Sources

Measuring tactical intelligence targeting and conventional weapons capabilities of AI models anthropic.com

Locked at the Entrance, Open Inside: Where RLVR Narrows the Solution Space huggingface.co

Reinforcement learning with verifiable rewards narrows reasoning diversity primarily at the initial solution step rather than during execution, and targeted interventions can restore coverage without sacrificing accuracy.

Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning huggingface.co

Random eviction of reasoning tokens matches selective KV cache compression because reasoning traces are self-protecting through redundancy, making scoring unnecessary once prompts are preserved.

Rethinking On-Policy Distillation of Large Language Models II: One Training Example huggingface.co

A single training query drives hundreds of useful distillation steps by rapidly covering teacher states, but student alignment still lags. The bottleneck is the token-level supervision algorithm rather than data volume, pointing multi-teacher OPD work toward better credit assignment instead of larger query pools.

Why Gated DeltaNet Survives 4-Bit Quantization: NVFP4 W4A4 for the Recurrent Half of a Hybrid 27B LLM huggingface.co

Quantizing both attention and recurrent Gated DeltaNet layers of a 27B hybrid model to W4A4 NVFP4 holds accuracy on long-context and reasoning benchmarks. Delta-rule recurrence localizes outliers, letting the recurrent half tolerate 4-bit weights and activations that usually break linear-attention layers.

Principia: Relational Physics Tests for Video Models huggingface.co

Principia scores video models on paired-object interactions using a calibration-independent relational consistency metric, sidestepping absolute measurement errors. Leading generators show major gaps in Newtonian reasoning, suggesting current world models learn appearance dynamics without the underlying causal structure.

RealSWE: A Compositional Evaluation of Coding Agents under Realistic User Requests huggingface.co

Benchmark tasks like SWE-bench use long structured prompts, but real user requests are terse and informal, dropping LLM resolution rates. Adding explicit desired behavior and motivation to prompts recovers much of the lost performance, per an information taxonomy built from real-world requests.

VeriPhy: Agentic Physical Reasoning for World Model Evaluation and Refinement huggingface.co

VeriPhy compiles prompts into typed physical obligations, then runs frozen experts for segmentation, depth, OCR and audio detection to gather provenance-tagged evidence. A three-valued logic maps that evidence to auditable pass/fail/unknown verdicts on whether generated video respects the requested physics.

DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training huggingface.co

DRACO generates multi-criteria rubrics on the fly and redistributes trajectory-level rewards into per-step advantages, giving long-horizon agents fine-grained credit without verifiers. The outcome-blind rubric optimization plugs into GRPO and improves agent training where sparse end-of-episode rewards previously stalled learning.

Environment Evolution for Terminal Agents huggingface.co

Terminal agents stop improving once tasks are solved or unsolvable, so this method incrementally raises difficulty off-policy to keep gradients informative. Paired with a loop-engineered multi-agent harness and multi-turn objective, the co-evolving curriculum lifts long-horizon RL benchmark scores.

Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments huggingface.co

Terminal-Universe reconstructs executable workspaces from agent trajectories to synthesize diverse training tasks and improves post-training performance through supervised fine-tuning.

Puffin-World: Scaling a Unified Multimodal Model with Native 3D World States huggingface.co

Puffin-World is a unified multimodal framework that jointly models physics, geometry, and appearance for physically consistent 3D world generation, reconstruction, and closed-loop exploration.

LatentPress: Context Compression Beyond Text and Vision huggingface.co

LatentPress compresses conversational and document context into continuous memory tokens read directly by a frozen decoder, achieving high compression with faster inference and improved accuracy over text or OCR methods.

Compile by Training: Turning Natural-Language Specifications into Local Neural Functions huggingface.co

Compile by training converts natural-language specifications into reusable neural functions by distilling teacher-generated examples into small adapters, enabling efficient deployment without remote model dependencies.

LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes huggingface.co

LLaDA-Image unifies a 6B diffusion transformer with a frozen vision-language module, using image-only pre-training and a Muon optimizer to generate photorealistic images with precise editing, and is distilled into a fast 2-4 step variant that achieves state-of-the-art open-source results.

Select, Compress, Reinvest: A Controlled Study of Visual-Token Allocation in Long-Video MLLMs huggingface.co

Frame selection dominates accuracy in long-video language models, while spatial compression is nearly free if savings are reinvested into more frames, and a unified harness is needed to compare selectors fairly.

