Judge-Free Test-Time Scaling Tops WebArena Infinity at 74.6%: Salesforce Introduces CLIFT, Alongside MMSearch-OPD and SCOPED-8B Beating 32B VLMs

Core Background & Industry Pain Points

Training autonomous web agents, long-horizon multimodal deep-search agents, and unified Vision-Language Models (VLMs) with reinforcement learning faces three persistent bottlenecks: (1) binary task completion rewards in browser environments are too sparse for credit assignment, while calling frontier LLM judges at every step is prohibitively expensive and unavailable at deployment; (2) long-horizon multimodal search (spanning up to 50 retrieval turns) suffers from a scarcity of hard multi-hop questions and token-level process guidance; and (3) jointly optimizing heterogeneous VLM skills (perception, OCR, symbolic math, multi-image reasoning) triggers severe cross-capability gradient conflicts and reward-hacking biases.

Architectural Highlights & Core Mechanisms

Three new frameworks—Salesforce AI Research's CLIFT (arXiv:2610.06829), MMSearch-OPD (arXiv:2610.22240), and Northeastern University's SCOPED (arXiv:2610.22217)—introduce principled solutions:

• Salesforce's CLIFT: Conformal Self-Verification & Judge-Free Test-Time Scaling: During RL training, the web agent answers natural-language verification questions about each visited URL state; a Compositional Conformal Certifier (CCC) retains only URL-conditional signals that agree with a training-time judge via polarity-aware lift weights, blending them into dense per-step rewards. At test time, the frozen certified bank powers Conformal Trajectory Selection (CTS)—sampling a greedy rollout plus diverse retries and selecting the best trajectory via conservative majority voting without calling any external judge.
• MMSearch-OPD: Controllable Data Refinement + Reward-Gap Joint RL–OPD: Expands simple multimodal questions via parameterized transformations (Refine) to enforce deep cross-modal evidence chains, then combines outcome RL with On-Policy Distillation (OPD) in veRL (Qwen3.5-27B teacher guiding a Qwen3.5-9B student), where the reward gap modulates token-level teacher supervision alongside an exponentially decaying distillation coefficient.
• SCOPED: Skill-Decoupled On-Policy Expert Distillation: Trains capability-specific RL experts and extracts explicit procedural skills via reward-guided reflection. During on-policy distillation, these explicit skills guide the expert teachers while remaining hidden from the unified student, forcing the student to internalize conflict-free multi-domain expertise into its weights with zero skill-prompt overhead at inference.

Authoritative Benchmarks & Empirical Comparisons

• CLIFT Sets SOTA on WebArena Infinity & VisualWebArena (arXiv:2610.06829): Trained on Gemma 4 31B + LoRA across 1,886 WAI and 1,272 VWA synthesized tasks, CLIFT + CTS (K=3) achieves 74.6% overall success across the 9-app WebArena Infinity (WAI) leaderboard (+12.8 points over the 61.8% Gemma-4 base, winning 7 of 9 apps including 91.7% on Elation Clinical/Prescription and 90.7% on PayPal), beating Gemini 3 Flash + BU (70.1%), Qwen 3.5+ (50.7%), and Kimi K2.5 (46.4%). Transferred without weight updates to GPT-5.5 on VisualWebArena (VWA), CLIFT + CTS (K=4) reaches a new 53.7% SOTA (vs. WALT at 52.9% and SGV at 50.2%).
• MMSearch-OPD Gains +15.3 Points Across 7 Search Benchmarks (arXiv:2610.22240): Across 7 multimodal deep-search benchmarks (MMSearch, LiveVQA, BrowseComp-VL, MM-BrowseComp, MMSearch-Plus, BrowseComp-V3, VisBrowse-Bench), MMSearch-OPD-9B reaches 60.2% average accuracy, beating its Qwen3.5-9B agentic workflow baseline by +15.3 points and DeepVoyager-VL-8B by +9.6 points.
• SCOPED-8B Beats 32B & Closed Frontier Models (arXiv:2610.22217): Across 8 perception, OCR/chart, symbolic, and multi-image benchmarks, SCOPED-8B (built on Qwen3-VL-8B) achieves 73.13% average accuracy (+5.83 points over its 67.30% base), surpassing the 4x larger Qwen3-VL-32B (72.01%), Claude-Opus-4.6 (72.86%), Claude-Sonnet-4.6 (72.79%), and GPT-5.4 (69.85%).

Developer Practical Guide & Quickstart

Web agent builders can inspect Salesforce's certified URL question bank and CTS selector at https://huggingface.co/spaces/CoderDoge/CLIFT, while VLM post-training teams can clone https://github.com/NEUIR/SCOPED to consolidate specialist RL checkpoints into a single compact student model without multi-task gradient interference.