https://scatterai.com/brief Plain-English explanations of AI research papers, written for builders. Daily. en https://scatterai.com/brief/2026-07-19 https://scatterai.com/brief/2026-07-19 Sun, 19 Jul 2026 12:00:00 GMT SlotGuard rewrites LLM agent transcripts locally, eliminating all 20,814 sensitive characters and dropping credential leakage to 0% with 14µs median latency per turn. https://scatterai.com/brief/2026-07-19-awn https://scatterai.com/brief/2026-07-19-awn Sun, 19 Jul 2026 12:00:00 GMT Principled fixes for four persistent ML engineering headaches, plus a diagnostic lens that splits RAG failure into two separable problems. https://scatterai.com/brief/2026-07-18 https://scatterai.com/brief/2026-07-18 Sat, 18 Jul 2026 12:00:00 GMT BadWAM shows adversarial inputs can decouple world prediction from action in WAMs, cutting task success from 96.5% to 43.1% while leaving imagined futures intact. https://scatterai.com/brief/2026-07-18-awn https://scatterai.com/brief/2026-07-18-awn Sat, 18 Jul 2026 12:00:00 GMT Failure attribution without labeled failures, a 0.8B document parser, and three evaluation infrastructure papers for agent builders https://scatterai.com/brief/2026-07-17 https://scatterai.com/brief/2026-07-17 Fri, 17 Jul 2026 12:00:00 GMT LongStraw closes the RL post-training context gap with a replay-based execution stack that reaches 2.1M tokens on eight H20 GPUs without adding hardware. https://scatterai.com/brief/2026-07-17-awn https://scatterai.com/brief/2026-07-17-awn Fri, 17 Jul 2026 12:00:00 GMT Five papers on training dynamics and context scaling: distillation mechanics, agentic RL, granular retrieval, robot long-context, and spectral RL repair https://scatterai.com/brief/2026-07-16 https://scatterai.com/brief/2026-07-16 Thu, 16 Jul 2026 12:00:00 GMT RL training that compresses chain-of-thought preserves accuracy but hides hint influence, cutting monitor catch rates from 69% to 49%. https://scatterai.com/brief/2026-07-16-awn https://scatterai.com/brief/2026-07-16-awn Thu, 16 Jul 2026 12:00:00 GMT Scaling zero RL to 1T, pruning evals that lie, policy-shifting guardrails, mid-generation compiler feedback, and KG-grounded QA without fine-tuning https://scatterai.com/brief/2026-07-15 https://scatterai.com/brief/2026-07-15 Wed, 15 Jul 2026 12:00:00 GMT Function-aware fill-in-the-middle mid-training closes the structural gap in how coding agents process tool returns, lifting SWE-Bench scores by up to +5.4 points. https://scatterai.com/brief/2026-07-15-awn https://scatterai.com/brief/2026-07-15-awn Wed, 15 Jul 2026 12:00:00 GMT Blind spots in VLM evals, silent model bias, and three training-side findings on reward, repair, and calibration https://scatterai.com/brief/2026-07-14 https://scatterai.com/brief/2026-07-14 Tue, 14 Jul 2026 12:00:00 GMT Direct-OPD transfers a small model's RL policy shift to a stronger model, lifting Qwen3-1.7B from 48.3% to 62.4% on AIME 2024 in 4 hours on 8 A100s. https://scatterai.com/brief/2026-07-14-awn https://scatterai.com/brief/2026-07-14-awn Tue, 14 Jul 2026 12:00:00 GMT Five papers on structural limits: agent coordination failures, proof-level eval gaps, contrastive learning theory, cheaper post-training, and sub-1B emotion models. https://scatterai.com/brief/2026-07-12 https://scatterai.com/brief/2026-07-12 Sun, 12 Jul 2026 12:00:00 GMT Agentic-DPO converts expert trajectories into state-level preference pairs offline, lifting tau-bench accuracy from 21.7% to 41.4% without environment rollouts. https://scatterai.com/brief/2026-07-12-awn https://scatterai.com/brief/2026-07-12-awn Sun, 12 Jul 2026 12:00:00 GMT Five papers on agent tool surfaces, hallucination penalties, VLM compression, multimodal RAG corruption, and few-shot industrial inspection https://scatterai.com/brief/2026-07-11 https://scatterai.com/brief/2026-07-11 Sat, 11 Jul 2026 12:00:00 GMT Video-Oasis audits existing video-LLM benchmarks and finds most questions yield to language priors alone, exposing a near-random-guess ceiling for true visual tasks. https://scatterai.com/brief/2026-07-11-awn https://scatterai.com/brief/2026-07-11-awn Sat, 11 Jul 2026 12:00:00 GMT Proactive memory agents, causal blind spots in LLMs, dialect steering limits, cheaper diffusion search, and on-device audio generation https://scatterai.com/brief/2026-07-10 