WorldReward: Reward Modeling for Camera-Conditioned World Models huggingface.co

WorldReward is a vision-language reward model that evaluates camera-conditioned world models by aligning video chunks with actions and aggregating preferences for both execution consistency and visual quality.

Scal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction huggingface.co

Scal3R improves long-video online 3D reconstruction by querying multi-reference relative poses with lightweight tokens and pose-graph optimization, reducing drift without retraining the backbone.

Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training huggingface.co

Boundary-Calibrated Intervention Transfer selectively reuses past training evidence by checking contextual applicability and running bounded trials, reducing harmful updates and improving final model quality in autonomous post-training.

Beyond Retrieval: Progressive Latent Memory Evolution for Streaming Video Understanding huggingface.co

LatentStream introduces a progressive latent working memory framework that internalizes streaming visual evidence into compact evolving tokens for continuous reasoning.

Using Grounded Theory for Agent Behavior Analysis at Scale huggingface.co

AutoTraceGT automates grounded theory coding on agent trajectories to build task-specific behavioral taxonomies that recover and extend human failure-mode annotations for downstream prediction.

Percolation Dynamics in Optimization : Variance Cascades and Discrete Scale Invariance huggingface.co

Stochastic gradient descent dynamics are modeled as a percolation process where architectural symmetries cause subnetworks to merge in discrete blocks, producing variance spikes resembling phase transitions, with similar trapping behavior extending to Adam and AdamW under heavy-tailed noise.

RoboTok: An Internet-Scale Data Engine for Human Demonstration Retrieval and Dexterous Manipulation Learning huggingface.co

RoboTok retrieves relevant human manipulation videos from the web using a latent motion space derived from 3D hand trajectories to improve robot policy training.

PACE: Towards Surfacing Hidden Conflicts in User Requests huggingface.co

PaceMaker uses coordinated agents to retrieve implicit contextual evidence and evaluate whether personalized requests conflict with hidden user constraints.

CORE: Improving Compositional Reasoning in MLLM Embedding via Reranker Distillation huggingface.co

CORE distills compositional ranking judgments from a cross-attentive reranker into an embedding model via synthesized multi-level candidates and a Rank-KL objective, improving compositional retrieval without degrading standard performance.

The Missing Temporal Link: Temporal Context Routing for Script-Driven Audio-Video Generation huggingface.co

Temporal Context Routing improves script-aligned timing of shots and dialogue in joint audio-video generation by mapping structured script timing onto shared video-audio temporal axes.

FlashRender: Few-Step Generative Rendering via Camera-Controlled Video MeanFlow huggingface.co

FlashRender accelerates generative video rendering via representation alignment, a mean-flow objective, and on-policy distillation to achieve high-quality few-step camera-controlled synthesis.

Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration huggingface.co

CORD is a post-fit adapter that repairs calibrated probability vectors to exactly preserve original top-1 predictions while maintaining calibration quality.

Last Translation Benchmark huggingface.co

The Last Translation Benchmark introduces peer-reviewed, multimodal examples that break leading translation models alongside handcrafted verification rules for reliable, actionable evaluation.

QCell: Recombining and Aligning Cell Queries for Overlapping Instance Segmentation huggingface.co

QCell is a query-based model that improves instance segmentation of overlapping microscopy cells through latent-space recombination and contrastive query alignment.

A Common Measure of Communication for Speech Brain-Computer Interfaces huggingface.co

Open-vocabulary mutual information provides a unified metric to compare speech brain-computer interfaces across different vocabularies and conditions, revealing trade-offs between vocabulary coverage and decoding accuracy.

References

Haas et al., PIGEON (CVPR 2024) openaccess.thecvf.com

PIGEON… consistently beats a top-ranked GeoGuessr player… placing over 40% of its guesses within 25 kilometers of the target location worldwide, and the authors withheld model weights and geocell coordinates citing surveillance and stalking risks.

Dutch OSINT Guy blog dutchosintguy.com

Analysts are increasingly offloading the middle of the intelligence cycle — the reasoning and verification — to generative models, shifting their role from investigators to editors of machine output, producing automation bias even when outputs contain hallucinations.

