From Softmax Correspondence to Identification: The Observation Gap and a Diagnostic Template―One-Step Log-Score rEFE and RSA―
Fuente:
Zenodo
Salvato in:
| Autore principale: | |
|---|---|
| Natura: | Recurso digital |
| Lingua: | inglese |
| Pubblicazione: |
Zenodo
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866901757683040256 |
|---|---|
| author | Saito, Ryotaro |
| author_facet | Saito, Ryotaro |
| contents | <p>Softmax choice rules are shared by Active Inference policy selection (often written as a softmax over expected free energy, EFE) and Rational Speech Act (RSA) pragmatics, encouraging cross-context interpretations of logit/softmax slopes as differences in “precision/temperature.” In open-ended corpora, however, the counterfactual utterance menu that fixes the softmax normalizer is unobserved. As a result, utterance policies and temperatures are not identifiable from realized strings and success outcomes alone.</p> <p>This preprint formalizes the observation gap and distinguishes it from a diagnostic gap (policy assumptions are assessed only through outcome-level GLMs) and a proxy gap (unobserved cost/decoder terms require imperfect proxies). As a controlled case study, we define rEFE, a restricted one-step instrumental objective whose hard-goal log-score (reverse-KL) term matches RSA informativeness under channel identification. The rEFE–RSA softmax correspondence is treated as an identification target, not as a general “EFE = RSA” claim.</p> <p>We derive two outcome-level diagnostics for heterogeneous contexts:</p> <p>A2_outcome: a comparability gate for cross-context coefficient/temperature readings (slope×condition block tests).</p> <p>A5_outcome: a residual-baseline check consistent with offset-style calibration (condition-intercept block tests, especially conditional on slope heterogeneity).</p> <p>These are interpretability checks for outcome models under missing menus, not policy-level correspondence tests.</p> <p>Using Colors in Context, pre-specified multiverse robustness and multiway cluster-robust block tests frequently reject A2_outcome, indicating that cross-condition slope/precision interpretations are not licensed in these success models without explicit scale anchors or calibration models. Held-out gains are dominated by geometry and coarse condition baselines rather than a length proxy.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18785620 |
| institution | Zenodo |
| language | eng |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | From Softmax Correspondence to Identification: The Observation Gap and a Diagnostic Template―One-Step Log-Score rEFE and RSA― Saito, Ryotaro Observation Gap Identification Softmax Choice Active Inference Rational Speech Acts Colors in Context <p>Softmax choice rules are shared by Active Inference policy selection (often written as a softmax over expected free energy, EFE) and Rational Speech Act (RSA) pragmatics, encouraging cross-context interpretations of logit/softmax slopes as differences in “precision/temperature.” In open-ended corpora, however, the counterfactual utterance menu that fixes the softmax normalizer is unobserved. As a result, utterance policies and temperatures are not identifiable from realized strings and success outcomes alone.</p> <p>This preprint formalizes the observation gap and distinguishes it from a diagnostic gap (policy assumptions are assessed only through outcome-level GLMs) and a proxy gap (unobserved cost/decoder terms require imperfect proxies). As a controlled case study, we define rEFE, a restricted one-step instrumental objective whose hard-goal log-score (reverse-KL) term matches RSA informativeness under channel identification. The rEFE–RSA softmax correspondence is treated as an identification target, not as a general “EFE = RSA” claim.</p> <p>We derive two outcome-level diagnostics for heterogeneous contexts:</p> <p>A2_outcome: a comparability gate for cross-context coefficient/temperature readings (slope×condition block tests).</p> <p>A5_outcome: a residual-baseline check consistent with offset-style calibration (condition-intercept block tests, especially conditional on slope heterogeneity).</p> <p>These are interpretability checks for outcome models under missing menus, not policy-level correspondence tests.</p> <p>Using Colors in Context, pre-specified multiverse robustness and multiway cluster-robust block tests frequently reject A2_outcome, indicating that cross-condition slope/precision interpretations are not licensed in these success models without explicit scale anchors or calibration models. Held-out gains are dominated by geometry and coarse condition baselines rather than a length proxy.</p> |
| title | From Softmax Correspondence to Identification: The Observation Gap and a Diagnostic Template―One-Step Log-Score rEFE and RSA― |
| topic | Observation Gap Identification Softmax Choice Active Inference Rational Speech Acts Colors in Context |
| url | https://doi.org/10.5281/zenodo.18785620 |