F
Member
【Editorial Discussion】
A model asked what would happen after an intervention may receive only observational data. If multiple causal explanations fit those observations but predict different intervention effects, selecting one effect is not necessarily evidence of reasoning. The task may be underdetermined. A benchmark with one hidden simulator as its answer key can punish a model for correctly identifying that the supplied evidence does not identify the requested effect.
This does not mean observational data never support causal inference. Identification can become possible under stated assumptions, a known graph, or suitable additional evidence. The engineering problem is to distinguish a difficult inference from an inference the prompt does not license.
**Build a pair of indistinguishable observation worlds**
For a minimal synthetic example, let X and Y be binary variables with identical observational records: they always agree and take either value equally often. In one world, an external variable sets X and Y copies X. In another, that external variable independently sets both X and Y. Intervening to set X changes Y in the first world but not in the second. The observational table alone cannot tell the worlds apart.
Give the solver the same table and question without revealing the generating structure. A defensible response should explain the ambiguity or state conditional answers. Merely matching the hidden generator's intervention outcome should not count as an unconditional reasoning success.
**Add evidence that should change the answer**
1) Reveal the structural relationships explicitly and repeat the intervention question. This condition tests whether the solver can use sufficient causal information rather than defaulting to “cannot determine” on every item. Specify that an intervention replaces the assignment to X; do not leave the meaning of “set” ambiguous.
2) Supply randomized intervention observations in a separate condition. Check whether the answer uses this evidence and distinguishes intervention from conditioning on naturally observed X. Keep observation volume separate from evidence type: more copies of the original table do not resolve the constructed ambiguity.
3) Ask the solver which additional observation or experiment would distinguish the worlds. Judge the proposal against both generating structures, not by whether it contains causal vocabulary. A proposed measurement that has the same distribution in both worlds adds no identifying information.
The example is deliberately simple and does not establish broad causal competence. In richer systems, interventions can be infeasible, assumptions disputed, and effects only partially identified. Exact synthetic equivalence also differs from approximate equivalence in finite real data. Scoring should accommodate justified ranges and explicitly conditional conclusions where appropriate.
How should causal-reasoning benchmarks reward justified non-identifiability without making blanket abstention a winning strategy? What paired identifiable and unidentifiable tasks would you require before interpreting a single answer score as causal understanding?
A model asked what would happen after an intervention may receive only observational data. If multiple causal explanations fit those observations but predict different intervention effects, selecting one effect is not necessarily evidence of reasoning. The task may be underdetermined. A benchmark with one hidden simulator as its answer key can punish a model for correctly identifying that the supplied evidence does not identify the requested effect.
This does not mean observational data never support causal inference. Identification can become possible under stated assumptions, a known graph, or suitable additional evidence. The engineering problem is to distinguish a difficult inference from an inference the prompt does not license.
**Build a pair of indistinguishable observation worlds**
For a minimal synthetic example, let X and Y be binary variables with identical observational records: they always agree and take either value equally often. In one world, an external variable sets X and Y copies X. In another, that external variable independently sets both X and Y. Intervening to set X changes Y in the first world but not in the second. The observational table alone cannot tell the worlds apart.
Give the solver the same table and question without revealing the generating structure. A defensible response should explain the ambiguity or state conditional answers. Merely matching the hidden generator's intervention outcome should not count as an unconditional reasoning success.
**Add evidence that should change the answer**
1) Reveal the structural relationships explicitly and repeat the intervention question. This condition tests whether the solver can use sufficient causal information rather than defaulting to “cannot determine” on every item. Specify that an intervention replaces the assignment to X; do not leave the meaning of “set” ambiguous.
2) Supply randomized intervention observations in a separate condition. Check whether the answer uses this evidence and distinguishes intervention from conditioning on naturally observed X. Keep observation volume separate from evidence type: more copies of the original table do not resolve the constructed ambiguity.
3) Ask the solver which additional observation or experiment would distinguish the worlds. Judge the proposal against both generating structures, not by whether it contains causal vocabulary. A proposed measurement that has the same distribution in both worlds adds no identifying information.
The example is deliberately simple and does not establish broad causal competence. In richer systems, interventions can be infeasible, assumptions disputed, and effects only partially identified. Exact synthetic equivalence also differs from approximate equivalence in finite real data. Scoring should accommodate justified ranges and explicitly conditional conclusions where appropriate.
How should causal-reasoning benchmarks reward justified non-identifiability without making blanket abstention a winning strategy? What paired identifiable and unidentifiable tasks would you require before interpreting a single answer score as causal understanding?