Search posts, papers, and topics
This empirical study dissects coding agent harness components like planning, action space, and context management. It finds context management is critical for tight windows, rule-based elision is efficient, and planning shifts from accuracy to cost savings for stronger models.
The article content is unavailable, providing no details about HarnessTax or its findings. Therefore we cannot extract any conclusions or takeaways.
BITCOS is a distribution‑aware storage layout for ternary LLM weights that replaces the standard five‑trit packing. By storing a presence bitmap and a compact sign vector, it reduces the effective bits‑per‑weight to 2 − z (z = zero density), achieving as low as 1.485 b/w on sparse models. The authors provide AVX‑512, AVX2, and Xe2 GPU unpacking kernels and show up to 1.28× speedup in matrix‑vecto…
The paper studies long‑horizon memorization where a language model must learn 100 tasks via continual fine‑tuning without retaining data or task IDs. By composing data, function, and weight anchors with merged LoRA, they boost final retention from 1.2% to 34.9%, a 28× improvement.
JEPA-Anything extends joint‑embedding predictive architectures with orthogonal predictive factorization, letting a single model learn complementary latent factors that can be recombined for prediction across disparate domains. The paper shows consistent performance gains on ten dynamics tasks, molecular simulations, and clinical event forecasting, plus experimental validation of a biologically‑de…
Alibaba announced the release of Qwen 3.8 Omni Flash, a new model in their Qwen series. The article provided no specific technical details, features, or performance metrics regarding this release.
Claude merges its Cowork and chat interfaces into a single assistant and adds beta Docs, Slides, and Design capabilities, rolling out first to Pro and Max plans. The unified UI lets users ask Claude to generate reports, presentations, and other artifacts within the same conversation, with optional auto‑action settings. No new technical details or implementation specifics are provided.
Alibaba's Damo Academy open-sourced Damo Radar, an AI model that detects nearly 150 abdominal conditions, including cancers, from CT scans. It achieved an average AUC of 0.913 across 146 clinical findings and outperformed most human radiologists in a comparative study.
IntBMoE introduces block‑level conditioning to MoE, decoupling token participation, compute execution, and memory materialization. A hypernetwork merges all experts into a composed expert per block, while routing remains sparse. Dual‑Path Residual Gating further mixes two composed paths. Experiments show consistent gains on vision, language, and recommendation tasks, and the model is live in AMap…
Numberwang is a tiny character‑level CNN packaged as a 1.8 MB JSON model and a ~100‑line pure‑Python inference script that classifies numeric strings in eleven languages without any external dependencies. It reaches 88.9 % macro‑F1 on a held‑out set, but struggles with unseen arithmetic expressions.
The authors audit six popular physics benchmarks by having domain experts re‑grade model outputs, fixing reference answers and removing ambiguous items. After correction, GPT‑5.6‑Sol’s mean@4 jumps from ~47 % to ~79 % on HLE‑Physics and from ~61 % to ~87 % on CMT‑Benchmark, with a corrected pass@4 of 94 % on 54 vetted CritPt challenges. The work shows current benchmarks severely under‑report LLM…
StepAudio 3 Gen is a general‑purpose audio generation model that replaces diffusion with a discrete autoregressive generator over residual vector quantization tokens. Using a 16‑layer RVQ tokenizer and progressive pretraining, it reaches state‑of‑the‑art zero‑shot TTS and voice‑design performance while handling speech, vocals, sound effects, and music.
PACT is a new benchmark designed to measure how well enterprise LLM agents follow compliance rules, especially when under user pressure. It found substantial variability across 22 models, with even the strongest assistants misapplying rules 6-10% of the time, and user pressure increasing violation rates by 65% on average.
The author argues that mathematics should reward “motivated explanations” – narrative, intuition‑driven expositions that clarify why a theorem is interesting and how it fits into broader context – on par with traditional proof‑oriented work. He defines the concept, contrasts it with proofs, cites examples (Princeton Companion, Thurston’s essays, Chow’s exposition paper), and suggests formalizing…
ActObs is a supervised fine-tuning method that trains agents to predict both actions and environmental observations from trajectories. This joint supervision improves subsequent reinforcement learning performance, leading to better exploration and task completion on benchmarks like Terminal-Bench 2.0 and aider-polyglot.
