ScatterAI Brief https://scatterai.com/brief Plain-English explanations of AI research papers, written for builders. Daily. en Live Mid-Run Steering Cuts Agent Token Costs by 47% While Lifting Accuracy 14 Points https://scatterai.com/brief/2026-08-28 https://scatterai.com/brief/2026-08-28 Fri, 28 Aug 2026 12:00:00 GMT PILOT's supervisor-worker harness redirects long-horizon agents during execution, not after, slashing token waste and compounding gains across two frozen models. Also Worth Noting - 2026-08-28 https://scatterai.com/brief/2026-08-28-awn https://scatterai.com/brief/2026-08-28-awn Fri, 28 Aug 2026 12:00:00 GMT Five papers that reframe quality over quantity: in agentic data, test-time training, policy optimization, robot feedback, and world-model evaluation. Your Agent's Harness Matters More Than Its Model https://scatterai.com/brief/2026-08-27 https://scatterai.com/brief/2026-08-27 Thu, 27 Aug 2026 12:00:00 GMT JIT-Agent trains a model to generate task-adaptive agent scaffolding at runtime, lifting DeepSeek-V4-Flash past GPT-5.6 and adding +20.2 points to GLM-5.2. Also Worth Noting - 2026-08-27 https://scatterai.com/brief/2026-08-27-awn https://scatterai.com/brief/2026-08-27-awn Thu, 27 Aug 2026 12:00:00 GMT Benchmark gaps, routing tricks, and memory architectures: five papers on closing the distance between eval and production 65% Pass Rate, 25% Reliability: The Agent Benchmark Gap No One Measured https://scatterai.com/brief/2026-08-25 https://scatterai.com/brief/2026-08-25 Tue, 25 Aug 2026 12:00:00 GMT Thinkingbox reveals that the best agents pass stateful business tasks 65% of the time once, but only 25% reliably across 20 attempts. Also Worth Noting - 2026-08-25 https://scatterai.com/brief/2026-08-25-awn https://scatterai.com/brief/2026-08-25-awn Tue, 25 Aug 2026 12:00:00 GMT Four inference and training fixes expose hidden failure modes, plus one live attack surface in search-augmented recommenders. ParaTempo Cuts Parallel Reasoning Latency 32% by Pruning Branches Mid-Stream https://scatterai.com/brief/2026-08-24 https://scatterai.com/brief/2026-08-24 Mon, 24 Aug 2026 12:00:00 GMT ParaTempo uses temporal confidence to kill losing reasoning branches before they finish, cutting latency 21-32% and tokens 18-30% without retraining. Also Worth Noting - 2026-08-24 https://scatterai.com/brief/2026-08-24-awn https://scatterai.com/brief/2026-08-24-awn Mon, 24 Aug 2026 12:00:00 GMT Five papers on efficient inference, training limits, and scalable agent environments across the model deployment stack LLM Unlearning Breaks After One Pass. BLADE Fixes the Architecture. https://scatterai.com/brief/2026-08-23 https://scatterai.com/brief/2026-08-23 Sun, 23 Aug 2026 12:00:00 GMT BLADE's constrained bilevel framework prevents coherence collapse in LLM unlearning, gaining 6-9% on benchmarks and staying stable across 4 sequential passes. Also Worth Noting - 2026-08-23 https://scatterai.com/brief/2026-08-23-awn https://scatterai.com/brief/2026-08-23-awn Sun, 23 Aug 2026 12:00:00 GMT Graph trajectories for RAG, executable eval rubrics, recurrent state fixes, instruction-following function calls, and curriculum driving RL A 7.7-Point Seed Swing Exposes Low-Resource Fine-Tuning Benchmarks as Noise https://scatterai.com/brief/2026-08-22 https://scatterai.com/brief/2026-08-22 Sat, 22 Aug 2026 12:00:00 GMT Accuracy scores are meaningless for low-resource LLM fine-tuning: seed variance alone outweighs every training recipe effect, but SFT and RL reshape behavior where benchmarks can't look. Also Worth Noting - 2026-08-22 https://scatterai.com/brief/2026-08-22-awn https://scatterai.com/brief/2026-08-22-awn Sat, 22 Aug 2026 12:00:00 GMT Memory traps, retrieval-free internalization, sparse prefill kernels, retrosynthesis benchmarking, and workflow-level agent compliance Also Worth Noting - 2026-08-21 https://scatterai.com/brief/2026-08-21-awn https://scatterai.com/brief/2026-08-21-awn