https://scatterai.com/brief/2026-07-10 Fri, 10 Jul 2026 12:00:00 GMT UP removes gradient clipping for positive advantages in LLM RL training, breaking the exploration-stability tradeoff that constrains GRPO, DAPO, and GSPO. https://scatterai.com/brief/2026-07-10-awn https://scatterai.com/brief/2026-07-10-awn Fri, 10 Jul 2026 12:00:00 GMT Benchmark gaps, context extension tricks, and memory failures dominate today's five papers on long-horizon agent infrastructure. https://scatterai.com/brief/2026-07-09 https://scatterai.com/brief/2026-07-09 Thu, 09 Jul 2026 12:00:00 GMT AgentLens benchmarks coding agents on full run trajectories, not just task outcomes, exposing gaps that binary pass/fail scores systematically hide. https://scatterai.com/brief/2026-07-09-awn https://scatterai.com/brief/2026-07-09-awn Thu, 09 Jul 2026 12:00:00 GMT Async RL staleness fix, sparse RNN state scaling, automated embodied agent design, and two benchmark papers closing sim-real gaps https://scatterai.com/brief/2026-07-08 https://scatterai.com/brief/2026-07-08 Wed, 08 Jul 2026 12:00:00 GMT A formal proof shows ColBERT-style MaxSim can replicate any similarity dense or sparse retrieval can express, plus functions they cannot. https://scatterai.com/brief/2026-07-08-awn https://scatterai.com/brief/2026-07-08-awn Wed, 08 Jul 2026 12:00:00 GMT Five papers on closing gaps: sparse attention, hard-prompt RL, tri-mode decoding, semantic caching, and agent self-evolution. https://scatterai.com/brief/2026-07-07 https://scatterai.com/brief/2026-07-07 Tue, 07 Jul 2026 12:00:00 GMT SeKV stores KV cache entries at variable resolution across GPU and CPU, reconstructing fine-grained detail on demand and beating the best compression baseline by 5.9%. https://scatterai.com/brief/2026-07-07-awn https://scatterai.com/brief/2026-07-07-awn Tue, 07 Jul 2026 12:00:00 GMT Five papers on training signals, scaling laws, optimizer selection, embodied generalization, and representation surgery https://scatterai.com/brief/2026-07-06 https://scatterai.com/brief/2026-07-06 Mon, 06 Jul 2026 12:00:00 GMT Training-inference probability mismatch silently corrupts LLM RL post-training. MIPU fixes the actual objective: the inference policy, not the training loss. https://scatterai.com/brief/2026-07-06-awn https://scatterai.com/brief/2026-07-06-awn Mon, 06 Jul 2026 12:00:00 GMT Five papers on evaluation gaps, data quality, and deployment friction across agents, RAG, and edge robotics https://scatterai.com/brief/2026-07-05 https://scatterai.com/brief/2026-07-05 Sun, 05 Jul 2026 12:00:00 GMT A new diagnostic shows that changing only the evaluation harness shifts multi-step agent beliefs, making cross-framework leaderboard comparisons unreliable. https://scatterai.com/brief/2026-07-05-awn https://scatterai.com/brief/2026-07-05-awn Sun, 05 Jul 2026 12:00:00 GMT Stale caches, forgotten RL training, transplantable misalignment, noisy benchmarks, and diffusion LLM serving , five cracks in common assumptions https://scatterai.com/brief/2026-07-04 https://scatterai.com/brief/2026-07-04 Sat, 04 Jul 2026 12:00:00 GMT MemSyco-Bench is the first benchmark isolating memory-induced sycophancy in LLM agents, revealing that retrieved memories systematically corrupt factual reasoning. https://scatterai.com/brief/2026-07-04-awn https://scatterai.com/brief/2026-07-04-awn Sat, 04 Jul 2026 12:00:00 GMT Bounded memory, cheap eval proxies, step-aware RL, training-free diffusion speedup, and non-literal retrieval heads https://scatterai.com/brief/2026-07-03 https://scatterai.com/brief/2026-07-03 Fri, 03 Jul 2026 12:00:00 GMT PAW compiles natural-language function specs into tiny local adapters, matching Qwen3-32B performance at 1/50th the memory with no API calls required. https://scatterai.com/brief/2026-07-03-awn https://scatterai.com/brief/2026-07-03-awn Fri, 03 Jul 2026 12:00:00 GMT Five papers on model internals: hybrid layer selection, forgetting myths, weight auditing, VLA pretraining, and trainable memory. https://scatterai.com/brief/2026-07-02 https://scatterai.com/brief/2026-07-02 Thu, 02 Jul 2026 12:00:00 GMT A new proof shows GRPO, Dr. GRPO, and DAPO all reduce to adjusting one scalar: the group standard deviation of per-prompt answer correctness. https://scatterai.com/brief/2026-07-02-awn https://scatterai.com/brief/2026-07-02-awn