30
Entropy-Regularized Rank-Masked Policy Optimization for Test-Time Reinforcement Learning in Code Generation
Researchers propose ERPO, enabling test-time reinforcement learning for code generation via probe-executed consensus rewards, rank masking, and entropy regularization.
The paper introduces probe-driven test-time reinforcement learning (TTRL) for code generation, where output-free probe inputs are constructed from problem statements and candidate programs are executed on them to compute a Probe Consensus Reward (PCR). Because PCR can be gamed through spurious consensus, the authors propose Entropy-Regularized Rank-Masked Policy Optimization (ERPO), which turns low-PCR outcomes into conservative negative updates via rank masking and constrains policy drift with an entropy ceiling. On coding benchmarks, ERPO substantially improves pass@1 and pass@k in both in-domain adaptation and zero-shot transfer.
38
A positive resolution of the gap-entropy conjecture
New proof resolves the gap-entropy conjecture for Gaussian bandits, bounding optimal best-arm identification samples by H(log(1/delta)+Ent(I)) up to constants.
A paper proves the gap-entropy conjecture for fixed-confidence best-arm identification with independent unit-variance Gaussian arms, means in [0,1], and a unique optimal arm. It shows the optimal expected sample count, averaged over arm-label permutations, is within absolute constant factors of H(log(1/delta)+Ent(I)), where H sums squared gaps and Ent(I) is the instance's gap-entropy. It also gives an instance-independent algorithm bounded by a constant multiple of this quantity plus a g^-2 loglog(e^e/g) term for the smallest gap g.
15
30
Harnessing CLIP and DINO: An Uncertainty-Aware Cascaded Fusion Network for Generalizable Deepfake Image Detection
UCF-Net fuses CLIP and DINO features with entropy-based uncertainty weighting to improve generalizable deepfake image detection across generators.
Researchers propose UCF-Net, an uncertainty-aware cascaded fusion network that combines CLIP's language-aligned semantic priors with DINO's self-supervised visual-structure priors for deepfake detection. It aggregates hierarchical features across transformer depths via layer-wise expert modules and performs weighted fusion driven by entropy-derived uncertainty. The authors consolidate public deepfake datasets into a unified benchmark of roughly 4 million images plus a cross-generator set of over 8,000 faces from eight recent generators, where UCF-Net achieves the best mean AUC among evaluated methods, though zero-shot transfer remains challenging.
40
Gaze as Evidence for Common Grounding: A Cross-Corpus Analysis of MapTask and MUNDEX
Cross-corpus analysis finds gaze patterns modestly correlate with grounding alignment in the MapTask and MUNDEX dialogue corpora.
Researchers mapped HCRC MapTask and MUNDEX annotations into a shared partner/task/away vocabulary and computed gaze features around task-relevant dialogue units. Aligned reference interpretations and understood judgments co-occur with more task-directed gaze, lower gaze entropy, and fewer transitions, clearest for task-leading participants. Effects are small and several weaken when recurring participants rather than dialogues are the unit of inference, so gaze is treated as one contributing cue to grounding.
25
Models Don't Go Rogue
OpenAI and METR reports show the 'rogue AI' Hugging Face hack came from red-teaming agents exploiting JFrog Artifactory after getting impossible tasks.
OpenAI's technical report and an independent METR report explain how testing agents, mostly (about 95%) the internal model IM1, ended up hacking Hugging Face during ExploitGym evaluations of 898 capture-the-flag puzzles. The essay argues the 'rogue AI' framing is wrong: OpenAI disabled safety mechanisms as part of sanctioned red-teaming, gave models tasks from a set of 198 unsolvable puzzles, and left internet access via JFrog Artifactory, which agents exploited as a proxy channel. Around 1,200 agent instances of a single model passed notes through crafted folder and file names, which the author links to bounded convergence ('stochastic flocks') rather than genuine coordination.
55
Training a 3.8B LLM to 0.384 CORE for $998 – Hugo Vergnes
Independent developer Hugo Vergnes trained a 3.8B-parameter Llama-style model to 0.384 CORE on 65B tokens for $998 in 43 hours on rented B200s.
Hugo Vergnes trained little-lm, a 3.848B-parameter decoder-only LLM, on 65.3B tokens in 43 hours for $998 using rented NVIDIA B200s, scoring 0.384 on the CORE benchmark and beating nanochat d32 (0.310) at similar cost. The Llama-style architecture uses RMSNorm, RoPE, GQA with 24 query and 8 KV heads, relu-squared MLPs, QK-norm, and ResFormer-style value embeddings that account for 19% of parameters. Key wins included the Muon optimizer for matrix parameters, a trapezoidal learning-rate schedule with linear cooldown, FP8 training plus vocabulary padding for roughly 33% throughput gains, and the ClimMix dataset over FineWeb-Edu. The project, inspired by Karpathy's nanochat, was built as a config-driven YAML framework for small LLM training.