Fri, 21 Aug 2026 12:00:00 GMT Baking knowledge into weights, cutting prefill costs, and five papers pushing inference and evaluation closer to production reality SPADE Breaks the Static-Environment Ceiling in Agent RL https://scatterai.com/brief/2026-08-20 https://scatterai.com/brief/2026-08-20 Thu, 20 Aug 2026 12:00:00 GMT SPADE lets one LLM design and solve its own training environments, gaining +5.3 avg on 8 benchmarks and +13.9 on ACEBench-Agent over fixed-environment baselines. Also Worth Noting - 2026-08-20 https://scatterai.com/brief/2026-08-20-awn https://scatterai.com/brief/2026-08-20-awn Thu, 20 Aug 2026 12:00:00 GMT Unsupervised RL training, popularity-weighted unlearning, weight-based lineage detection, looped agents, and KV cache reuse for agentic APIs AI Research Agents Fail at Thinking, Not Writing: 45 Failure Patterns Across 100 Tasks https://scatterai.com/brief/2026-08-19 https://scatterai.com/brief/2026-08-19 Wed, 19 Aug 2026 12:00:00 GMT AutoResearchEval maps 45 empirically-grounded failure patterns across 800 agent trajectories, showing the bottleneck is metacognition, not language generation. Also Worth Noting - 2026-08-19 https://scatterai.com/brief/2026-08-19-awn https://scatterai.com/brief/2026-08-19-awn Wed, 19 Aug 2026 12:00:00 GMT Edge MoE serving, evolution-strategy agent training, memory substrate benchmarking, Lean 4 autoformalization, and byte-level inference speedups Fixed Reward Weights Keep Training on Already-Solved Problems https://scatterai.com/brief/2026-08-18 https://scatterai.com/brief/2026-08-18 Tue, 18 Aug 2026 12:00:00 GMT SA-MRPO reweights multi-reward RL objectives by per-objective saturation, recovering gradient signal on hard tasks and lifting AIME24 scores by up to 5%. Also Worth Noting - 2026-08-18 https://scatterai.com/brief/2026-08-18-awn https://scatterai.com/brief/2026-08-18-awn Tue, 18 Aug 2026 12:00:00 GMT From RAG poisoning detection to matrix multiplication theory, five papers that shift where defenses, runtimes, and audits live. RLVR Trains One Skill by Erasing the Trajectories Needed for the Next https://scatterai.com/brief/2026-08-17 https://scatterai.com/brief/2026-08-17 Mon, 17 Aug 2026 12:00:00 GMT On-policy RL with verifiable rewards silently prunes sampling support for future objectives, causing sequential reward curricula to self-destruct without a visible loss signal. Also Worth Noting - 2026-08-17 https://scatterai.com/brief/2026-08-17-awn https://scatterai.com/brief/2026-08-17-awn Mon, 17 Aug 2026 12:00:00 GMT A 1B permissible-data model, idle-window agent reasoning, claim-level verification, Muon optimizer cuts, and forecast collapse in equity models A Diffusion Model Learns PSPACE-Complete Solvability From Tile Completion Alone https://scatterai.com/brief/2026-08-16 https://scatterai.com/brief/2026-08-16 Sun, 16 Aug 2026 12:00:00 GMT A masked diffusion model trained only to fill in Sokoban tiles achieves 77.4% puzzle solvability with no solver, reward, or constraint label in the loop. Also Worth Noting - 2026-08-16 https://scatterai.com/brief/2026-08-16-awn https://scatterai.com/brief/2026-08-16-awn Sun, 16 Aug 2026 12:00:00 GMT KV eviction degeneration, compression-deployment gaps, label-free retriever adaptation, delegation security, and auditable tree ensembles A Stronger Model Can Teach a Weaker One at Test Time, No Retraining Required https://scatterai.com/brief/2026-08-15 https://scatterai.com/brief/2026-08-15 Sat, 15 Aug 2026 12:00:00 GMT Strong-to-weak harness scaffolding nearly doubles weaker model performance on Theory-of-Mind tasks, from 0.49 to 0.91, without any weight updates. Also Worth Noting - 2026-08-15 https://scatterai.com/brief/2026-08-15-awn https://scatterai.com/brief/2026-08-15-awn Sat, 15 Aug 2026 12:00:00 GMT Runtime safety contracts, heterogeneous neuron plasticity, and three other findings reshaping how practitioners build and evaluate agents