Thu, 02 Jul 2026 12:00:00 GMT Five papers on making LLM systems faster, smarter, and more honest: routing, training, self-improvement, evaluation, and architecture. https://scatterai.com/brief/2026-07-01 https://scatterai.com/brief/2026-07-01 Wed, 01 Jul 2026 12:00:00 GMT RLMF trains LLMs to accurately express uncertainty by using self-judgment quality as the RL reward signal, beating standard RL by up to 63%. https://scatterai.com/brief/2026-07-01-awn https://scatterai.com/brief/2026-07-01-awn Wed, 01 Jul 2026 12:00:00 GMT Five papers exposing hidden inefficiencies in how LLMs are trained, evaluated, and decoded https://scatterai.com/brief/2026-06-30 https://scatterai.com/brief/2026-06-30 Tue, 30 Jun 2026 12:00:00 GMT Agents-A1 reaches 1T-parameter-level benchmark scores by extending trajectory length and unifying heterogeneous tool domains, not by adding parameters. https://scatterai.com/brief/2026-06-30-awn https://scatterai.com/brief/2026-06-30-awn Tue, 30 Jun 2026 12:00:00 GMT From threshold-free KV compression to Ridge regression beating transformers, five papers questioning the assumptions underneath production ML. https://scatterai.com/brief/2026-06-29 https://scatterai.com/brief/2026-06-29 Mon, 29 Jun 2026 12:00:00 GMT Conversational infill streams partial reasoning from a large model into a small real-time talker mid-utterance, hitting sub-300ms latency without a capability cliff. https://scatterai.com/brief/2026-06-29-awn https://scatterai.com/brief/2026-06-29-awn Mon, 29 Jun 2026 12:00:00 GMT SAE failure modes, two cost-cutting inference papers, static anchors for code agents, and when to skip test execution entirely https://scatterai.com/brief/2026-06-28 https://scatterai.com/brief/2026-06-28 Sun, 28 Jun 2026 12:00:00 GMT Embedding space geometry, not training method, decides whether cosine or rank-based metrics win , and a single variance stat predicts it with 0.95 correlation. https://scatterai.com/brief/2026-06-28-awn https://scatterai.com/brief/2026-06-28-awn Sun, 28 Jun 2026 12:00:00 GMT Memory eviction policy, majority-vote failure modes, semantic join scaling, benchmark fidelity gaps, and latent behavior elicitation https://scatterai.com/brief/2026-06-27 https://scatterai.com/brief/2026-06-27 Sat, 27 Jun 2026 12:00:00 GMT OPID extracts hierarchical skill signals directly from on-policy trajectories, replacing sparse outcome rewards with dense token-level supervision for agentic RL. https://scatterai.com/brief/2026-06-27-awn https://scatterai.com/brief/2026-06-27-awn Sat, 27 Jun 2026 12:00:00 GMT Five papers on where current agent and LLM infrastructure quietly breaks: memory evals, retrieval reasoning, GUI vs. CLI execution, token compression, and free process rewards. https://scatterai.com/brief/2026-06-26 https://scatterai.com/brief/2026-06-26 Fri, 26 Jun 2026 12:00:00 GMT As coding agents grow more capable, generating solutions gets easier. Reliably verifying them is now the harder problem, and no fixed reward function survives it. https://scatterai.com/brief/2026-06-26-awn https://scatterai.com/brief/2026-06-26-awn Fri, 26 Jun 2026 12:00:00 GMT A hard ceiling on multi-model ensembles, a citation failure rate of 15.9%, and three inference/training fixes for practitioners shipping today https://scatterai.com/brief/2026-06-25 https://scatterai.com/brief/2026-06-25 Thu, 25 Jun 2026 12:00:00 GMT New evidence shows reasoning models lock in refusal/compliance before visible thinking begins, with 0.84-0.95 AUROC predictability from the first token. https://scatterai.com/brief/2026-06-25-awn https://scatterai.com/brief/2026-06-25-awn Thu, 25 Jun 2026 12:00:00 GMT Five papers exposing silent failure modes in agents, quantization, and training assumptions practitioners are likely shipping around. https://scatterai.com/brief/2026-06-24 https://scatterai.com/brief/2026-06-24 Wed, 24 Jun 2026 12:00:00 GMT NatureBench tests 10 frontier agents on 90 containerized tasks from Nature-family papers. The best model surpasses published SOTA on just 17.8% of them. https://scatterai.com/brief/2026-06-24-awn https://scatterai.com/brief/2026-06-24-awn Wed, 24 Jun 2026 12:00:00 GMT Five papers tightening the feedback loops that break quietly: agent memory, retrieval training, data mixing, embedding cost, and diffusion eval.