A method designed to test whether AI answers stay tied to the evidence in front of a model led the reported F1 and Accuracy comparisons on HotpotQA and MIRAGE. Called Twin Worlds, it had the highest F1 and Accuracy in all six combinations of those two benchmarks and three tested backbones, the highest Reliability Score in four, and an abstention rate of no more than 2.9 percent. RC-RAG's reported abstention rate ranged from 27 percent to 54 percent, while Context Perturbation reached 25 percent. The findings come from a preprint dated 28 August 2026 and describe benchmark comparisons, not a promise that the same pattern will hold elsewhere.
How the test works
At the heart of the method is a controlled name swap. Twin Worlds delexicalizes entity mentions, turning them into placeholders that record each entity's type, then relexicalizes them with synthetic entities of matching types. The replacements are linked through one-to-one reversible mappings, so an answer from an altered version can be mapped back and compared with the original. If the answer changes in a way that does not track the substitution, the system treats the equivariance failure as a warning and may abstain. The design aims to preserve the relationships in the evidence while reducing lexical cues that might activate parametric priors, or information the model draws from its learned internal patterns.
The default setup generated three twin worlds for each query and used a 0.60 threshold for the equivariance score. That threshold was tuned on a held-out validation set of 500 instances to maximize the Reliability Score. The primary evaluation contained 8,000 instances, with 2,000 from each of HotpotQA, MIRAGE, FaithEval and FEVER, and tested three model backbones.
Strong results across several tests
On FaithEval, where every question is unanswerable, Twin Worlds had the best Accuracy and abstention rate on LLaMA-4 and Mistral-Small, and the second-best results on GPT-5.1. On FEVER, it had the best F1 across all three backbones.
The researchers also tested whether the substitutions preserved the relationships they were meant to preserve. In a human evaluation of 200 MIRAGE and HotpotQA instances, three independent annotators gave structure preservation an average rating of 4.61 points, with 99.1 percent of ratings at least 4. Twin Worlds agreed with human judgments on 86 percent of answerable cases and 81 percent of unanswerable cases, for 83.5 percent overall agreement. Fleiss' kappa, an agreement coefficient, was 0.78.
What happened when the design was changed
Behavioral probes pointed to the same distinction. World Tracking Rate, or WTR, was higher with faithful evidence than with unfaithful evidence for all three backbones. The faithful and unfaithful values, in that order, were 0.804 and 0.215 for GPT-5.1, 0.780 and 0.230 for LLaMA-4, and 0.761 and 0.243 for Mistral-Small. A separate activation probe found lower entity-associated activation after synthetic substitutions, while original entities paired with unfaithful evidence showed the highest activation, especially in later layers.
The ablation tests also favored the full design. In a balanced sample of 500 instances, including 250 answerable and 250 unanswerable cases, the reported Reliability Score reductions were 0.044 with one twin world, 0.048 to 0.051 without type constraints, 0.103 to 0.115 when Wikipedia entities were used, and 0.067 to 0.076 without back-mapping. Using six twin worlds produced negligible gains and slightly lowered GPT-5.1's score. The reported comparisons therefore favored the default combination of three typed twin worlds and reversible back-mapping.
The largest abstention gap appeared in a hard subset of 200 unanswerable instances drawn from all four datasets. The cases were selected because Self-Consistency agreement was at least 0.8 and the Sufficient Context score was at least 0.5, even though the cases were unanswerable. Twin Worlds abstained on 71 percent of them, compared with 9 percent for Self-Consistency and 11 percent for Sufficient Context. The method used four passes and had a mean wall-clock latency of 3.8 seconds, equal to RC-RAG and below Contrastive Decoding's 4.1 seconds. At the same four-pass budget, Self-Consistency's reported Reliability Score was 0.046 lower.
A narrow brief
Those results come with a narrow brief. Twin Worlds is most applicable to entity-grounded reasoning with named entities. The evaluation is limited to English-language benchmarks, while mathematical and procedural tasks are outside the stated scope. Entity-recognition failures may complicate the method. The work is an arXiv preprint dated 28 August 2026.
Paper data and sources
Original title: Twin Worlds: Equivariance-Based Abstention for Evidence-Grounded Reasoning
Authors: Vy Nguyen, Ziqi Xu, Jeffrey Chan et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text