RESEARCH

Research

173 paper takes since 2026-08-19

RESEARCHYuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality · arXiv · Sep 30

Composing an explicit, editable score before rendering audio demonstrably improves what expert listeners hear—and makes the composition itself available for revision.

YuE2 writes a readable score before generating audio, beating Suno v4.5 in expert listening and nearly matching Suno v5 on a best-of-8 basis.

Read by the desk

↑216hf upvotes

RESEARCHBeyond Teacher Assignment: Domain-Normalized Multi-Teacher On-Policy Distillation · arXiv · Sep 30

Instruction-following feedback dominates multi-teacher distillation gradients; rescaling by batch-wise spread recovers the mathematics gain that label routing loses.

DN-MOPD rescales each specialist's feedback by its measured spread, fixing the imbalance that lets instruction-following dominate multi-teacher distillation.

Read by the desk

↑140hf upvotes

RESEARCHSelf-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence · arXiv · Sep 30

Wrapping a VLA inside an executable world-and-policy program beats the same VLA alone by attributable, reproducible margins—the mechanism is the key result.

HexaAnything wraps VLA/WAM policies in code-represented task state so verified execution traces feed back as training data, beating direct VLA on unseen tasks.

Read by the desk

↑98hf upvotes

RESEARCHDuplex-MPE: Benchmarking Multi-Party Interaction in Full-Duplex Dialogue · arXiv · Sep 30

Four deliberately non-aggregable scores expose failure modes that a single response rate would bury in any multi-party spoken assistant evaluation.

Duplex-MPE is a benchmark of 2,000 scenarios testing whether a full-duplex speech assistant answers, stays silent, or stops in multi-party conversations.

Read by the desk

↑84hf upvotes

RESEARCHCoWindow Attention: Full Causal Coverage Is a Collective Property · arXiv · Sep 30

Distributing long-range attention across KV heads rather than duplicating it cuts 32K training FLOPs by 28.5% at 14B with no meaningful quality loss.

CoWA splits causal-history access across KV heads for full collective coverage, cutting 128K training latency 7-8x while tracking FullAttn accuracy at scale.

Read by the desk

↑60hf upvotes

RESEARCHHow Far Are We from Removing the Visual Encoder? Scaling Laws for Encoder-Free Multimodal Pretraining · arXiv · Sep 30

Encoder-free multimodal models are predicted to match encoder-based ones around 10²³ FLOPs—and the decoder's internal reorganization explains why.

Scaling-laws analysis predicts encoder-free MLLMs close the multimodal gap with encoder-based models at ~10^22 FLOPs, within practical pretraining budgets.

Read by the desk

↑57hf upvotes

the other track

182 more tool takes on the wire

Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.

feder-cr/dots JoasASantos/Offensive-Security-AI-Models CaptureGrubEnchant/SolidWorks

RESEARCHImproving Test-Time Scaling with Adaptive Looped Transformers · arXiv · Sep 30

TaH2 turns the looped transformer's steeper scaling slope into actual accuracy gains by learning per-token iteration depth jointly with the backbone, not as an afterthought.

TaH2 adds a learned iteration decider to looped transformers, lifting the accuracy-compute slope 53% over a non-looped baseline on AIME benchmarks.

Read by the desk

↑49hf upvotes

RESEARCHRecursive Harness Distillation across Agents for Robot Manipulation · arXiv · Sep 30

A playbook distilled from a strong agent's failures can make a cheap agent outperform the strong agent running blind — and the numbers hold on a physical robot.

A method that distills a strong agent's robot manipulation experience into a reusable playbook, improving success from 37.3% to 64.0% in real-world tasks.

Read by the desk

↑35hf upvotes

RESEARCHKnowing When Thinking Is Not Enough: Teaching Small Reasoning Models to Reason Beyond Their Parametric Knowledge · arXiv · Sep 30

Thinking longer only helps when the answer is already reachable; this paper proves the distinction empirically and trains small models to act on it.

FlyBy trains 4B/8B models to detect knowledge bottlenecks mid-reasoning and query stronger models, letting FlyBy-4B beat Qwen3-14B at 2.7x lower serving cost.

Read by the desk

↑34hf upvotes

RESEARCHFuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders · arXiv · Sep 29

Randomizing which encoder layers form the latent during training turns a fixed architectural choice into a regularizer that benefits both reconstruction and generation without extra cost.

FuseReg trains over random encoder-layer subsets so one decoder handles any fusion without retraining, cutting unguided gFID by up to 29%.

Read by the desk

↑118hf upvotes

RESEARCHDisaggregated Quantization: Specializing LLM Prefill and Decode · arXiv · Sep 29

Phase-aware quantization that treats prefill and decode as distinct hardware problems — and solves both without requiring a single unified compromise.

Disaggregated quantization uses separate formats for prefill and decode, yielding 32+ point accuracy gains and 1.78x faster first-token latency on 27B models.

Read by the desk

↑63hf upvotes

RESEARCHBlock Sparse Attention with Log-Linear Complexity · arXiv · Sep 29

Solves the overlooked quadratic bottleneck inside block-sparse attention — block selection itself — with a log-depth pyramid that keeps overall complexity at O(N log N).

PISA cuts block-sparse attention's quadratic block-selection bottleneck to O(N log N) via a coarse-to-fine pyramid Top-K strategy with fused Triton kernels.