The paper presents Infinite-Parameter LLMs, where a compact hypernetwork creates feed‑forward weights from live user data and updates a Bayesian latent code online, keeping the stored model size constant while effectively having infinite parameters. This design aims to improve over standard in‑context learning and retrieval by persisting knowledge in weights and freeing context space.
PPO critics in reinforcement learning for LLMs suffer from "Value Flattening," where predicted state values are too flat compared to actual values. This paper identifies the causes as an implicit variance penalty and redundant updates, and proposes SParse Proximal Policy Optimization (SP3O) to mitigate it by supervising only a few well-separated states.
This paper introduces a novel evaluation framework to assess the physical world reasoning capabilities of omni-modal generative models like MiniMax-H3. It found that MiniMax-H3 achieved an overall success rate of 41.97% across 517 instances, with significant performance variations depending on the input modalities and reasoning tasks.
The article discusses techniques for exfiltrating machine learning model weights. It highlights the security implications for ML deployments.
Agora treats a Git repository as a shared, append‑only memory for autonomous research agents, recording each claim as an immutable commit in a DAG. In a 12‑day run with 13 language‑model workers it generated 1,703 reproducible contributions and closed 62 % of the gap to a trained GPT‑2‑124M, demonstrating that shared research state can accelerate discovery.
On-policy distillation (OPD) can lead to excessively long student responses, a phenomenon called length inflation. This paper identifies "termination-token mismatch" between base students and post-trained teachers as a key source, where models place stopping probability on different EOS tokens. Treating functionally equivalent EOS tokens as a shared semantic stopping action substantially mitigate…
Feyospace‑v1 presents a data‑centric training pipeline for cyber‑security agents, combining five systems (Choulea, SkyReal, Hongzwang, PSBreakup, Kreator) to generate and verify 164 k long‑context trajectories across diverse exploit environments. The resulting checkpoints improve baseline performance by ~24% on CyberGym and achieve a 63% verified success rate, ranking top among similarly‑sized op…
The paper introduces Designer‑RSI, a continual‑adaptation system that couples a frozen design‑software‑controlling LLM with an external procedural memory of natural‑language design skills. Over five adaptation rounds on real user briefs, the memory grows from 76 to 139 procedures and lifts execution success from 72.7% to 99.3%, showing that skill accumulation and selective replay can dramatically…
Action tokenizers for autoregressive VLA models often fail to preserve subtle action adjustments, despite good pointwise reconstruction. This paper introduces Physical Rank Consistency (PRC) to measure relational fidelity and ActionPiece, a new tokenizer that uses joint supervision to preserve these physical relationships, significantly improving policy success on robotics benchmarks.
Video DeltaNet (VDN) replaces full‑softmax attention in video diffusion models with a hybrid: per‑frame local Softmax for fine detail and a bidirectional linear memory (Video Delta Attention) for long‑range context. A teacher‑alignment schedule injects the linear branch into a pretrained MiniMax H3 model, preserving Softmax for text/audio streams. On eight NVIDIA B200 GPUs VDN‑H3 denoises a 14.3‑…
Fathom introduces a per-query read depth mechanism for sparse decoding over offloaded KV caches, allowing each query to adaptively decide how many bits of each key channel to read. This method significantly speeds up decoding for large language models with long contexts by reducing host memory traffic, achieving 1.67x faster GPU decoding on Qwen3-8B at one million tokens.
FAMOS is a feed-forward model that predicts 3D articulation, including movable-part segmentation and joint parameters, from sparse, unordered point clouds. It leverages multiple observations through a Multi-state Articulation Transformer and a novel observed articulation span objective, outperforming existing baselines.
The paper defines recursive self‑improvement (RSI) for AI, introduces the Headroom‑Closed Index to expose current LLM limits, and proposes a staged roadmap toward full RSI. It surveys application scenarios and outlines practical challenges.
XConf (eXperiential Confidence) is a novel method for estimating language model confidence by leveraging the model's accumulated experience from past graded episodes. It significantly outperforms existing methods like self-consistency in discrimination and calibration, at a fraction of the computational cost, across various tasks.
When2Think introduces a post‑training framework that lets a large reasoning model decide per‑instance how much reasoning depth to allocate, using difficulty‑aware reward shaping (IDAC) and verifier rewards. It cuts token usage by ~28% while boosting Pass@3 by 10% on AIME24 and reaches 40% Pass@3 on AIME25, outperforming compression and routing baselines.
