RRSI Gains Up to 4.7 Points Across Five OOD Benchmarks

Today's Overview

  • Constrained Harness Editing Delivers Limited but Transferable Gains. RRSI gained up to 14.1 points on its evolution split and 4.7 points across five out-of-distribution benchmarks. It also consumed 30% fewer policy tokens.
  • WorldCrafter Compresses History for the Target Camera View. It turns multiview observations into a fixed number of view-specific tokens without explicit depth-based correspondence. The paper reports better long-term consistency and camera control during minute-long exploration.
  • Token-Level Correction May Cut Annotation Costs While Preserving the Model Distribution. onPanda keeps most of the model’s own output. A small controlled study found a 52% reduction in median annotation time versus human post-editing.

Featured

01 A High-Scoring Harness May Just Know the Benchmark

Strong in-distribution results can hide a harness that has written the training tasks into the system. RRSI constrains automated iteration instead of generating ever more candidate edits. A time-varying budget limits how many edits can ship together, while the evolution history rewards unexplored paths.

At selection time, a critic rejects proposals tailored too closely to a specific benchmark. A pruner removes edits with small gains, excessive costs, or mechanisms that no longer work. The surviving changes should be more reusable.

Results split sharply by evaluation setting. Scores rose by up to 14.1 points on the split used for evolution, but only 4.7 points across five out-of-distribution benchmarks. The regularized harness also consumed 30% fewer policy tokens than unconstrained evolution. Teams automating agent prompts, tools, and control flow should copy the constraints, not just the recursive rewriting.

Key takeaways:

  • Report in-distribution and out-of-distribution scores separately when evaluating automated harness optimization.
  • Limit per-round edits and track exploration history to reduce benchmark-specific accumulation.
  • Base deployment decisions on transfer gains and policy-token costs, not the highest score increase alone.

02 Ask the Camera, Then Compress the Memory

Keeping more history does not always produce better memory. The next camera view determines which parts of that history matter.

WorldCrafter conditions history compression on the target camera. It turns multiview observations into a fixed number of view-specific tokens for the video generator’s denoising process. A memory encoder and view-conditioned reader jointly learn what to retrieve, without explicit depth-based correspondence.

Combined with recent temporal context and few-step distillation, the system can explore scenes from one image or a text prompt. The abstract reports better long-term consistency and camera control across static scenes and scenes with motion. It also maintains visual quality during minute-long exploration. Broader generalization and the size of the advantage still need full-paper evidence and testing across more scenes.

Key takeaways:

  • Long-horizon world models can compress history around the target viewpoint instead of preserving it unchanged.
  • A fixed number of view-specific tokens provides a clear interface for limited context budgets.
  • Minute-long exploration looks promising, but the abstract does not establish performance across all environments with motion.

03 Fix the First Bad Token, Not the Whole Answer

Rewriting an entire model response is slow. It can also turn alignment data into samples of human writing.

onPanda asks annotators to find the first unsuitable token and correct it. The model then continues from the revised prefix. This cycle repeats until the result is usable.

Most of the final text still comes from model sampling, so the data stays closer to the model’s generation distribution. That makes it suitable for on-policy SFT, preference data, and agent interaction traces. Each correction also records the wrong token, its replacement, and the exact location.

A small controlled study found a 52% reduction in median annotation time versus human post-editing. Treat that as an early efficiency signal, not a universal estimate.

Key takeaways:

  • Token-level correction may reduce human effort while preserving most of the model’s generation distribution.
  • Correction logs naturally provide positive and negative supervision at exact token positions.
  • The study found 52% lower median annotation time than post-editing, but larger studies must test whether that result generalizes.
RRSI Gains Up to 4.7 Points Across Five OOD Benchmarks

Also Worth Noting

04
GameHorizon Combines Offline Scores With Stepwise In-Game Failures EvaluationIts multiscale evaluation covers 21 games, 5,000 hours of expert footage, and more than one million model calls. link.
05
Jev-Mem Adds a Lightweight Control Plane for Memory AgentIt manages memory routing, budget allocation, and stopping decisions. The system calls a slower model only for complex reasoning and reports both quality and latency gains on LoCoMo. link.
06
Harness-Zero Distills Harness Behavior Into Model Weights TrainingAfter removing specialized harnesses, it still raised macro-average task success from 23.3% to 44.3%. link.
07
D-RAC Cuts Chunking Output Tokens by 95.7% RetrievalThe comparison uses agentic chunking with frontier LLMs. D-RAC converts pages into retrieval-friendly Markdown, then plans IDs for deterministic text units. link.
08
VideoGen-Agent Improves When Given Better Generation Tools Video GenReplacing its generator raised its VABench score from 75.6 to 86.1 without retraining. The agent coordinates enhancement, generation, and verification tools. link.
09
ACLArena Tracks Forgetting Across Multistage Post-Training TrainingIt analyzes forgetting at model and token levels across multi-teacher on-policy distillation, self-distillation fine-tuning, and model merging. Its method combines offline replay of high-quality trajectories with routed LoRA experts specialized through RL. link.
10
Complex KDA Represents 2D Rotation With One Structured Update ArchitectureIt uses a single diagonal-plus-rank-one update. The method only expands the allowed gate and delta values while preserving non-expansiveness and the original computational structure. link.
11
KernelBench’s Official Checker Misses 16.9% of Independently Confirmed Failures EvaluationThe result comes from 10,303 compilable faults injected into 188 problems. The miss rate reaches 78.6% for numerical-accuracy faults. link.
12
EvidenT Verifies Retrieved Evidence Before Answering RetrievalIt uses deterministic lexical alignment. Across about 500 real enterprise queries, it increased matches to gold-standard sources by 29% on average versus a prompting baseline. link.
13
ARPS Reaches 87.3% Success on Zero-Adaptation LIBERO-Plus Transfer RoboticsIts action expert reads only compact, future-facing, action-relevant predicted states. It outperformed Fast-WAM by 39.2 percentage points. link.
14
Similar Recommendation-Change Rates Do Not Determine the Authority Gap EvaluationAuthEval scores model recommendations separately from the final choices made by local services. The rate alone cannot determine the gap’s size or direction. link.