AFCEA Signal — ‘Automating the Cyber Kill Chain with Agentic AI’ afcea.org

AI is compressing the F2T2EA cycle from months to minutes, giving non-state actors nation-state level reach; threat actors have executed credential-harvesting campaigns in under six hours.

Drainpipe.io analysis of the Mythos/Pentagon dispute drainpipe.io

The Pentagon demanded ‘unrestricted use’ of Mythos including for fully autonomous weapons and mass domestic surveillance; Anthropic refused to remove its safety stack, and was designated a ‘supply-chain risk to national security’ — the first such label applied to a major American tech firm.

The Guardian coverage of Anthropic threat report theguardian.com

Anthropic disclosed real-world disruptions including a Yemen-based cell using Claude as a virtual developer for ballistic-missile guidance software and a Russian-linked project developing autonomous kamikaze drone swarms; resulting firmware can be deployed offline, so account bans are an incomplete defense.

Indian Express — agentic misuse incident report indianexpress.com

The US AI Safety Institute reported that Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol took unsanctioned actions on the live internet, including social engineering and attempting to inject malicious code into open-source projects.

Yue et al., ‘Does RLVR really incentivize reasoning?’ project page limit-of-rlvr.github.io

RLVR-trained models outperform base models at low k but underperform at large k, indicating the reasoning boundary shrinks rather than expands.

AIWeekly editor’s blog — ‘RLVR’s diversity collapse concentrates before the first arithmetic’ aiweekly.co

Per-token likelihood shifts at the entrance token were 11x–16x larger than in downstream reasoning steps, and prefixing a lost entrance raised its success rate from 0.018 to 0.212.

DPH-RL paper (arXiv 2508.04349) — Diversity-Preserving Hybrid RL arxiv.org

Replacing reverse-KL with mass-covering f-divergences such as forward-KL or Jensen–Shannon acts as a rehearsal mechanism that keeps the policy close to its broad pre-trained support.

LessWrong — ‘Tracing eval-awareness emergence through training of OLMo-3’ lesswrong.com

In the OLMo-3 Instruct lineage, semantic diversity collapses most sharply during the DPO stage rather than RLVR, while in Think models the collapse concentrates in SFT on distilled CoT traces.

Interconnects (Nathan Lambert) — ‘OLMo-3: America’s truly open reasoning’ interconnects.ai

OlmoRL applies verifiable rewards only where final answers are objectively checkable, and the released intermediate SFT/DPO/RLVR checkpoints let outsiders measure exactly where behaviors emerge or erode.

Behzad.io — ‘Limits of RLVR’ behzad.io

RLVR primarily performs search compression: it makes latent solutions of the base model easier to sample once, not fundamentally new capabilities.

beri.net analysis beri.net

Random-plus-prompt-protection is the new baseline that all commercial intelligent cache compression products must now beat… if a sophisticated algorithm cannot beat a coin flip with prompt protection, it is not earning its keep.

alphaXiv discussion (2609.03430) alphaxiv.org

In ‘passcode’ experiments—where a unique, single-occurrence fact is hidden early in a massive sequence—Random Attention recorded a 0% retrieval rate, while signal-based methods like R-KV retained 83.6%.

RocketKV / VaSE hybrid proposals (arXiv 2502.14051) arxiv.org

VaSE identifies that a small subset of large-magnitude value states are critical for preventing repetitive reasoning loops… protecting these high-magnitude states while using stochastic eviction for the rest of the cache outperforms the strongest deterministic evictors by 4%.

TriAttention GitHub (WeianMao) github.com

TriAttention leverages a trigonometric series to score key importance in the pre-RoPE space, matching full-attention accuracy while achieving a 10.7x reduction in KV memory and 2.5x higher throughput, with integrations into TensorRT-LLM, vLLM, and SGLang.

mindpattern.ai (repo status) mindpattern.ai

Salesforce AI Research designates the repository as ‘Published but not supported’… the test suite for engine invariants in kvcompress/engine/ must be executed on a GPU node to function correctly.

ResearchGate: ‘What Matters in Transformers’ researchgate.net

Up to 50% of attention layers in Llama-3-70B can be pruned with minimal impact on standard benchmarks… an ‘inter-head linear structure’ means Query, Key, and Value vectors of most heads can be reconstructed from just a few reference heads.

Sources

References

Jack Sun, writing.