Sources

How Claude is uplifting biomolecular modeling anthropic.com

An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics huggingface.co

A natural-language proof-generation pipeline using post-trained Nemotron 3 Ultra checkpoints achieves gold-medal performance on IMO 2026 through iterative verification and refinement without external tools.

EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents huggingface.co

EvoSafeHarness optimizes deployable safety harnesses by jointly searching natural-language policies and executable logic tailored to a frozen model and target domain, improving safety-utility trade-offs across agent benchmarks.

IdeaAMBIG: Benchmarking Implementation-Critical Gaps in Research-Idea Specifications huggingface.co

IdeaAMBIG tests whether research method descriptions are concrete enough for a language model to implement, and finds that spotting the missing pieces, not filling them in, is where models struggle most. The benchmark scores defect localization and clarification actions across research idea specifications.

SenseNova-U1.5: Towards Native Unified Visual Intelligence huggingface.co

SenseNova-U1.5 handles visual understanding, reasoning, and generation in a single 8B network without image encoders or VAEs, relying on spatially coherent patch reconstruction. Training combines curated data, multi-expert on-policy distillation, and reinforcement learning to support bilingual text rendering and infographic generation.

Negative Self-Distillation: Learning to Reason by Avoiding Flaws huggingface.co

Negative Self-Distillation pushes a model away from its own bad reasoning traces rather than toward good ones, using a dynamic gate that fires only on reasoning-critical tokens. The gating preserves general linguistic ability while unlearning error patterns during self-bootstrapping reinforcement learning.

Beyond Solver Verdicts: Generative Reward Models for Autoformalization huggingface.co

Neurosymbolic pipelines often accept formal translations that hit the right answer for the wrong reasons, a failure mode dubbed Verdict-Preserving-Unfaithfulness. A generative verifier scores reference equivalence without needing a Z3 oracle, lifting downstream accuracy on autoformalization tasks.

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction huggingface.co

NCP-ArchPreview trains a latent-space language model to predict both the next token and the next concept, drawn from a product-quantized concept vocabulary. The joint objective improves pretraining efficiency and downstream performance, with a DFlash2 drafter enabling faster autoregressive generation.

World in World: Explore the World with World Models huggingface.co

World in World steers frozen autoregressive video world models with flexible camera and time control by routing extra visual evidence through native self-attention. A correspondence router and per-channel attention guidance support long-horizon revisiting without any fine-tuning of the base model.

Recursive Code World Models: Building Complex Worlds through Recursive Scene Programs huggingface.co

Recursive Code World Models turn a single image into a complex 3D scene by emitting executable scene programs, alternating global layout with local detail passes under shared camera alignment. A vision-language coding agent drives the recursive reconstruction loop.

SpatialBlock: Enhancing Spatial Intelligence in LVLMs via Synthetic Block-Stacking Problem huggingface.co

Large vision-language models trained on synthetic block-manipulation tasks improve 3D spatial reasoning and generalize to real-world visual tasks.

Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning huggingface.co

Optimized supervised fine-tuning data composition enables reasoning models to consistently process and respond in diverse non-English languages without requiring reasoning supervision in each target language.

Memory as Plans: World-Action Modeling with Memory-Grounded Planning huggingface.co

MaP-WAM improves non-Markovian robotic manipulation by separating memory-grounded planning from plan-conditioned execution, using compact episodic segment records and progress-calibrated action chunks to maintain fixed inference latency.

TempCloze: Can Video-LLMs Identify the Missing Middle? huggingface.co

TempCloze evaluates visual temporal reasoning in Video-LLMs by requiring identification of missing video segments from distractors targeting semantics, alignment, and progression.

HyQuant: Hybrid-Precision Quantization for LLM Attention huggingface.co

HyQuant improves low-bit LLM attention quantization by preserving critical vertical-line tokens and local windows in high precision while quantizing the rest, maintaining accuracy with low overhead.

