1. DINOv3
DINOv3 is a selfâsupervised vision foundation model that scales data and model size, introduces a Gram anchoring loss to preserve dense patch consistency during long training, and adds postâhoc tweaks for resolution, size, and text alignment. With a frozen backbone, it sets new results across dense and global tasks without taskâspecific fineâtuning.
Key idea: Gram anchoring. Regularize patch features by matching the studentâs patchâfeature Gram matrix to that of an earlier âGram teacher,â improving local consistency while leaving global features flexible. Implemented late in training, it can repair degraded dense features.
Immediate dense gains. Applying Gram anchoring quickly boosts VOC and ADE20k segmentation and strengthens robustness on ObjectNet; using higherâresolution teacher features adds further improvements.
Highâres teacher trick. Compute teacher features at 2Ă input resolution, then downsample to smooth patch similarities; this yields better Gram targets and extra mIoU on ADE20k.
Frozenâbackbone SOTA. A lightweight PlainâDETR decoder on top of a frozen DINOv3 backbone reaches 66.1 mAP on COCO, rivaling or beating specialized detectors with hundreds of millions of trainable parameters.
General, scalable suite. The release covers multiple model sizes and training recipes designed to serve diverse resource and deployment needs while outperforming prior selfâ and weaklyâsupervised foundations.
2. Capabilities of GPT-5 on Multimodal Medical Reasoning
A controlled, zeroâshot CoT evaluation positions GPTâ5 as a generalist medical reasoner across text and image inputs. Using standardized prompts and splits, GPTâ5 consistently beats GPTâ4o and smaller GPTâ5 variants, with especially large gains on multimodal expertâlevel QA.
Unified zeroâshot CoT setup. The authors fix prompts, exemplars, and answer formats for both QA and VQA to isolate the model upgrade.
Text benchmarks: new highs. On MedQA (US 4âoption) GPTâ5 reaches 95.84% (+4.80% vs GPTâ4o). On MedXpertQAâText, GPTâ5 improves reasoning by +26.33% and understanding by +25.30% over GPTâ4o. MMLUâMedical is at or near ceiling, with notable gains in Medical Genetics and Clinical Knowledge.
USMLE practice sets: strong clinical management. GPTâ5 tops all baselines on Steps 1â3, with the largest margin on Step 2 (+4.17%), averaging 95.22% (+2.88% vs GPTâ4o).
Multimodal reasoning: big leap and humanâplus. On MedXpertQAâMM, GPTâ5 gains +29.26% in reasoning and +26.18% in understanding over GPTâ4o and surpasses preâlicensed human experts by +24.23% and +29.40%, respectively. A worked case shows coherent synthesis of CT findings and clinical context to recommend a Gastrografin swallow.
Caveats and outlook. GPTâ5 is slightly below GPTâ5âmini on VQAâRAD, possibly due to calibration on small radiology datasets. The discussion cautions that standardized tests differ from messy clinical reality and calls for prospective studies and deployment calibration.
3. M3-Agent
A framework for agents that watch and listen to long videos, build entity-centric memories, and use multi-turn reasoning to answer questions. M3-Agent stores both episodic details and distilled semantic knowledge in a multimodal memory graph, then learns a retrieval-reasoning policy with RL. The authors also introduce M3-Bench, a long-video QA benchmark with robot-view and web videos. Results show consistent gains over strong prompting baselines.
Entity-centric long-term memory. Builds a multimodal graph with nodes for text, faces, and voices, plus edges for relations and cross-modal identity links. Episodic entries record events, and semantic entries capture attributes and world knowledge. Conflicts are resolved by weight-based voting to keep memory consistent.
Online memorization with identity tools. Processes video streams clip by clip, using face recognition and speaker identification to maintain persistent character IDs, then writes grounded episodic and semantic memories keyed by those IDs.
RL-trained control for retrieval and reasoning. A policy model decides when to search memory and when to answer, performing iterative, multi-round queries over the memory store rather than single-shot RAG.
M3-Bench for long-video QA. 100 robot-perspective videos and 929 web videos with 6,313 total QA pairs targeting multi-detail, multi-hop, cross-modal, human understanding, and general knowledge questions.
State-of-the-art results and ablations. M3-Agent beats a Gemini-GPT-4o hybrid and other baselines on M3-Bench-robot, M3-Bench-web, and VideoMME-long. Semantic memory and identity equivalence are crucial, and RL training plus inter-turn instructions and explicit reasoning materially improve accuracy.
Editor Message
I am launching a new hybrid course on building effective AI Agents with n8n. If you are building and exploring with AI agents, you donât want to miss this one. You can use code DAIRX20 for an extra 20% off (available only for the next 24 hrs).
