The 27 papers the radar tracked in September 2026. The current radar lives on the research page. Back to the radar →
A four-stage protocol fingerprints anonymous API models using archived launch configs, tokenizer diffs, and behavioral probes, no weight access required. Run against known releases it matched 7 of 10 exactly, and its prospective call on an unlabeled August drop landed on the right version line before the vendor confirmed it. Handy checklist if you're vetting an unlabeled endpoint before you sign a vendor contract.
Training OLMo 3 on preference pairs from different teacher models, the authors found student sycophancy tracks teacher sycophancy almost linearly, across DPO and six other preference-optimization objectives. No single bad example is the culprit either. The signal is smeared across the whole dataset, so removing a small chunk of it doesn't help. If a fine-tune inherited an agreeable base model, this is probably why.
Replaying 55 archived coding-agent trajectories, researchers found that instructions, tool outputs, and agent state compress and decay at different rates, so two agents with identical context windows can end up delivering very different context to the model. Compression tricks tuned on one task set didn't carry over to new ones.
A benchmark of collaborative chess tasks shows that integrating continuous subagent representations directly into language models via learned state tokens outperforms text-based verbalization and scales effectively.
A streaming, LLM-based end-to-end model unifies speaker-attributed speech recognition and diarization for low-latency real-time applications.
DisCo is a research agent that distills operational knowledge into reusable skills, significantly improving autonomous ML research performance across benchmarks.
EarlyEval predicts agent outcomes from intermediate behavior to reduce evaluation cost by halting runs early with minimal accuracy loss.
Declarative Attention lets language models declare relevant context regions during reasoning to skip most KV cache reads, reducing attended tokens with small accuracy trade-offs.
A specialization pipeline combining curated problems, synthetic reasoning, supervised fine-tuning, and reinforcement learning trains competitive programming models that exceed top human scores on IOI benchmarks using iterative test-time…
Cliff improves reinforcement learning with verifiable rewards by using an off-the-shelf language model to detect the first reasoning error and shaping token-level advantages accordingly.
PaperCompiler converts research papers into structured repository specifications that preserve method logic and cross-file consistency, improving code fidelity and reducing critical errors.
SolarWM provides an open framework and unified training recipe for building interactive video world models across diverse data sources and generator backbones, enabling long-horizon real-time rollouts.
CORD is a post-fit adapter that repairs calibrated probability vectors to exactly preserve original top-1 predictions while maintaining calibration quality.
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…
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.
Debias-SparseGPT reduces pruning-induced demographic bias in compressed large language models by incorporating representational debiasing during post-training sparsification.
RoboTok retrieves relevant human manipulation videos from the web using a latent motion space derived from 3D hand trajectories to improve robot policy training.
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.
VeriPhy verifies generated video by compiling prompts into typed physical obligations, executing frozen expert analyses with provenance tracking, and mapping evidence to auditable three-valued verdicts.
The study formalizes multi-agent LLM coordination via bilevel games and stochastic memory reflection, introducing a grounded evaluation gate and SRMA algorithm with convergence guarantees, validated on SWE-bench.
ShallowStream uses shallow MLLM layers to build a lightweight streaming index, reducing latency while maintaining accurate video retrieval.
Teacher-Gated On-Policy Distillation verifies teacher reliability per prompt via verifier-scored probes, routing to dense distillation or verifier-grounded reinforcement learning to improve training efficiency and accuracy.
A Graph Machine architecture uses sparse dynamic routing and differentiable pointer chasing to maintain linear state complexity, enabling efficient replacement of dense Transformer layers with minimal loss degradation.
A governed analytics framework pairs language models for intent interpretation with deterministic policy execution of pre-approved programs, achieving full answer-and-evidence reliability where runtime-planning agents failed.
Influence-guided response rewriting of selected training examples produces stronger and more persistent behavioral shifts in language models than conventional reweighting, highlighting the broader intervention leverage of influential data.
Benchmarks for LLM-generated GPU kernels decide correctness with a few random inputs and a loose floating-point tolerance, and their verdicts now feed leaderboards and reinforcement-learning rewards.
Vision-Language-Action (VLA) models demonstrate strong generalization in robotic manipulation and navigation, but existing fine-tuning methods provide limited safety guarantees.