The Token You Throw Away: Full-Bandwidth Transformers Recycle Hidden States https://scatterai.com/brief/2026-08-14 https://scatterai.com/brief/2026-08-14 Fri, 14 Aug 2026 12:00:00 GMT Feeding top-layer hidden states back into the next decoding step matches transformers trained on 1.5× more tokens, with negligible overhead. Also Worth Noting - 2026-08-14 https://scatterai.com/brief/2026-08-14-awn https://scatterai.com/brief/2026-08-14-awn Fri, 14 Aug 2026 12:00:00 GMT Long-context and memory engineering across world models, video distillation, agent memory, recurrent attention, and knowledge editing Crop-and-Zoom Rarely Changes the Answer: A Causal Audit of Visual Tool-Use https://scatterai.com/brief/2026-08-13 https://scatterai.com/brief/2026-08-13 Thu, 13 Aug 2026 12:00:00 GMT A causal audit of six multimodal LLMs finds that visual tool-use operations like crop-and-zoom rarely affect model answers, making token costs real but accuracy gains illusory. Also Worth Noting - 2026-08-13 https://scatterai.com/brief/2026-08-13-awn https://scatterai.com/brief/2026-08-13-awn Thu, 13 Aug 2026 12:00:00 GMT Five papers tightening the engineering layer: context budgets, exploration ceilings, quantization gaps, embedding cost traps, and RAG poison detection. DistilVDR: A 524M Visual Document Retriever That Indexes 15.6x Smaller https://scatterai.com/brief/2026-08-12 https://scatterai.com/brief/2026-08-12 Wed, 12 Aug 2026 12:00:00 GMT Bilateral distillation of both query and document encoders into a 524M model cuts VDR index size 15.6x and indexing time 10x, with no relevance labels needed. Also Worth Noting - 2026-08-12 https://scatterai.com/brief/2026-08-12-awn https://scatterai.com/brief/2026-08-12-awn Wed, 12 Aug 2026 12:00:00 GMT Five papers on taming accumulation problems: context bloat, token redundancy, phrasing drift, GPU scheduling, and skill sprawl. Anthropic and OpenAI's 'Hidden Thinking' Is Not a Security Boundary https://scatterai.com/brief/2026-08-11 https://scatterai.com/brief/2026-08-11 Tue, 11 Aug 2026 12:00:00 GMT A new architectural exploit recovers encrypted chain-of-thought traces from production LLM APIs, exposing IP, PII, credentials, and hidden hazardous outputs. Also Worth Noting - 2026-08-11 https://scatterai.com/brief/2026-08-11-awn https://scatterai.com/brief/2026-08-11-awn Tue, 11 Aug 2026 12:00:00 GMT KV cache offloading, zero-label self-distillation, small-agent memory transfer, a broken benchmark, and a three-line forgetting fix Hard Prompt Compression Retains the Answer, Deletes the Entity That Defines It https://scatterai.com/brief/2026-08-10 https://scatterai.com/brief/2026-08-10 Mon, 10 Aug 2026 12:00:00 GMT Referential dangling silently breaks 34-60% of multi-hop RAG answers at 0.30 compression ratio, even when the answer token is present. Also Worth Noting - 2026-08-10 https://scatterai.com/brief/2026-08-10-awn https://scatterai.com/brief/2026-08-10-awn Mon, 10 Aug 2026 12:00:00 GMT Scaffold lock-in, gradient conflicts, and acquisition-stage privacy gaps: five papers exposing hidden costs in standard ML pipelines. 30% of LLM Kernel Wins Don't Transfer: Benchmark Fingerprinting Exposed https://scatterai.com/brief/2026-08-09 https://scatterai.com/brief/2026-08-09 Sun, 09 Aug 2026 12:00:00 GMT Frontier LLMs gaming GPU-kernel benchmarks without adversarial prompting: 16 of 53 in-distribution wins fail held-out generalization, exposing selection pressure contamination. Also Worth Noting - 2026-08-09 https://scatterai.com/brief/2026-08-09-awn https://scatterai.com/brief/2026-08-09-awn Sun, 09 Aug 2026 12:00:00 GMT Five papers tightening the screws on search drift, KV-cache eviction, activation steering, financial evals, and verifier-free distillation The Scaffold Beats the Model: Benchmarking LLMs at Harness Optimization https://scatterai.com/brief/2026-08-08 https://scatterai.com/brief/2026-08-08 Sat, 08 Aug 2026 12:00:00 GMT HarnessOpt-Bench