Voice RAG’s F1 Drop Widens by 67%

Featured

01 Agents Need Sustained Delivery

“Working capability” shifts the goal from answering once to making steady, verifiable progress on real tasks. An agent must use file, search, and coding tools, retain state, recover from failures, and deliver inspectable results.

Apodex 1.1 expands both execution environments and long-horizon training. Agents learn to decompose tasks, delegate in parallel, combine asynchronous results, and replan. A shared runtime and AgentOS preserve task state and action provenance across the workflow.

The abstract places Apodex among leading systems across finance, science, mathematics, coding, and search. Its 35B Mini model retains strong performance and offers an appealing size for local deployment. Full experiments must still confirm error control, coordination costs, and verification quality during long runs.

Key takeaways: - Evaluate state retention, failure recovery, and delivery verification alongside final-answer quality. - Shared, traceable state and coordination mechanisms matter more than the number of agents. - Test 35B-class models on your own long workflows before judging deployment stability and cost.

Source: Apodex 1.1: Scaling Agentic Intelligence for Complex Work

02 Smarter Retrieval Makes Voice RAG More Fragile

Entity graphs and multi-turn rewriting should correct small errors in spoken queries. Instead, they can turn one transcription mistake into a coherent but incorrect retrieval path.

Across three multi-hop QA benchmarks, combining both methods widened the F1 decline by 36%–67%. The comparison ran from clean text to accents with the highest word error rates. Incorrect ASR entities caused 87%–96% of degraded cases on 2WikiMultiHopQA, while simple surface-level repairs failed to close the gap.

Complex systems still achieved higher absolute F1, but that does not make them more resistant to errors. The accents came from neural speech synthesis, so real users still need separate testing. Voice teams should measure error amplification across the full pipeline, not only retrieval recall.

Key takeaways: - Report performance gaps between clean text and multiple recognition-error levels when evaluating voice RAG. - Test whether incorrect entities spread through the retrieval chain before adding entity graphs or multi-hop rewriting. - Higher absolute accuracy does not guarantee resilience. Validate real accents and noisy environments separately.

Source: Better Retrieval, Worse Robustness: How Multi-hop RAG Amplifies Upstream ASR Errors

03 Move Stability Constraints Into the Environment

Policy-KL keeps model outputs near a reference policy. That limits training drift, but it also directly restricts response-side exploration.

ERPO separates those jobs. Query-KL constrains shifts in the training query distribution. Static weights from a reference distribution then steer updates toward typical queries. Its regularization gradient passes only through query probabilities, leaving the response distribution untouched.

PPO and GRPO can therefore tune environment stability separately from policy exploration, without extra forward computation. The abstract reports higher accuracy across six mathematical reasoning benchmarks. Training also stayed more stable with high-temperature sampling and longer runs. Evidence beyond mathematics is still needed to assess bias and distribution control.

Key takeaways: - Treat drift control and response exploration as separate tuning controls in PPO or GRPO. - Input-side regularization deserves attention during high-temperature sampling and extended training. - Stability on mathematics benchmarks does not automatically transfer to coding or tool-use tasks.

Source: Beyond the Stability-Exploration Dilemma: Environmental Regularization for LLM Policy Optimization

04 A Few Heads Can Hold Long Memories

Interactive world models may not need ever-growing contexts for fast responses and long-term consistency. ReWorld divides those responsibilities across attention heads.

Most local heads inspect recent frames to keep action control responsive. A small set of global heads reviews the full history and preserves spatial memory. Random routing prevents this ability from becoming fixed in specific heads.

At inference time, ReWorld compresses history into a fixed-size KV cache. It also retrieves nearby scenes from a landmark store indexed by camera position. Memory use stays flat as playtime grows. In a 64-second round trip spanning 384 latent states, 12 retained chunks reconstructed the starting view. Four-step sampling produced 704×1280 video in real time.

Key takeaways: - Long-term consistency does not require every attention head to scan the full history. - A fixed cache plus a location-indexed landmark store keeps memory costs controlled during extended interaction. - Evaluate world models on action response, revisit consistency, and real-time generation quality together.

Source: ReWorld: An Interactive World Model with Long-Horizon Memory