A lot of past neuro-symbolic work struggled because the integration layer was too tightly coupled.
Either the symbolic part constrained the model too much, or the neural part bypassed it entirely.
It feels like the real opportunity now is treating symbolic reasoning more like a verification or constraint layer rather than part of the generation process itself.
Discussion (1 Comments)Read Original on HackerNews
Either the symbolic part constrained the model too much, or the neural part bypassed it entirely.
It feels like the real opportunity now is treating symbolic reasoning more like a verification or constraint layer rather than part of the generation process itself.