4. TRImodal Brain Encoder
A triâmodal, multiâsubject, nonlinear encoder that fuses text, audio, and video features with a transformer to predict timeâvarying fMRI responses to natural movies. It took 1st place in the Algonauts 2025 brain encoding competition and shows the strongest gains in the highâlevel associative cortex.
Model recipe. Extracts timed embeddings from Llamaâ3.2â3B (text), Wav2VecâBERTâ2.0 (audio), and VâJEPAâ2 (video), groups layers, projects to shared width, and feeds a windowed sequence to an 8âlayer transformer with subject embeddings and adaptive pooling to 1,000 cortical parcels. Modality dropout trains robustness when inputs are missing.
State of the art. Ranked 1st of 263 teams on the public leaderboard; large margin over the field. Mean Pearson inâdistribution on Friends S7 is 0.3195, and it generalizes to outâofâdistribution films, including cartoons, documentaries, and silent blackâandâwhite clips.
Noise ceiling. Achieves a normalized Pearson of about 0.54 on average, near ceiling in auditory and language cortices, indicating more than half of the explainable variance captured.
Why multimodal matters. Unimodal encoders trail the triâmodal model; combining any two helps, and all three help most. Biggest gains appear in associative PFC and parietoâoccipitoâtemporal regions, while primary visual cortex can favor visionâonly features.
Ablations and scaling. Removing multiâsubject training or the transformer hurts performance; more sessions keep improving results, and longer LM context windows up to 1,024 words steadily boost textâdriven encoding, supporting the role of highâlevel semantics.
5. OdysseyBench
A new benchmark and dataâgeneration pipeline to test agents on realistic, multiâday office tasks across Word, Excel, PDF, Email, and Calendar. It introduces OdysseyBench (two splits) and HOMERAGENTS (an automated multiâagent generator), with evaluations showing large gaps between human and model performance and clear benefits from semantically compressed memory.
Whatâs new: OdysseyBench targets longâhorizon, contextâdependent workflows instead of atomic tasks. Two splits: OdysseyBench+ (300 tasks distilled from real OfficeBench cases) and OdysseyBenchâNeo (302 newly synthesized, more complex tasks). Tasks require retrieving key facts from multiâday dialogues and coordinating actions across apps.
How itâs built: HOMERAGENTS has two paths. HOMERAGENTS+ iteratively turns atomic OfficeBench items into rich multiâday dialogues via a generatorâverifier loop. HOMERAGENTSâNEO plans, explores an app environment, generates tasks (intent, subtasks, eval criteria), and then synthesizes 5âday dialogues. All agents use GPTâ4.1; at least five calendar days of dialogue are produced per task.
Data & evaluation: 602 total tasks: 153 singleâapp, 166 twoâapp, 283 threeâapp. Neo conversations are longer and denser (â49% more tokens) than Plus. Execution steps cluster around 3â15. Automated checks (exact/fuzzy/executionâbased) compute pass rate after running agents inside a Dockerized office stack; LLMâjudge and human curation raise data quality.
Main results: Performance drops as apps increase; even top models struggle on 3âapp tasks. Example: on OdysseyBench+, o3 goes 72.83%â30.36% from 1âapp to 3âapp; GPTâ4.1 goes 55.91%â12.50%. Humans exceed 90% across settings. RAG with semantic summaries beats raw retrieval at far lower token budgets; chunkâlevel summaries reach â56% on Neo vs. 52% longâcontext with ~20% tokens. Execution steps remain similar or shrink with summarized memory.
Where agents fail: Typical errors include missing referenced files, skipping required actions, wrong tool choice (e.g., trying to âcreate PDFâ directly instead of writing in Word then converting), and poor planning order. File creation/editing in docx/xlsx is particularly errorâprone. The authors argue that semantic compression and coherent aggregation are essential for multiâstep reasoning in long contexts.
6. Beyond Ten Turns
This paper introduces ASearcher, an open-source framework for training LLM-based search agents capable of long-horizon, expert-level search. It addresses two major limitations in prior open-source approaches: short turn limits (â€10) and lack of large-scale, high-quality QA data.
Key points:
Fully asynchronous RL for long-horizon search â Unlike batch generation RL, ASearcher decouples trajectory execution from model updates, avoiding bottlenecks from long trajectories. This enables relaxed turn limits (up to 128), with training showing >40 tool calls and >150k tokens in a single trajectory.
Scalable QA synthesis agent â An autonomous LLM agent generates complex, uncertainty-rich QA pairs by injection (adding external facts) and fuzzing (obscuring key info), followed by multi-stage quality checks. From 14k seeds, 134k high-quality QAs were created, 25.6k requiring tool use.
