User feedback helps models repair targeted faults, but LLM judges often miss those repairs. Agent evals should retain human feedback and avoid treating model preference as ground truth.FEED7 SUMMARY

Language-based monitoring cannot fully reveal an agent’s internal computation, the paper argues. Secure agent runtimes still need isolation and data-flow controls independent of self-reporting.FEED7 SUMMARY

A coding-specialized model paired post-training with an iterative generate-evaluate-refine loop to exceed the top IOI 2026 human score. The reusable idea is feedback-driven test-time search.FEED7 SUMMARY

A UE5M3 block-scaling recipe trained an 8B model in FP4 without Hadamard transforms or BF16 final layers, while reporting better losses and downstream estimates than the compared recipe.FEED7 SUMMARY

Cliff trains on the location of a rollout’s first reasoning error, rewarding the valid prefix and penalizing the suffix. It offers finer supervision without a specialized reward model.FEED7 SUMMARY

A telecom diagnosis pipeline structures telemetry before forcing a decision path and evidence-grounded explanation. The pattern is relevant to agents reasoning over heterogeneous operational data.FEED7 SUMMARY

A label-free Dutch-book test finds internally inconsistent probabilities from language models, especially when prompts add logical complexity or irrelevant context.FEED7 SUMMARY

SafeEvolve turns completed agent trajectories into reversible harness updates and policy training, linking runtime safeguards with learned multi-step behavior.FEED7 SUMMARY

Mechanistic tests suggest two open LLM judges route local errors early and form ratings late, giving eval builders concrete layers and components to probe for judge failures.FEED7 SUMMARY

CordisBench finds that models lose reliability as harness lifecycle interactions grow, while deterministic execution semantics can replace costly reasoning on controlled cases.FEED7 SUMMARY

This survey separates verbal feedback by when it acts and what it changes, helping agent builders distinguish task specification, test-time guidance, and parameter updates.FEED7 SUMMARY

StudentSim turns sparse user histories into individualized simulators that model both current behavior and response to guidance, a pattern for testing adaptive agents before live deployment.FEED7 SUMMARY

When quantizing an open model, spend a small extra precision budget across the network before protecting a few “important” layers; causal tests found the damage was usually diffuse.FEED7 SUMMARY

Small proxy models may be enough to choose an SFT-versus-RL annotation split: the paper finds broad near-optimal ranges that transfer to larger models.FEED7 SUMMARY

A production-derived post-training recipe consolidated more than 200 internal apps onto one self-hosted model by training separate experts for distinct quality gaps, then merging them.FEED7 SUMMARY

SAGE uses an expensive VLM only when a learner is uncertain, tests its advice against environment rewards, and distills useful guidance into a policy that needs no VLM at runtime.FEED7 SUMMARY

A two-stage ensemble reconciles outputs from string, graph-embedding, and RAG-based ontology aligners. Composition matters: mixed paradigms favor precision, while LLM-only groups more often favor F1.FEED7 SUMMARY

DIASENTINEL combines deterministic extraction, guideline retrieval, risk prediction, and hybrid verification on-premise. It is a useful architecture reference for auditable agents handling sensitive data.FEED7 SUMMARY

A controlled ontology-learning study finds model size is a weak selector on its own. Dense 27B models beat larger sparse models on one task, while MoE models led another.FEED7 SUMMARY

BLOOM-WILT makes rare model behaviors easier to surface by adapting both the auditor’s strategy and target decoding. Its results also show that elicitation method can reverse safety rankings.FEED7 SUMMARY

An industrial code-model case study frames post-training as maintaining data mixtures under fixed budgets. Raising usable supervision yield improved coding benchmarks without changing the teacher.FEED7 SUMMARY

S3Gym tests whether agents can explore, judge outcomes, and reuse experience. Raw history, compressed memory, and training each help in some tasks, but none is a reliable default.FEED7 SUMMARY

DARTS targets representation drift in merged decoder LLMs with entropy-weighted, position-aware correction, adding 0.1% parameters in the reported Llama-2-7B tests.FEED7 SUMMARY

CE-MoE concentrates routed experts in fewer layers and restores depth with dense layers, cutting reported GPU-hours by 33.3% at 31.5B parameters without worse benchmark results.FEED7 SUMMARY

A study of 8,351 Claude Code plugins finds instruction files and scripts often require coordinated maintenance, treating skills as evolving software rather than static prompts.FEED7 SUMMARY

CritICL turns recurring failures from smaller models into critique examples for stronger models, aiming to improve inference with fewer generations and lower token cost than repeated sampling.FEED7 SUMMARY

SWE-Prime finds that filtering coding-agent traces by process and segment quality can beat training on every resolved trajectory, reducing noisy imitation from redundant or risky steps.FEED7 SUMMARY

TTPO adapts a model at test time without labels by treating agreement and disagreement differently, reducing the damage from wrong majority-vote pseudo-labels.FEED7 SUMMARY

RedEvoAgent turns prior jailbreak trajectories into a compact attack skill, then keeps only validated improvements. It offers a more interpretable way to probe tool-using agents for unsafe actions.FEED7 SUMMARY

An unbiased sampler makes prefix probabilities tractable when a transducer maps many source strings to one target prefix. It matters mainly for specialized text or biological sequence systems.FEED7 SUMMARY

Persona-Execution Separation isolates mutable agent behavior from audited, stateful work through a governed bridge. It is an architecture option for multi-user agents under compliance constraints.FEED7 SUMMARY