Generative Late-Interaction Embeddings For Visual Document Retrieval huggingface.co

Generative Late-Interaction Embeddings compress visual document retrieval vectors by learning a small basis set that regenerates full embeddings on demand, improving accuracy under tight storage limits without retraining the encoder.

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models huggingface.co

Using a controlled autoregressive testbed, the study analyzes task-specific validation losses during multimodal pretraining to evaluate how image tokenizer design affects joint text-image modeling and downstream performance.

ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation huggingface.co

ActReview is a rebuttal-guided post-training framework that generates diagnostic claims and concrete revision suggestions for peer review by leveraging author responses as latent supervision.

CARDEA: Auditable Reasoning Grounded in Spatial Evidence for End-to-End Coronary Angiography Interpretation huggingface.co

A unified vision-language model for coronary angiography uses chain-of-box reasoning and reinforcement learning with verifiable rewards to provide auditable diagnoses and improve zero-shot report generation.

X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation huggingface.co

X-AuT progressively prunes audio-encoder layers in speech large language models and restores accuracy via behavioral probes, representation alignment, cross-scale distillation, and LoRA adaptation.

FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation huggingface.co

FreeFlow is a hierarchical transformer for optical flow that eliminates task-specific inductive biases and achieves state-of-the-art accuracy using window, shifted-window, and global attention.

Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity Linking huggingface.co

A reasoning-capable vision-language model that iteratively retrieves and reasons over Wikipedia improves multimodal entity linking for rare entities defined by knowledge-graph structure.

UniH^3: Unifying Hierarchical Homogeneity and Heterogeneity for All-in-One Medical Image Restoration huggingface.co

UniH3 unifies hierarchical homogeneity and heterogeneity for universal medical image restoration via memory-based homogeneity priors and a heterogeneity balancer.

Mi-Ripple: Restoring Images Degraded by Iterative AI Editing huggingface.co

Mi-Ripple reduces digital ripple artifacts in edited images by separating lattice artifacts from texture and applying targeted spectral filtering and reference cleaning.

MetroLLM-Bench: Evaluating Language Models as Transit Kiosk Runtimes huggingface.co

A benchmark for transit-kiosk policy reasoning shows that small parameter-efficiently fine-tuned language models can match larger models on structured tool-use and fare-quoting tasks.

DRG-MAPPO: Hierarchical Dynamic Role-Graph Multi-Agent Reinforcement Learning for Cooperative Air Combat huggingface.co

A hierarchical multi-agent reinforcement learning framework combining graph attention and dynamic role assignment improves tactical coordination and win rates in air combat.

Adaptive Bridge: A Proxy-Based Decoupling Layer for Mitigating DDS Backpressure in ROS 2 huggingface.co

A proxy layer isolates critical ROS 2 subscribers from degraded ones via topic splitting and dynamic rate control to eliminate DDS backpressure.

References

Towards AI — ‘Claude’s protein design hit rate was 26.8%. One target returned 0 for 90’ pub.towardsai.net

A binomial test of the per-target counts strongly rejects the idea of a single shared hit rate (p ≈ 1.8 × 10⁻⁶³)… for a novel, unseen target there is a 38% chance the success rate will fall below the 10% industry floor, with a 20.1% probability of returning zero binders in a standard 30-design batch.

AllSci — Autonomous Protein Binder Design at Anthropic allsci.com

Claude produced zero binders for MBP out of 90 total designs… the confidence scores for the ‘dead’ MBP designs were virtually identical to those for high-performing binders against other targets.

Tamarind Bio — The open models that beat AlphaFold3 tamarind.bio

In the critical area of antibody-antigen prediction, independent benchmarks show that models like ESMFold2 and OpenDDE actually outperform AF3, reaching success rates of over 70% compared to AF3’s ~48% on identical inputs.

ARI — Reward Hacking: How AI Exploits the Goals We Give It (context on KernelBench-Verified) ari.us

Frontier models frequently engage in sophisticated reward hacking when writing GPU code… while some models claimed a 1.43× speedup under standard protocols, their actual performance dropped to 0.88× when tested against hidden distributions and more robust TF32-enabled baselines.

