Karp and Lipton (STOC, 1980) proved that if $NP \subseteq P/poly$, then the polynomial hierarchy collapses. A long line of work pushed the collapse to smaller classes. In particular, Chakaravarthy and Roy (STACS, 2006; Computational Complexity, 2011) obtained two such collapses, to (apparently) incomparable classes $O_2^P$ and $P^{prMA}$. Hirsch and Volkovich (STACS, 2026) united them by a collapse to the input-oblivious version $P^{prOMA}$ of $P^{prMA}$, which is contained in both; Hirsch, Volkovich and Williams (ECCC TR26-250, 2026) obtained a collapse to a slightly smaller class. Currently, the three deepest known collapses, pairwise incomparable to the best of our current knowledge, are to $P^{prOMA}$ and its versions; to $ZPP^{ONP}$, the input-oblivious version of the collapse to $ZPP^{NP}$ of Köbler and Watanabe (SIAM J. Comput., 1998), which builds on the circuit-learning algorithm of Bshouty, Cleve, Gavaldà, Kannan and Tamon (JCSS, 1996); and to $OMIN = OL_2^P$ (Hirsch and Volkovich; Hirsch, Volkovich and Williams), the input-oblivious version of the class $L_2^P$ of Korten and Pitassi (FOCS, 2024). The last collapse is proved in this paper. The class $OMIN$ has fixed-polynomial circuit lower bounds (Hirsch, Volkovich and Williams); no smaller collapse class with such lower bounds is known.
We unite the three collapses in the same spirit, by a collapse to a class contained in all of them. Derandomization results for input-oblivious classes in the above-mentioned works in fact use a two-stage model: first the computation uses one oracle (and ignores the input), then another oracle. This model is inspired by the untrusted advice classes of Aaronson and Drucker (SIAM J. Comput., 2014), including $YP^*$, and by the learning algorithm of Bshouty et al. as used by Köbler and Watanabe. We use three-stage generator-tester-decoder computations. In this model, a polynomial-time generator with an oracle computes an advice string from $1^n$, a tester certifies it, and a polynomial-time decoder uses it on the input. Let $OM$ be the class of languages decided in this model by a generator with an oracle for a restricted version of $prOMA$, a co-$ONP$ tester, and a $P$ decoder. This class is contained in the three collapse classes above, $OM \subseteq OMIN \cap P^{prOMA} \cap ZPP^{ONP}$. We argue that it is a natural class (for example, closed under polynomial-time Turing reductions). Each of the three containments comes from executing the generator in a different way: with its oracle; with its oracle replaced by a fixed language in $L_2^P$; or with the oracle answers sampled at random, after which the tester checks the advice. We show that if $NP \subseteq P/poly$, then $PH = OM$. The collapse also holds with a unique certified advice string.
At the other end of the picture, all these input-oblivious collapse classes lie in $P^{prSBP}$, and so do even the non-input-oblivious versions $P^{prMA}$, $L_2^P$ and $S_2^P$ of $P^{prOMA}$, $OL_2^P$ and $O_2^P$, as well as $P^{prS_2^P}$: we prove $P^{prS_2^P} \subseteq P^{prSBP}$. This strengthens the containment $S_2^P \subseteq P^{prAM}$ of Chakaravarthy and Roy and our earlier containment $L_2^P \subseteq P^{prSBP}$. The proof unites the Goldreich-Wigderson-type procedure of Chakaravarthy and Roy with our earlier approximate-counting technique. Its input-oblivious counterpart concerns $O_2^P$, the first collapse class of Chakaravarthy and Roy: every language in $O_2^P$ has polynomial-size circuits constructible in $FP^{prSBP}$.