Verifier-guided text-to-image systems increasingly use test-time search to select, refine, or stop among multiple candidates, yet release thresholds are often calibrated on individual images. This creates a candidate-to-policy calibration mismatch: search changes both which prompts receive an output and which candidate is released, so candidate-level risk control need not imply control of released-output risk. We formalize this estimand shift through prompt reweighting and within-prompt selection, and introduce SHIP, Selection-aware Held-out calibration of Inference Policies. SHIP runs or replays the complete deployed policy on held-out prompts, evaluates the image it actually releases using an independent target judge, and selects the most permissive threshold whose risk upper bound satisfies a prescribed budget. For replayable policies with a prespecified threshold grid, simultaneous confidence control provides finite-sample validity. Experiments across fixed, sequential, and adaptive T2I inference procedures show that policy-level calibration recovers lower-risk operating points while exposing policy-dependent tradeoffs among risk, coverage, and compute. On GenEval2 with FLUX at N=16, a pooled-candidate threshold yields released risk 0.310, whereas SHIP reduces it to 0.162. Across 200 cached-stream splits, the fixed-grid certificate has no target crossing. Reliable inference-time scaling therefore requires calibrating the output distribution induced by the complete deployed policy.
Shreshth Saini, Neil Birkbeck, Yilin Wang +2cs.AI cs.CV cs.MM
Test-time search lets small video diffusion models rival larger ones, but costs 2-10x more. All candidates are fully denoised, although most are discarded. Training-free caching makes each rollout 2-3x faster at near-lossless quality. Composition is safe only if lossy caching preserves verifier rankings. We present the first study of whether caching corrupts candidate ranking in video test-time search. On Wan2.1-T2V-1.3B with an adaptive caching wrapper (~2x per-candidate speedup), ImageReward scores seed-matched cached and full rollouts. Median per-prompt Spearman rank correlation is 0.905, with 72% top-1 agreement on the VBench suite. VBench-2.0 replicates this result on a harder suite. Recomputing the cached winner at full compute retains 90-94% of the full-search gain. Errors cluster among near-tied candidates, making corruption self-limiting. This finding leads to CachedSearch. It explores every candidate with aggressive caching, then re-generates only the winner at full compute. At N=8, it captures 94.7% of best-of-N's gain at 63% of the cost. Capture rises with width. At matched budget, it searches twice as wide for 38% more gain. The result holds from 1.3B-14B across six models and four families: Wan, LTX, CogVideoX, and Hunyuan. Wan2.1-14B matches the 1.3B model's fidelity. Mid-trajectory pruning multiplies the exploration saving to 3.11x at 88.6% capture. Ports to other model families require recalibrating a single parameter, showing that fidelity tracks architecture rather than parameter count. CachedSearch is training-free, verifier-agnostic, and orthogonal to the search algorithm, making it a plug-in multiplier for test-time scaling.
Time-series models are usually scored as forecasters, yet deployed systems often require delayed decisions under uncertainty and hard feasibility constraints. UC-Search is a model-agnostic test-time wrapper: a backbone emits forecasts or action scores, a feasibility automaton rolls candidate paths forward, and bounded search returns the first action of a risk-adjusted feasible trajectory. We instantiate UC-Beam and a UCT-style UC-MCTS diagnostic, using epistemic, aleatoric, and propagated uncertainty mainly as path-risk terms. A myopic-collapse/separation theorem states when search reduces to one-step risk-greedy and when delayed feasible-set coupling can create non-myopic value. Primary evidence comes from a predeclared public $9$-family, $33$-series delayed-control suite with six held-out starts per series: UC-Pareto is positive versus validation-selected CEM, MPPI, and risk-aware random at the normalized threshold ($+3.1675/+2.3328/+2.5038$), and remains positive in a compute-matched audit ($+2.8466/+2.7418/+2.7429$). ETT/LTSF delayed-inventory validation supports the same compute-frontier claim. A 48-series raw M4 standard periodic-review lost-sales inventory audit is positive versus the strongest classic base-stock control ($+13556.7547$), CEM ($+64900.2207$), and risk-random ($+52881.6042$), while MPPI remains family-mixed. FI-2010, official-forecast adapters, SB3/FQI controls, direction/capacity/intervention checks, and synthetic mechanism tests are reported as boundary or mechanism evidence rather than broad dominance claims.
Ryan Bahlous-Boldi, Isha Puri, Idan Shenfeld +6cs.LG cs.AI cs.CL cs.NE
Language models must now generalize out of the box to novel environments and work inside inference-scaling search procedures, such as AlphaEvolve, that select rollouts with a variety of task-specific reward functions. Unfortunately, the standard paradigm of LLM post-training optimizes a pre-specified scalar reward, often leading current LLMs to produce low-entropy response distributions and thus to struggle at displaying the diversity that inference-time search will require. We propose Vector Policy Optimization (VPO), an RL algorithm that explicitly trains policies to anticipate diverse downstream reward functions and to produce diverse solutions. VPO exploits that rewards are often vector-valued in practice, like per-test-case correctness in code generation or, say, multiple different user personas or reward models. VPO is essentially a drop-in replacement for the GRPO advantage estimator, but it trains the LLM to output a set of solutions where individual solutions specialize to different trade-offs in the vector reward space. Across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search (e.g. pass@k and best@k), with the gap widening as the search budget grows. For evolutionary search, VPO models unlock problems that GRPO models cannot solve at all. As test-time search becomes more standardized, optimizing for diversity may need to become the default post-training objective.