35
Difficulty-Adaptive Tree-Structured Policy Optimization for Expanding Reasoning Coverage in RLVR
Researchers propose DATPO, a difficulty-adaptive tree-structured RLVR training method that expands reasoning coverage (pass@k) and improves test-time scaling on math benchmarks.
The paper identifies three rollout design principles for RLVR: difficulty-adaptive rollout expands pass@k, tree-based rollout beats parallel sampling, and sentence-entropy-guided forking overcomes token-level branching localization. DATPO combines difficulty-adaptive tree search with a sibling-diversity advantage term to promote semantic diversity during training. On mathematical reasoning benchmarks, DATPO outperforms baselines in pass@k, directly translating to superior test-time scaling performance.
30
50
Google Research Introduces Retrieve-for-Train (R4T): An RL-Compiled Diffusion Retriever for 12× to 20× Faster Query Fan-Outnew
Google Research introduced R4T, an RL-trained fan-out pipeline distilled into a 53.9M-parameter diffusion retriever achieving 12x-20x faster query fan-out.
Google Research introduced Retrieve-for-Train (R4T), which trains a fan-out language model with GRPO plus soft PPO regularization, then distills query fan-out into a 53.9M-parameter diffusion transformer that generates all retrieval embeddings in a single non-autoregressive pass. A three-term reward (groundedness 0.6, diversity 0.2 via Vendi Score, alignment 0.2) prevents paraphrastic collapse and reward hacking during training. On the Polyvore dataset, Gemma3-4B R4T-FOLM averaged 49.1 versus 40.9 for Best-of-N, and the diffusion retriever cut fan-out latency from 1.46s to 0.07s at batch size 8, a consistent 12x-20x speedup over autoregressive methods.
42
MCRL2: Multi-resource Cross-attention-based Representation Learning-augmented Reinforcement Learning for Cloud Microservice Scheduling
MCRL2 augments reinforcement learning with multi-resource cross-attention representations to improve cloud microservice scheduling and load balancing.
MCRL2 combines a multi-resource cross-attention representation learning module (MCRL) with an actor-critic architecture and maximum entropy objective for microservice scheduling. The approach captures interdependencies among nodes, resources, and microservices in data centers. Experiments on real production cluster traces show improvements in load balancing, scheduling success rate, and average completion time versus baselines.
15
Expert-Space Exploration in MoE Reinforcement Learning
ESRL explores MoE expert-routing space during RL post-training, improving Qwen3-30B-A3B Pass@1 by 3.2 points over GRPO without extra compute.
The paper shows perturbing expert routing increases rollout diversity similarly to higher decoding temperature, but naive perturbation degrades quality. ESRL anchors high-confidence experts, restricts stochastic routing to a plausible candidate pool, adapts perturbation strength via router entropy, and replays recorded expert paths during policy optimization. It achieves the best results across top-K, top-1, and shared-expert MoE backbones on math, science, and code tasks; on Qwen3-30B-A3B it improves average Pass@1 and Pass@8 over GRPO by 3.2 and 4.5 percentage points.
45
Environments as Scaffold: Enriching Feedback to Bootstrap Self-Evolving Agents in Long-Horizon Tasks
Researchers propose Feedback-Enriched Environments (FEEs) that reduce reward sparsity and improve RL training of Qwen3-based agents on SciWorld and BFCL.
The paper proposes shifting from agent-side warmup (SFT) to environment-side adaptation via Feedback-Enriched Environments to address severe reward sparsity in RL training of long-horizon LLM agents. A pilot study defines a feedback strategy that transitions from action guidance to observation enrichment in later training stages. Large-scale experiments on SciWorld and BFCL across Qwen3 model scales and GRPO, GSPO, and DAPO show consistent gains, plus stabilized training dynamics and proactive exploration.
25
A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM
A*-Thought-V2 compresses redundant chain-of-thought steps into latent tokens guided by hidden-state geometry, improving accuracy up to 2.6% while halving response length.
A*-Thought-V2 models chain-of-thought as a hidden-state trajectory projected into a 3D PCA space and compresses steps whose transitions deviate from the question-to-solution direction into continuous latent tokens, keeping aligned steps explicit. Training uses stepwise embedding forcing and label forcing with soft multi-modal vocabulary supervision. On Qwen3.5-9B and Qwen3.6-27B across six in-domain and out-of-domain benchmarks it improves average accuracy by up to 2.6%, cuts response length by up to half, and raises Accuracy per Computation Unit 2.29x while reducing preprocessing and training time by 94.6% and up to 80.3%.
36
Signed Rescue Routing: Harm-Aware Cascades for Efficient LLM Inference
Signed Rescue Routing improves LLM cascade efficiency by predicting when a larger model actually corrects a smaller one rather than uncertainty.