Simple but powerful agent design â Uses only search and browsing tools (no external LLM), with end-to-end RL optimizing reasoning and summarization. Tailored prompting and history management are applied for both base LLMs (Qwen2.5-7B/14B) and large reasoning models (QwQ-32B).
Expert-level search behaviors â Through RL, ASearcher-Web-QwQ learns uncertainty-aware reasoning, precise key info extraction from noisy content, cross-document inference, and rigorous verification, outperforming Search-R1-32B and Search-o1(QwQ) in case studies.
State-of-the-art performance â Achieves Avg@4 of 42.1 on xBench-DeepSearch and 52.8 on GAIA, with significant RL gains (+46.7% and +20.8% respectively). Local KB-trained agents generalize well to web search, surpassing stronger baselines.
Training efficiency â Asynchronous rollouts and decoupled updates maintain high GPU utilization, handling large variance in tool call count and token length per trajectory.
7. Illusion of Progress
The paper argues that common QA hallucination detectors look better than they are because evaluations lean on ROUGE. In humanâaligned tests, many detectors drop sharply. Simple responseâlength heuristics rival sophisticated methods, revealing a core evaluation flaw.
ROUGE misaligns with humans. In a human study, LLMâasâJudge matches human labels much better than ROUGE.
Reâscoring detectors collapses headline results. When replacing ROUGE with LLMâasâJudge, AUROC drops are large: up to â45.9% for Perplexity and â30.4% for Eigenscore on NQâOpen with Mistral; PRâAUC gaps are even larger. Correlation between ROUGEâ and LLMâbased AUROC is only r = 0.55.
Length is the hidden confounder. Hallucinated answers are typically longer with higher variance. Many detectors are strongly correlated with length, not semantics. ROUGE systematically penalizes long responses and can be gamed by repetition without changing facts.
Simple baselines rival complex methods. Length features like mean and std across samples achieve competitive AUROC, sometimes matching or beating Eigenscore and LNâEntropy.
Fewâshot helps format, not truth. Fewâshot examples reduce some ROUGE vs LLMâasâJudge discrepancies and stabilize outputs, but method rankings still shift and model effects persist; Semantic Entropy is relatively more stable.
8. GLM-4.5
An open MixtureâofâExperts family that targets a single model excelling across agentic tool use, complex reasoning, and realâworld coding. GLMâ4.5 (355B total, 32B active) introduces hybrid âthinking vs directâ modes, multiâstage pretrain + midâtrain to 128K context, and extensive RL for reasoning, agents, and instruction following. It ranks near the top on a 12âbench ARC suite and releases weights and eval tooling.
Results at a glance â On the 12âbenchmark ARC suite, GLMâ4.5 averages 3rd overall and 2nd on agentic tasks; Key scores: TAUâBench 70.1, BFCLâV3 77.8, BrowseComp 26.4; AIME24 91.0, GPQA 79.1; SWEâbench Verified 64.2, TerminalâBench 37.5.
Architecture and scaling choices â MoE with lossâfree balance routing and sigmoid gates, GQA with partial RoPE, QKâNorm, 96 attention heads at 5,120 hidden, and an MoE MultiâToken Prediction layer for speculative decoding.
Training recipe â 23T tokens pretrain with qualityâbucketed web, code, math, science, and multilingual data; midâtraining adds repoâlevel code sequences, synthetic reasoning traces, 128K context, and large synthetic agent trajectories. Optimizer is Muon with cosine decay and sequenceâlength extension plus RoPE base adjustment.
Postâtraining and RL â Twoâstage expertâthenâunified SFT + RL. Reasoning RL uses a difficulty curriculum, singleâstage 64K outputâlength RL, tokenâweighted loss for code, and strict filtering for science. Agentic RL covers web search and SWE with outcome rewards, strict format penalties, iterative selfâdistillation, and turnâscaling benefits. A new XMLâtagged functionâcall template reduces escaping overhead.
RL infrastructure for agents â Slime provides synchronous training for general RL and decoupled asynchronous rollouts for longâhorizon agent tasks, with FP8 inference for faster data generation.
9. A Survey on Efficient Architectures for LLMs
This survey reviews advances in efficient LLM architectures beyond traditional transformers, including linear and sparse sequence models, efficient attention variants, sparse MoEs, hybrid designs, and diffusion-based LLMs. It highlights cross-modal applications and outlines a blueprint for building scalable, resource-efficient foundation models.
10. A Deep Dive into RL for LLM Reasoning
This paper reviews and rigorously re-evaluates reinforcement learning techniques for LLM reasoning, addressing inconsistencies caused by varied setups and unclear guidelines. It offers a unified open-source framework, practical selection guidelines, and shows that a minimalist two-technique combo with vanilla PPO can outperform methods like GRPO and DAPO.