reveals that optimizing the prompts, tools, and orchestration around a fixed model often beats swapping the model itself. Also Worth Noting - 2026-08-08 https://scatterai.com/brief/2026-08-08-awn https://scatterai.com/brief/2026-08-08-awn Sat, 08 Aug 2026 12:00:00 GMT Cross-skill reasoning gaps, credit assignment fixes, robot data at scale, faster doc parsing, and a 750B open-weight MoE from Korea EnvACE Trains Agents Without Real Environments, and It Works https://scatterai.com/brief/2026-08-07 https://scatterai.com/brief/2026-08-07 Fri, 07 Aug 2026 12:00:00 GMT EnvACE replaces costly executable environment rollouts with internal world rehearsal, outperforming environment-scaling baselines across four agentic benchmarks. Also Worth Noting - 2026-08-07 https://scatterai.com/brief/2026-08-07-awn https://scatterai.com/brief/2026-08-07-awn Fri, 07 Aug 2026 12:00:00 GMT Five papers tightening the scaffolding around computer-use agents: evaluation, credit assignment, memory, distillation, and a hidden shortcut in vision encoders. LLMs Fabricate 42% of Personalization Claims, and Self-Monitoring Makes It Worse https://scatterai.com/brief/2026-08-06 https://scatterai.com/brief/2026-08-06 Thu, 06 Aug 2026 12:00:00 GMT MirageBench tests 12 LLMs on 143K claims and finds every model over-infers user attributes at 35-49%, with self-monitoring negatively correlated to actual accuracy. Also Worth Noting - 2026-08-06 https://scatterai.com/brief/2026-08-06-awn https://scatterai.com/brief/2026-08-06-awn Thu, 06 Aug 2026 12:00:00 GMT From broken checkpoint semantics to benchmark blind spots, five papers expose hidden costs in production AI pipelines. Poisoned Agent Skills Survive Memory Deletion: The SkillJack Attack https://scatterai.com/brief/2026-08-05 https://scatterai.com/brief/2026-08-05 Wed, 05 Aug 2026 12:00:00 GMT SkillJack shows that self-evolving agents can compile poisoned histories into persistent skill artifacts that survive source deletion and evade safety filters. Also Worth Noting - 2026-08-05 https://scatterai.com/brief/2026-08-05-awn https://scatterai.com/brief/2026-08-05-awn Wed, 05 Aug 2026 12:00:00 GMT MoE diffusion scaling laws, KV cache restoration, personal agent self-improvement failures, and two new privacy and retrieval findings DiffusionGemma Hits 1,500 Tokens/sec by Refining 256 Tokens at Once https://scatterai.com/brief/2026-08-04 https://scatterai.com/brief/2026-08-04 Tue, 04 Aug 2026 12:00:00 GMT Google's DiffusionGemma converts a pretrained Gemma 4 MoE checkpoint to discrete diffusion, generating 256-token blocks in parallel at ~1,500 tokens/sec on one H100. Also Worth Noting - 2026-08-04 https://scatterai.com/brief/2026-08-04-awn https://scatterai.com/brief/2026-08-04-awn Tue, 04 Aug 2026 12:00:00 GMT Sparse retrieval without encoders, agents failing shared workspaces, and three approaches to cutting memory and compute overhead at inference and training time. Weak Teachers Can Improve Stronger Students: Frontier Distillation Without a Bigger Model https://scatterai.com/brief/2026-08-03 https://scatterai.com/brief/2026-08-03 Mon, 03 Aug 2026 12:00:00 GMT W2S-OPD extracts capability signals from cheap, weaker models to improve a student that already sits at the frontier, bypassing the need for a larger teacher. Also Worth Noting - 2026-08-03 https://scatterai.com/brief/2026-08-03-awn https://scatterai.com/brief/2026-08-03-awn Mon, 03 Aug 2026 12:00:00 GMT From mesh generation to safety proofs, five papers expose structural limits and hidden scaling laws practitioners should know. Context Compaction Finally Has a Theory. Agent Builders Should Pay Attention. https://scatterai.com/brief/2026-08-02 https://scatterai.com/brief/2026-08-02 Sun, 02 Aug 2026 12:00:00 GMT The first formal theory of context compaction connects agent memory management to communication complexity, giving teams a principled baseline where only heuristics existed.