The abstract reports that the tax is prevalent across 14 base/post-trained pairs from four model families and three agentic benchmarks (42 cases). It says the tax can be estimated from a few rollouts. It also proposes PTGS, a per-prompt temperature sampler used during RL training. The abstract says PTGS pays a smaller tax than a fixed-temperature baseline and also improves pass@1, in two agentic environments.

The stated mechanism is that post-training pushes tasks toward two extremes, always solved or never solved. That improves sampling efficiency and consistency but reduces coverage. The authors call the loss in test-time scalability the 'Sharpening Tax'.

The abstract says that base LLMs with a light inference harness have far lower pass@1 on agentic tasks. Given enough test-time samples (large K), they often beat their post-trained counterparts on pass@K. This is the authors' own claim. I did not check the experiments.

The abstract says LoopCD is training-free. One variant, LoopCD-Logits, adds one extra output pass. The other, LoopCD-Hidden, adds no output overhead. Both use an earlier recurrent pass as the 'weak' prediction and the final pass as the 'strong' one.

The abstract also claims that halving the recurrent loops with LoopCD still matches or beats full-depth unguided baselines. It puts the forward-FLOPs saving at 22.5% to 48.2%.

The abstract reports these gains across four looped model families. Ouro-2.6B-Thinking on AIME 2024 pass@1 rises from 61.88% to 73.33%. Huginn on HumanEval pass@1 rises from 22.56% to 31.71%.

Reported mitigation comparison, with Qwen3.5-4B and 9B. Multi-sample verification (MSV) cuts false-agreement mass only from 6.1% to 5.7% (4B) and from 8.8% to 7.2% (9B), and costs 6 extra labeler generations per candidate. CrossFit cuts it to 3.0% and 3.7%.

Reported downstream result: average gains of +8.8 (4B) and +8.4 (9B) points over standard coupled self-evolution, and +8.7 and +7.8 over Search-R1, across seven search benchmarks. These are the authors' claims, and I did not verify them.

Mechanism as described: source documents are split into folds A and B. Questions from A are scored by an auxiliary solver trained only on B, and the reverse. A same-source pseudo-label therefore cannot be echoed back as reward. Replaying identical proposals with source-excluded feedback gives 0.4% and 0.1% false agreement. The authors use this to separate the effect of feedback ancestry from curriculum changes.

Claim: in self-evolving search agents, the proposer and solver converge on shared errors ('co-cheating'). A post-hoc audit against source evidence reportedly shows this worsening over successive rounds. Pseudo-label correctness stagnates or declines while the in-loop reward improves.

The proposed E-MoE treats the expert-routing decisions of an MoE backbone as a shared discrete latent, giving a mixture of factorized distributions. The authors say it adds no active parameters over the factorized baseline. They report improved few-step generation on synthetic multi-modal benchmarks, binarized MNIST and LM1B. This is an author claim. The abstract page gives no numbers.

The abstract says masked diffusion models normally factorize the reverse process over positions, which limits sample quality in the few-step regime. It says earlier continuous-latent VAE approaches are prone to posterior collapse, where the latent is ignored.

The abstract names two VLM-based metrics, Logic-Render Alignment and Interaction Success Rate. It also covers entity control and long-horizon memory. The submitter's comment lists 12 metrics in total and says 'current video and world models' are evaluated across urban, indoor, nature, aquatic, aerial, racing and fantasy worlds. The page shows no scores, so I cannot say which models do well or badly.

The abstract says PROWBench has 170 programmatically built episodes and 600 proxy videos. Scene and event logs from a program run are replayed as 'world records'. Generated videos are then checked against what the program actually did, including events outside the camera's view. This is a described design, not a demonstrated result.

Reported: ablations suggest the gains depend on three parts. These are dynamically constructing targets, estimating their evolving utility, and balancing continued optimization with new failure discovery. A community comment links code on the microsoft/AutoSaddler GitHub repo (feat/activesaddler branch). I did not inspect the code.

Reported result: on GAIA2 and Terminal-Bench 2.0, test Pass@1 improves by 4.4 and 7.5 percentage points over the same harness optimizer with a scenario order fixed before optimization. The comparison is against a fixed-order variant of the same optimizer, not necessarily the strongest external baselines. The abstract gives no absolute scores, seeds or variance.

Claim: ActiveSaddler treats the curriculum as a non-stationary bandit. Recurring failures are abstracted into reusable 'failure-pattern' arms. It estimates the learning progress each arm could yield. It balances revisiting known weaknesses against exploring unseen scenarios.

Per the abstract, the authors define Rhetorical Robustness as two joint requirements: stability across content-preserving rewrites, and discrimination between different papers.

