanding are even meaningful for software. Others point to emerging hybrids—LLMs augmented with symbolic tools, reinforcement learning in text-based games, and systems like Cyc—as evidence that combining statistical and structured approaches may be the most promising path toward more robust reasoning.

Perception of the “neurosymbolic reasoner” claim

  • Some see the paper as confirming something obvious: LLMs can implement symbolic-like operations over text, but this doesn’t feel like a big step forward.
  • Others think it’s still useful to formalize and measure these capabilities, especially via interactive fiction (IF) games where actions and world state are in natural language.
  • There is debate over what “reasoner” means; the title is seen by some as rhetorically strong without a clear definition.

Are neural networks just lookup tables?

  • One long subthread argues any finite neural net is a numerical function and thus equivalent to a huge lookup table; therefore it “doesn’t think”.
  • Counterpoints:
    • The same reasoning would make humans and the universe lookup tables under determinism; if that’s allowed, “not thinking” becomes an unhelpful definition.
    • Lookup-table equivalence is mathematically true for any finite program but does not capture how generalization, compression, and interpolation work.
    • Recurrent/stateful nets, arbitrary precision, and connections to the halting problem show subtleties: some tables would be infinite or uncomputable.
  • Disagreement over whether physics is fundamentally computable; some invoke Bekenstein bounds and finite information, others argue continuous, real-valued models resist full discretization.

Humans vs machines: thinking, embodiment, and consciousness

  • One side insists humans cannot be reduced to tables “even in theory” because biology, embodiment, and subjective experience matter; software is “just numbers”.
  • Others argue humans are mechanistic systems of particles/fields; in principle all inputs/outputs can be numerically described, so refusing table equivalence is inconsistent.
  • Embodied cognition advocates emphasize movement, sensory-motor loops, plasticity, and goals in a shared world; LLMs only compress and replay text, lacking intentions or world-causal grounding.
  • Opponents claim similar “mystical” arguments were made against mechanistic minds historically, and see no reason silicon could not eventually support minds or agency.

Reasoning abilities and limits of LLMs

  • Many agree LLMs often fail at multi-step or “multi-hop” reasoning unless given chain-of-thought prompts; errors compound and hallucinations appear.
  • Some argue this shows they don’t truly reason, just recombine patterns from training; Advent of Code and subtle concurrency bugs are cited as failures on unseen problems.
  • Others counter that chain-of-thought is exactly how humans reason, and that self-critique or multi-agent setups already improve reliability.
  • There’s debate whether statistical pattern matching over text can yield genuine world models or only approximate the distribution of language.

Games as reasoning benchmarks

  • The paper’s success on text IF games is seen as promising, because rules and state are in natural language.
  • Several commenters propose harder tests: NetHack, Zork, complex roguelikes, or Diplomacy, where planning, exploration, and hidden information matter.
  • NetHack sparks a detailed side debate:
    • Some claim it’s effectively unwinnable without “spoilers” (external knowledge) due to opaque mechanics and heavy randomness.
    • Others note expert players with high win rates and suggest RL agents, given enough runs, might infer useful mechanics despite noise.
  • A cited line of work uses LLMs to provide intrinsic rewards for RL agents in NetHack; another system uses a text interface to files, where GPT‑4 can backtrack and search but GPT‑3.5 struggles.

Symbolic systems and hybrid approaches

  • One project injects explicit computational structures (lambda calculus, stacks, queues) into LLM internals to get more reliable, program-like reasoning over latent states.
  • A contrasting camp points to classic symbolic AI like Cyc, which uses explicit logical inference and a hand/derived knowledge base, arguing it “really reasons” and doesn’t hallucinate.
  • Critics question whether using NNs to synthesize programs is worthwhile given massive training costs and energy use, especially if the end product is a conventional program.
  • Supporters reply that automated program synthesis could still be cheaper and more scalable than paying humans, and that NNs let such systems leverage existing LLM knowledge.

Meta: learning resources and terminology

  • One commenter asks for good ways to keep up with LLM developments; others suggest using the site’s search to find introductory and tutorial content.
  • Several participants stress that loose use of terms like “information”, “entropy”, and “world model” across physics, CS, and ML causes confusion in these debates.