The paper introduces a controlled benchmark for deciding whether to acquire a third modality or abstain when audio, video, and text scores conflict. A threshold policy that observes two signed scores can request the third at a cost and abstain, achieving higher targeted decision accuracy and utility than always-answer fusion under certain reward settings. Experiments on synthetic and real data show that the policy’s advantage depends on the ambiguity mechanism, budget constraints, and the ranking of observed pairs, highlighting the trade‑off between full-modality accuracy and pre‑request selection value.
arXiv:2609.13150v1 Announce Type: cross Abstract: Reference-free quality predictors such as UTMOS, DNSMOS and SCOREQ are the de facto automatic evaluators for text-to-speech (TTS) and are increasingl...
arXiv:2610.20293v1 Announce Type: cross Abstract: Large language models increasingly generate multiple retrieval actions jointly, while downstream systems often provide only one reward for the result...
The paper evaluates Certo, a small non‑generative decision model that scores candidate actions based on text. It compares a joint scorer that processes state, rules, and candidates together with a cacheable encoder that pre‑encodes candidates to reduce cost. Experiments show the cacheable approach loses rule sensitivity, while targeted counterfactual supervision can recover performance on synthetic tasks; however, on real rules the joint scorer still outperforms the cacheable version, and cross‑domain mixtures do not improve accuracy.
The paper introduces State‑Conditioned Minimal Sufficient Evidence Recovery (SER), a method that, given a coding agent’s current state, reconstructs a compact set of evidence passages that collectively provide all facts needed for the agent’s next decision. Using the SERBench dataset of 500 held‑out states from 45 repositories, the authors show that their MSS‑Complement approach recovers a complete evidence set for 73.0 % of states with five items and 80.6 % with eight, outperforming baseline ranking methods. The study also demonstrates that this set‑level policy improves downstream performance on AMA‑Bench and highlights the importance of retrieving missing facts rather than merely re‑ranking similar passages.
arXiv:2607.27422v2 Announce Type: replace Abstract: Goal-conditioned action generators can produce many plausible actions. Best-of-\(K\) selection then uses a critic to rank those actions for the cur...
arXiv:2609.23665v1 Announce Type: cross Abstract: Constraint-following music generation asks a score to satisfy several user-specified properties at once, each checkable programmatically (key, meter,...
Verdict-MQ is an open encoder that processes a single piece of state once and then answers multiple typed questions about that state, avoiding separate passes for each question. By using a structured attention mask, it ensures the state encoding is independent of question order and prevents questions from influencing one another. On the typed‑decisions benchmark, Verdict‑MQ outperforms a larger specialist model, achieving higher accuracy and better calibration, while on other benchmarks it shows mixed results compared to fine‑tuned baselines.
The paper introduces Uncertainty‑Normalized Margin DPO (UNM‑DPO), a method that augments direct preference optimization with strength‑dependent margins and a learned prompt scale derived from a heteroskedastic Bradley‑Terry model. Two training objectives—advantage‑only (AO) and whole‑residual (WR)—compare implicit rewards from response log‑probability ratios, with WR enabling identification of the prompt scale and a practical learning procedure. Building on WR, ULNM‑DPO‑WR further normalizes rewards by response length, and experiments on HelpSteer2/3 and AlpacaEval show higher win rates, mean rewards, and shorter responses compared to baseline DPO and SimPO.
arXiv:2609.22910v1 Announce Type: new Abstract: Vision-language agents that crop and zoom are trained with rewards that credit a successful tool call, yet a successful call does not show that the mod...
arXiv:2606. 15420v1 Announce Type: cross Abstract: A constitution tells a language model what to value, but little tells us whether it does.
arXiv:2608. 19760v1 Announce Type: cross Abstract: Audited against causal ground truth from executed replay in a single-agent tool environment (ALFWorld), none of the step-level credit signals used to train LLM agents -- LLM-judge scores, outcome-conditioned logprob ratios, or the policy's own confidence -- identifies which steps causally matter better than chance.
arXiv:2607. 05904v1 Announce Type: new Abstract: Training a language model against its own reference-free judgments (the premise of self-rewarding, self-play, and LLM-as-a-judge pipelines) assumes a model's verdict on a shown answer tracks correctness.