The authors propose SciCore, which averages a full-manuscript judgment with a judgment on an extracted structured 'science core'. The abstract claims a leading joint stability-discrimination profile in the primary GPT-5.5 comparison, with competitive human alignment. This is a claim from the abstract. I did not verify it, and 'leading' is limited to the benchmarked reviewers in that one comparison.

Per the abstract, RobustReview is a benchmark of 1,260 manuscript versions on which 30 reviewer configurations were evaluated. The authors report 'false robustness' (low rewrite sensitivity together with score collapse across papers). They also report that human alignment and rhetorical robustness rank reviewers differently, and that content-focused prompting does not consistently help across backbones.

Claimed method, part 1: cross-modal influence-guided routing. It uses bidirectional cross-attention responses as a cheap proxy for cross-modal influence. That proxy reweights token-aware losses and scales gradients across cross-modal layers. It is applied during forward-process RL (DiffusionNFT).

The abstract claims consistent gains over strong RL baselines in modality quality, semantic consistency and synchronization, with supporting ablations. It gives no numbers. The page lists no linked models, datasets or Spaces, so I found no public artifacts to check the claim against.

Claimed method, part 2: preference-preserving modality-aware reweighting. Predefined reward weights stay as priors. After warm-up, branch-specific reward-gradient interactions add residual corrections, so dominant rewards do not suppress weak but essential objectives.

The problem, as the abstract states it: when RL optimizes several rewards together (modality quality, cross-modal semantic alignment, audio-video synchronization), fixed update routing and fixed reward weights fail to track how the model changes during training.

The abstract says APPL treats each skill policy's 'structural prior' as an interface between skill learning and skill composition. An example prior is that a grasp depends only on the gripper pose relative to the object. The prior is built into training and is also stated in language so a planner knows where the policy applies. This is the authors' description of their method.

As described, a construction agent segments demonstrations into skills, proposes several priors per skill, and trains and verifies one policy per prior. A runtime agent then picks among these policies and composes them toward new goals. The authors report better out-of-distribution skill generalization and previously unseen compositions on MetaWorld and long-horizon ManiSkill tasks. They also report that removing the interface information substantially reduces performance. These are the paper's own claims. I saw no numbers, baselines or seed counts.

Claims in the abstract: the 4B model is best among the compared methods on all eight streaming benchmarks. Ablations show that keeping generated captions improves historical QA without hurting real-time perception. PSTL beats dense state supervision while supervising only 27.5% of annotated state tokens. No absolute scores appear on this page, and I did not check which baselines were compared.

The paper's stated method has two parts. Proactive Hierarchical Caption Memory (PHCM) writes time-grounded local captions and summaries of completed events. A recent visual window is kept alongside them, so historical visual features are not revisited. Proactive State Transition Learning (PSTL) supervises state-change and state-persistence tokens, which reduces the dominance of repeated 'waiting' states.

Released artifacts listed on the page: the OneStreamer-4B model, the OneStreamer-1M dataset (over one million records, mixing synthesized captions and QA with cleaned open data), a demo Space, and code. This makes outside replication possible, but I did not test any of it.

The authors say they introduce world-space metrics and a benchmark covering real and synthetic scenes. They claim 'substantially improved' out-of-view dynamics while staying 'competitive' in visual fidelity, camera control and 3D adherence. The abstract page gives no numbers, baselines or variance, so the size of the gain is unverified. The claim comes from the authors' own benchmark.

The proposed method, World Observer, jointly generates a perspective actor view and one or more panoramic observers that watch chosen world regions. Both are grounded by warping from a shared panoramic source. An 'Observer Sink' of high-resolution perspective references restores fine appearance on re-entry. The authors say observers can be placed freely, extended to multiple locations, and driven by control signals.

The abstract says video world models are actor-centric. Once an object leaves the actor's view, the model loses evidence of how it evolves and often fails to preserve its state and dynamics when the object re-enters.

The abstract says discrete diffusion LMs sample each token independently from its marginal when decoding in parallel, which loses dependencies between tokens decoded together. It says continuous diffusion LMs avoid this, but their denoiser sees only the continuous state, which has no tie to a valid token configuration until the final decode.

The abstract reports gains over discrete and continuous diffusion baselines at matched model size. The gains are in puzzle accuracy on Sudoku and Countdown and in generative perplexity on LM1B. The abstract gives no numbers. The page lists no linked models, datasets or Spaces, though it shows a project page and GitHub link.

The proposed method, HC-DLM, makes a continuous latent the only persistent generative state. Tokens are read out from it at every step and fed back as a scaffold for the next latent update. The authors say the training objective is derived from a variational bound on token likelihood.

