Large Language Models
1
Are Dense Rewards Worth Their Tokens in GRPO?
The paper tests whether Monte-Carlo dense process rewards in RLVR are worth their cost, since the scoring rollouts make up 91% of generated tokens. With a matched-compute protocol over 1,668 runs on six models, the ranking reverses: on Qwen2.5-1.5B the Monte-Carlo reward leads per update but ends 14.3 points below GRPO at equal tokens and 7.8 points below at equal training time. None of four dense-reward variants beats plain GRPO. Selecting prompts with predicted success near 1/2 adds 1.7 points (Qwen3-0.6B) and 0.8 points (Qwen2.5-1.5B), though plain GRPO with a doubled learning rate comes within half a point.
Why it mattersMany RL-for-reasoning papers report gains per training step, which hides the true cost of extra rollouts. This large preregistered study shows that compute-matched comparisons can erase or reverse such gains, and offers a cheap protocol other teams can adopt.
arXiv ↗PDF ↓Code ↗
2
The Surprising Effectiveness of Deleting Weights in LLM Post-Training
The paper asks how little of a pretrained LLM must change to gain new capabilities. Deletion Tuning learns a binary mask over the gate projections of transformer blocks, zeroing under 0.05% of weights and changing nothing else. It matches full fine-tuning at scale in RLVR, SFT and on-policy distillation, with masks hundreds to thousands of times smaller than a full checkpoint. Under RL, prior-task accuracy stays near the base model, while full fine-tuning, LoRA and random sparse fine-tuning lose up to 5.4 points.
Why it mattersIt suggests post-training mostly selects capabilities that already exist rather than adding new ones. Tiny masks could make storing and serving many specialized variants far cheaper while reducing forgetting.
arXiv ↗PDF ↓
3
Tree-Ising: Repairing Language Models with Bit-Level Ising Edits
Tree-Ising repairs reasoning loops in LLMs by treating the fix as a sparse discrete optimization over BF16 weight bits, allowing only sign flips and exponent shifts. The problem splits into independent per-row QUBO (Ising) problems solved on GPU by simulated annealing or bifurcation. On Qwen3.5-0.8B it reaches 54.6% on MATH-500, versus 51.7% for an edit written by a frontier LLM agent and 7.9% for the base model. It needs four forward-pass requests, zero LLM tokens and about $0.90 of GPU time; on Qwen3.5-2B, greedy GSM8K rises from 21.6% to 69.4%.
Why it mattersIt turns weight-space model repair from an expensive agentic loop into a cheap, principled optimization. If it generalizes, targeted bit-level edits could become a practical tool for fixing specific failure modes without retraining.
arXiv ↗PDF ↓
4
One Byte, One Token: Byteifying a 27B Hybrid-Attention LLM with Self-Generated Data
The authors convert Qwen3.8-27B (hybrid Gated DeltaNet plus full attention) into a pure byte-level model, shrinking the vocabulary from 248,320 to 384 entries and leaving other weights unchanged. A two-stage curriculum on 4.77 GB of mostly self-generated data took 178.6 hours on 8xH200. Quality holds on MATH-500, HumanEval+, MMLU-Pro and non-thinking GSM8K, with 4–6 point losses on C-Eval, thinking GSM8K and IFEval. Character-level skills improve (CUTE 75.2 to 84.0), but Chinese output speed falls to about a quarter and 64K-byte prefill is 6.5x slower.
Why it mattersIt is one of the largest demonstrations that a trained subword LLM can be retrofitted to bytes without a patching architecture. The detailed cost and failure-mode accounting is useful for anyone weighing tokenizer-free models for non-English languages.
arXiv ↗PDF ↓Eval suite ↗
AI Agents
5
ExpertCrowd-SWE: Measuring Frontier Coding Agents on Expert-Authored Engineering Tasks in Unfamiliar Codebases
ExpertCrowd-SWE is a benchmark of 100 multi-file tasks written by practising engineers across 87 open-source projects and 11 languages, each with a sealed reference solution and a partial-credit grader. The median reference change is 471 lines, and nearly half the tasks include lazy, alternative and adversarial control patches to validate the grader. Across 3,495 scored runs of seven frontier models, Claude Opus 5.5 (67.3%) and GPT-6 Astra (63.7%) lead with no significant gap; Claude Fable 5.1 scores 46.5% and Gemini 3.8 Flash 40.0%, with Qwen 3.8 Max, DeepSeek V4.1 Flash and Grok 4.6 at 32–34%. On 80 tasks no model earns full credit in all five runs; 24 tasks are public and 76 held out.
Why it mattersIt measures coding agents on realistic, merge-quality work in unfamiliar repositories rather than short bug fixes. The held-out split and validated graders make it a credible check on vendor claims.
arXiv ↗PDF ↓Dataset ↗
6
Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
The paper shows external memory can hurt agents: they fail tasks they would solve without memory, usually by following a wrong lesson or misapplying a valid one. The authors trace this to lessons being stored without their source trajectories. WikiTrace stores each lesson as a short insight linked to a summary of why its source run succeeded or failed and to the raw trajectory. When an insight is incomplete, conflicting or from a failed task, the agent follows the links with terminal tools to correct it before acting; across five agentic benchmarks it beats existing memory methods and reduces memory-induced degradation (no headline numbers in the abstract).
Why it mattersPersistent memory is becoming standard in agent products, so knowing when it backfires matters. Keeping provenance with each lesson is a simple design pattern other agent frameworks can adopt.
arXiv ↗PDF ↓Code ↗
7
MMDM: Measuring Massive Multitask Decision Making
MMDM benchmarks System-1 decision models that pick among query-specified options in one forward pass, the kind used for agent state tracking, tool and skill selection, preference judgments and rubric scoring. It has 5,448 questions in 16 task categories, option spaces from 2 to 1,024 and inputs up to 64,817 tokens. Across 28 models, many handle common knowledge but struggle with interdependent options, constrained decisions and multi-select. Perplexity Decider-V1.2 scores 78.75% on world knowledge but 53.95% on option logic and 47.95% on constraint planning, and the best hard multi-select accuracy is 36.81%.
Why it mattersFast routers and selectors sit inside most production agent stacks, yet rarely get dedicated evaluation. MMDM gives a fresh, contamination-aware test of exactly that component.
arXiv ↗PDF ↓Project ↗
Model Releases
8
Muse Ember: Meta's smallest, most efficient Muse model
Meta released Muse Ember, the smallest and cheapest Muse model, aimed at high-volume agent work such as coding subtasks, tool calls, routing and document analysis. It accepts text, image, audio and video and costs $0.10 input / $0.50 output per million tokens (up to 200K input tokens) via the OpenAI-compatible Meta Model API. In Meta's own numbers it ties Claude Haiku 5.5 on the Artificial Analysis Intelligence Index (36), scores 40.9 on Terminal-Bench 4.0 and 69.5 on OSWorld 2.1, but trails Haiku 5.5 on FrontierCode 2.0-AA (27.2 vs 34.9). Meta positions it as a worker under larger planners like Muse Wonder; the post does not mention open weights.
Why it mattersCheap executor models are what make planner-plus-worker agent architectures affordable at scale. Ember adds price pressure in the small-model tier, though its benchmark claims are self-reported.
Meta blog ↗