Read by the desk

↑21hf upvotes

the other track

182 more tool takes on the wire

Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.

openai/mcp-extensions rehan-remade/universal-modder Barty-Bart/motion-graphics

RESEARCHInternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data · arXiv · Sep 29

A disciplined architecture that turns future-video supervision into inference-free predictive representations, backed by the largest open robot training corpus assembled to date.

InternW0-Δ unifies visual dynamics, 4D geometry, and action generation in a Mixture-of-Transformers WAM trained on 20K+ hours of open robot and egocentric data.

Read by the desk

↑18hf upvotes

RESEARCHTimeEvo: Failure-Driven Self-Evolution of a Time Series Agent · arXiv · Sep 29

Failure-driven tool synthesis with a paired admission gate that counts harm directly beats every human-curated and self-refinement baseline across all thirty task-backbone combinations.

TimeEvo grows a time series agent's tool library from scratch—clustering diagnosed failures into capability gaps—and improves accuracy on every task tested.

Read by the desk

↑16hf upvotes

RESEARCHJev in the Wild: A Data-Driven Analysis of the Jev Model's Functionality, Applications and Ecosystem · arXiv · Sep 29

A week-one ecosystem census showing Jev is already a multi-purpose judgment layer, not just a router, with attention and adoption pointing in opposite directions.

A data-driven analysis of 2,170 GitHub projects examining how the Jev decision model is used for judgments, scoring, and routing across application domains.

Read by the desk

↑9hf upvotes

RESEARCHEnhancing Photogrammetric Digital Surface Models with Pretrained Diffusion Models and Multimodal Conditioning · arXiv · Sep 29

A disciplined adaptation of Stable Diffusion 3 to elevation refinement that halves Dense Urban DSM error in held-out tests — with honest accounting of where it fails.

Modified Stable Diffusion 3 cuts satellite DSM elevation error nearly in half (6.00→3.45 m RMSE) by conditioning on both DSM and optical imagery.

Read by the desk

↑9hf upvotes

RESEARCHSLCA-GRPO: Resolving Cross-Segment Credit Misattribution in Tool-Calling RL · arXiv · Sep 29

Routing tool and summary advantages to separate token segments before the backward pass measurably fixes a real failure mode in tool-calling RL, with honest bounds on what it cannot fix.

SLCA-GRPO routes execution advantages to tool tokens and preference advantages to summary tokens to fix cross-segment credit misattribution in tool-calling RL.

Read by the desk

↑7hf upvotes

RESEARCHTraining Object Permanence in World Models · arXiv · Sep 26

A cognitively grounded benchmark that exposes two distinct failure modes in video models and shows fine-tuning on synthetic physics data meaningfully closes one of them.

WROP pairs a 1.5M-sample object-permanence corpus with a 300-question blind Elo exam; their 16B PWM-WROP ranks first among continuation video world models.

Read by the desk

↑193hf upvotes

the other track

182 more tool takes on the wire

Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.

moguzbulbul/blueprint-animation deepseek-ai/DeepGEMM-Ascend emir/claude-s40

RESEARCHAgent-Editing World Model: Rethinking World Modeling for LLM Agents · arXiv · Sep 26

Editing the agent's reasoning before execution—not simulating tool responses—is what actually breaks the compounding-error cycle in long-horizon tasks.

AEWM edits noisy reasoning and actions in task history rather than predicting tool responses, improving agent scores by 3.2–6.7 points across six benchmarks.

Read by the desk

↑16hf upvotes

RESEARCHParts-of-Speech as Emergent Categories in SAE Latent Space · arXiv · Sep 26

PoS categories live in compact latent groups, not single features — and the group size tracks how lexically variable the category is.

A study finding that part-of-speech distinctions are recoverable from SAE activations via compact latent groups, not one-to-one latent-to-category mappings.

Read by the desk

↑10hf upvotes

RESEARCHQwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents · arXiv · Sep 26

A rigorous account of building mobile agents where the harness and the model improve each other across iterations, with honest cost and memory numbers to back it up.

Qwen-Planner-Agent uses an agentic data flywheel and CARE reward shaping to build mobile planning agents in a closed AI-for-AI loop.

Read by the desk

↑10hf upvotes

RESEARCHIterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis · arXiv · Sep 26

Role-decoupled prompts and per-role advantage normalization let a single 8B model beat most 30B deep-search agents by fixing context collapse at the architectural level.

IterSynth splits deep search into a Planner and Synthesizer to reduce role coupling and context noise, training both with role-specific advantages via RDPO.

Read by the desk

↑9hf upvotes

RESEARCHExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds · arXiv · Sep 26

Exploration works, but it's brittle: the same model under the same budget can end 72 points apart, and stating a rule correctly still leaves a 30-point execution gap.

A benchmark that tests AI exploration in rule-based "Alien Worlds" whose logic conflicts with training data, so recall alone cannot solve the tasks.

Read by the desk

↑8hf upvotes

RESEARCHCoding Agents for Generalized Task and Motion Planning Problems · arXiv · Sep 26

Freeze-at-deployment coding agents beat hand-engineered TAMP planners without receiving any of the symbolic scaffolding those planners depend on.

Coding agents hit 56–95% success vs 47% for hand-engineered planners on generalized TAMP, using an order of magnitude less compute per instance.

Read by the desk

↑8hf upvotes