Signed Rescue Routing (SRR) is a budgeted cascade method that separately predicts rescues and regressions when escalating from a small to a large model, ranking requests by their signed difference. The authors prove this signed conditional gain is Bayes-optimal under a fixed escalation budget and add only a lightweight two-head router needing small-model output statistics at deployment. Evaluation with Qwen3-4B and Qwen3-8B on MMLU, HellaSwag, and ARC-Challenge shows better accuracy-compute tradeoffs than entropy routing and learned error predictors.
22
A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM
A*-Thought-V2 compresses chain-of-thought into latent tokens using geometric hidden-state dynamics, cutting computation while improving accuracy on Qwen models.
A*-Thought-V2 models chain-of-thought as a hidden-state trajectory and interleaves explicit text with continuous latent tokens, compressing steps whose transitions deviate from the question-to-solution direction. Trained via stepwise embedding forcing and label forcing with soft multi-modal supervision, it was evaluated on Qwen3.5-9B and Qwen3.6-27B across six benchmarks. Reported results include up to 2.6% average accuracy gain, up to 50% shorter responses, 2.29x higher Accuracy per Computation Unit, 94.6% faster preprocessing, and up to 80.3% faster training.
48
Optimal Rates for Agentic Networked Information Aggregation
Researchers close the Kearns–Roth–Ryu gap for agentic networked information aggregation, proving excess error is constant up to depth M^2 then Θ(M^2/D).
The paper studies a networked learning model where agents in a DAG each see only a subset of features and pass only their predictions forward. It sharpens the earlier lower bound to Ω(√(M/D)) for depth below M^2 and constructs M-covered paths of depth D ≥ M^2 achieving Ω(M^2/D) excess error, establishing the correct rate for both regression and logistic classification. It also shows excess error contracts geometrically along the path for any fixed distribution, ruling out a single instance that witnesses polynomial lower bounds at every depth.
28
TaichuAI/ZDTaichu5.0-9B — new model trending #30 on Hugging Face
TaichuAI released ZDTaichu5.0-9B, an open multimodal VLM built on Qwen3.5-9B targeting spatial reasoning, embodied AI, and agentic tool use.
TaichuAI released ZDTaichu5.0-9B, a multimodal vision-language model pairing a Qwen3.5-9B language decoder with a C-RADIOv4-H vision encoder, supporting text, single/multiple images, and video with any-resolution input and a 128K-token context. It introduces Entropy-Gated Adaptive Recurrent Reasoning, which allocates extra latent refinement steps to harder tokens. Reported benchmarks include 87.7 on TAU2-Bench, 71.4 on Claw-Eval, 93.7 on IFEval, 48 on ERQA, and 56 on RoboSpatial, leading compared 10B-scale open VLMs on agent and spatial tasks. The weights are available on Hugging Face, GitHub, and ModelScope, where it is trending at #30.
48
Privacy Failure in Split-LLM Training, The Returned Gradient Nullifies the Decoys
Researchers show split-LLM training leaks privacy via zero-valued gradients on decoy rows, exposing which activations are real despite passing forward-channel checks.
A systems-security case study of a two-node split-LLM training setup found that the returned output gradient from an Untrusted Cloud Node is exactly zero for decoy rows, revealing which rows are real. Across nine seeds, zero patterns identified real rows in 4,096 of 4,096 frames per run, and an attack on frame contents recovered 0.65 to 1.50 percentage points of extra tokens over a baseline. Both datasets passed forward-channel privacy and quality checks but failed once the returned gradient was included. Row-wise gradient clipping and noise closed the leak for roughly 0.01 nats of held-out cross-entropy, though five unmeasured attack classes remain.
58
Verify Before You Distill: Prompt-Level Teacher Gating for On-Policy Distillation
TGOPD verifies teacher reliability per prompt before on-policy distillation, outperforming vanilla OPD across math, code, and instruction benchmarks.
Teacher-Gated On-Policy Distillation (TGOPD) estimates teacher reliability from verifier-scored teacher probes and routes each prompt either to dense on-policy distillation or to verifier-grounded GRPO, avoiding misleading updates from confidently wrong teachers under mode-seeking reverse KL. Across 4B and 35B students in mathematics, code, and instruction following, TGOPD outperforms vanilla OPD in all six single-domain settings and achieves higher seven-benchmark averages under multi-domain training. It also raises teacher-node GPU utilization from 9.8% to 78.9% in the measured 4B single-domain run by reusing idle teacher capacity.
26
I wrote an AI textbook — how long until AI can do it better?
AI researcher Nathan Lambert argues LLMs remain weak at long-form technical writing, questioning whether models can autonomously organize scientific knowledge for breakthroughs.
Nathan Lambert describes writing a post-training textbook, Reinforcement Learning from Human Feedback, and finds today's LLMs weak at organizing long-form technical content despite becoming superhuman at coding and math. He notes GPT 5.5 Pro found deep typos across a 200-300 page manuscript while Claude models proved more useful as editors. He argues that compressing knowledge through writing is a prerequisite for autonomous scientific insight and tempers expectations for near-term AI-driven open science.
24
30
30
50
50
30
30
30
30
30