The authors' summary (a submitter comment on the page) reports that PoS scored highest in all 12 benchmark-backbone combinations (4 benchmarks x 3 LLMs). It reports relative gains over the strongest same-backbone baseline of up to 22.68% on ALFWorld and 37.89% on RCA-100 joint accuracy. The page also lists ablations on consistency validation and recovery, and context-scaling experiments. These are self-reported results from the abstract page only.

Limitations of this read: I did not check the number of runs, variance, absolute scores, baseline strength or compute overhead. Gains are described as relative, and 'highest overall' does not show statistical significance. The page does not link any model or dataset.

The abstract says PoS maintains an explicit belief state. The state combines an estimate of the current world with unresolved task requirements. It checks belief updates for consistency, and it monitors progress to detect 'Belief Trapping', where an agent keeps acting without advancing the goal. Recovery is tailored to the trapping pattern and the blocked requirement.

Reported zero-shot results from the introduction, on Qwen3.6-27B and GPT5.6-Sol: 11.4% higher accuracy with 21.5% fewer prefix-reuse FLOPs than the strongest baseline on BrowseComp-Plus. TerminalBench 2.1 accuracy matches the strongest baseline with 29.5% fewer FLOPs. On a 10-task subset of 12-hour EdgeBench, scores are 5% higher with 59% fewer FLOPs. These are the authors' numbers. The compute metric is the authors' own 'prefix-reuse FLOPs'.

Serving cost and mitigation: arbitrary mid-context edits break prefix caching and force re-prefilling. The authors propose Suffix Cache Reuse, which they report cuts server-side compute by 35% versus standard SGLang at matched performance. I only read the claim in the introduction and did not read the appendix details.

Training result: stepwise GRPO with a success-gated efficiency advantage lifted Qwen3.5-9B on BrowseComp-Plus from 28.8% to 42.5%. It beat a Codex-style summary harness trained with the same recipe by only 0.4 points, though with 38.8% fewer FLOPs. The accuracy edge over that baseline is small, and the main advantage is efficiency. No run counts or variance were visible in the text I read.

Method (stated by the authors): a standard LM appends tokens to its context. A CLM instead treats the context as an editable file, so c_{t+1} = f(c_t) can be an arbitrary edit. Multiple agents' contexts can coexist as files. This is a design claim. I read it in the introduction.

RRSI constrains both sides of the loop. The proposer has an annealed limit on how many edits one candidate can bundle, and it is pushed toward unexplored edits. The selector has a critic that screens benchmark-specific proposals, plus a pruner that removes changes that are too small, too costly or no longer useful. The authors compare these to L0, L2 and L1 regularization, and state the analogy is qualitative.

Setup details from the HTML: the policy is frozen (the paper names Claude Opus 4.8). Baselines are the unevolved harness plus four evolution methods (Meta-Harness, AHE, TTHE, HarnessX), all sharing the same starting harness, evolve set and candidate budget. Each domain is evolved on one suite and then run unchanged on held-out benchmarks. Several verifiers are LLM judges.

Reported results across 8 benchmarks in 3 domains (coding, agentic workspace, engineering design): up to +14.1 points on the evolve split, up to +4.7 points on the five out-of-distribution benchmarks, and a harness using 30% fewer policy tokens than unregularized evolution. The authors say all six held-out splits improved. These are author-reported figures that I could not verify against tables.

The paper says that when an LLM agent's harness (prompts, control flow, tools, memory) is evolved automatically against a finite evolve set, it can overfit. In-distribution gains can shrink or vanish on out-of-distribution benchmarks. This is the authors' framing, and they cite other work reporting evolve-to-held-out gaps.

The authors report that rollout policy does not necessarily play a central role. KL direction shapes task performance and output coverage more clearly, and learning rate governs forgetting and update sparsity. Forward KL is reported as robust to rollout policy. Reverse KL is more sensitive and favours student-generated rollouts.

On-policy data improved generalisation to harder Countdown variants under both KL directions. The abstract says this advantage does not reliably persist after later RLVR. The authors also say their conclusions hold without gradient clipping, with sampled KL estimators, and on longer reasoning chains.

Per the abstract, the authors vary rollout policy, token-level KL direction and learning rate independently. They test Llama3 and Qwen2.5 models on scientific, medical and arithmetic reasoning tasks. This is a controlled design, which the abstract says earlier SFT-vs-RL comparisons lacked.

The proposed mechanism is that post-training pushes tasks to two extremes, always solved or never solved. That improves sampling efficiency and consistency but costs coverage. The authors introduce a diagnostic metric called 'Sharpening Tax'. They say it appears in most of 42 cases (14 base/post-trained pairs, 4 model families, 3 agentic benchmarks) and can be estimated from a few rollouts.