Unite.ai — Anthropic Launches Life Sciences Verification Program in Beta unite.ai

The bio-optimized ‘Mythos 5.1’ is restricted to a ‘trusted access program’ developed in collaboration with the US government… some independent researchers have criticized this move, questioning whether restricted access will stifle academic innovation while favoring well-funded ‘trusted’ corporate partners.

Ace & Company — Adaptyv raises $40M Series A aceandcompany.com

Adaptyv Bio’s role in providing $1 million in automated experimental validation—capable of testing 5,000+ proteins at no cost to entrants—is viewed as a game-changer for democratizing access to high-throughput screening… Adaptyv’s recent $40 million Series A positions them as a central hub for AI-native drug discovery.

Inference by Sequoia (Substack) inferencebysequoia.substack.com

Huawei’s Celia and Xiaohongshu’s dots-note-3.0 received official scores of 42/42 from IMO organizers after completing the exam under strict competition conditions.

Digital Applied — ‘IMO 2026 Perfect Scores & AI Benchmark Saturation’ digitalapplied.com

OpenAI’s GPT-5.6 Sol, Claude Fable 5, and Kimi K3 all achieved perfect scores when tested in autonomous harnesses, though these self-administered results often use different grading standards, such as Claude-based automated agents, rather than the official IMO jury.

The Moonlight review of the Nemotron IMO paper themoonlight.io

Both the model’s internal verifier and independent LLM-based jury agents (using GPT-5.5 and Claude 4.8) significantly overestimated the quality of the failed proofs on Problems 3 and 6, marking them as correct when human graders identified subtle logical gaps.

EmergentMind — DeepSeekMath-V2 topic page emergentmind.com

A meta-verifier reviews the verifier’s critiques to prevent ‘hallucinated issues’—where a verifier wrongly penalizes a valid proof—improving verification quality scores from 0.85 to 0.96.

OrcaRouter blog — Nemotron-3 Labs Ultra Math RL orcarouter.ai

Running the full BF16 checkpoint demands approximately 1.5 TB of Blackwell-class HBM, typically necessitating an 8x B200 node or a cluster of 16x H100 GPUs; no outside lab has yet confirmed a successful end-to-end replication of the 30/42 score.

Stanford HAI — AI Index Report 2026 hai.stanford.edu

Seventeen Fields Medalists, including Terence Tao, issued a declaration warning that the ‘rush to produce answers’ might undermine the shared understanding essential to mathematical discovery… on Riemann-Bench, which evaluates models on novel research-level problems, all frontier AI systems currently score below 10%.

AgentDyn paper (arXiv:2602.03117) arxiv.org

On more complex, long-horizon tasks, almost all defenses that succeeded in AgentDojo failed to maintain security or utility

PromptArmor (alphaXiv 2504.11703) alphaxiv.org

PromptArmor uses a secondary LLM as a firewall to excise injections… claim near-zero ASR on AgentDojo by fundamentally altering the agent architecture

newsaint.id analysis of EvoSafeHarness architecture newsaint.id

the evolution process relies on a ‘two-level resource ledger,’ where the outer optimizer (often a high-frontier model like Claude 3.5 or GPT-4o) consumes significant inference costs… certain complex adversarial intents still bypass the evolved logic, and the framework currently lacks a human-in-the-loop review for the synthesized natural-language policies

cctest.ai review cctest.ai

Open questions remain regarding the ‘non-monotonic’ benefit variation across different base models, which may require users to perform extensive local searches to find the optimal harness for their specific deployment

Hacker News discussion news.ycombinator.com

remote harnesses are ‘backwards’ because they force agents to reach into local infrastructure from the outside, complicating security and management

AI Weekly digest aiweekly.co

EvoSafeHarness cuts agent attack success from 45.6% to 10%

Sources

References

Jack Sun, writing.