Feeding a model partial reasoning from a weaker model during RLVR preserved exploration and improved large-k coverage. This suggests diversity can come from cross-model perturbation, not only regularization.FEED7 SUMMARY

Three ways to combine RLVR domain experts perform similarly on average but diverge by task. Choose Merge for cheap reuse, Mix RL for training from pooled data, and MOPD for preserving expert gains.FEED7 SUMMARY

TraceML exposes why coding agents stall on long ML work: they repeat narrow optimization loops instead of alternating work types, revisiting discarded ideas, and pivoting like experts.FEED7 SUMMARY

Prefix Sliding drops stale reasoning tokens while retaining instructions and recent work, capping memory during long inference and reportedly making existing models 3× faster without training.FEED7 SUMMARY

Correct answers can hide invalid computation. Trace Integrity proposes executable, replayable contracts for data agents and reports 45.8–59.1% correct-answer/invalid-trace rates across three SQL modes.FEED7 SUMMARY

Self-refinement pipelines need not use equally capable models: invest capacity in generation and revision, while a small critic may preserve gains at lower compute cost.FEED7 SUMMARY

A tool-exposed ontology steered models closer to human therapy patterns without fine-tuning, showing how explicit action vocabularies can improve agent behavior.FEED7 SUMMARY

LLM compliance generation behaves differently under vague and strict schemas: vague artifacts need richer context, while rigid formats can stay consistent yet hallucinate.FEED7 SUMMARY

Re³Cap uses multimodal retrieval to find caption omissions and hallucinations before refinement, offering a concrete retrieval-and-review pattern for vision agents.FEED7 SUMMARY

CLEAR conditionally activates a safety adapter instead of applying safety tuning to every prompt, reducing harmful completions while limiting benign-task degradation.FEED7 SUMMARY

ConceptTS turns LLM-proposed concepts into executable labels and interpretable forecasting bottlenecks, enabling concept-level inspection and intervention.FEED7 SUMMARY

ConceptGuard tests whether model unlearning blocks harmful uses of a concept while preserving benign ones. Current methods show weak contextual control and sharp forgetting-versus-utility trade-offs.FEED7 SUMMARY

AI4AI-Bench tests whether agents can rewrite training algorithms, not merely tune runs. The best system scored 0.250, with most submissions never changing how the model learns.FEED7 SUMMARY

MidTool trains general tool use before post-training, using API, MCP, document, web, PDF, and code data. Qwen3 4B and 8B variants improved across three downstream tool-use benchmarks.FEED7 SUMMARY

Per-problem self-improvement claims can arise from inference and evaluation noise. This audit argues every transition statistic needs a measured null from frozen baseline replicates.FEED7 SUMMARY

IAR turns a fixed document corpus into parametric knowledge through separate injection, QA alignment, and recovery stages. It improved domain QA while retaining more general capability than vanilla SFT.FEED7 SUMMARY

LFU was the strongest simple semantic-cache default, but answer validity mattered far more than eviction policy. Raw hit rates above 50% fell near 2% after quality adjustment.FEED7 SUMMARY

Reusable agent skills transferred better when induced from subtasks and stored as text. Whole-task skills often hurt versus no memory, so retrieval alone is not evidence that a memory helps.FEED7 SUMMARY

A model learned to choose its own reasoning budget, cutting average output tokens 41% with a small MATH500 accuracy drop. Agent builders could replace fixed effort settings with task-level routing.FEED7 SUMMARY

TokEval links tokenizer properties to language, math, and code performance, offering cheaper screening signals before committing compute to pretraining sweeps.FEED7 SUMMARY

Users were far more willing to deploy their own conversational agent than engage someone else’s, making receiver consent and receptivity-aware routing core product constraints.FEED7 SUMMARY

A single-GPU SBERT beat the reported zero-shot LLM and vendor baseline for invoice coding, suggesting narrow, private classifiers can outperform broader models with modest local data.FEED7 SUMMARY

BATON explores robot subtasks independently, stores their solutions, and manages handoffs between them, replacing multiplicative whole-task search with additive composition.FEED7 SUMMARY

A controlled study encoded authenticated internal-state evidence into unchanged answers, suggesting generated text could carry provenance signals, but not that current models reveal them naturally.FEED7 SUMMARY

Tested compliance guards often ignored the governing rule and classified from scenario cues. Builders should counterfactually swap policies before trusting a detector as an audit control.FEED7 SUMMARY

Agent handoffs should preserve task-relevant predictive state, not blindly compress the transcript. The paper offers a three-part record for carrying decisions, evidence, and irreducible observations across sessions.FEED7 SUMMARY

Multi-source agents should separate source interpretation from decision aggregation. A fixed evidence tuple plus calibrated likelihood pooling avoids vote thresholds that drift as more sources arrive.FEED7 SUMMARY

Rollplex overlaps VLM prefix work with rollout decoding and shares weights across training phases. On 32 H800 GPUs, it reports higher throughput without changing synchronous on-policy updates.FEED7 SUMMARY

Twin has a coding agent build and continuously repair an executable simulator before acting. The harness lifts the same base model from 7.8 to 93.3 on ARC-AGI-3 by validating every observed transition.FEED7 SUMMARY

YOPO combines reasoning steering and answer abstention in one pass of a frozen model. Reconstructing the clean residual avoids interference and beats the paper's two-pass reference at three model sizes.FEED7 SUMMARY

SheetCompass keeps spreadsheet structure as a hierarchical relation graph instead of flattening cells into text. Its agent memory retains task-relevant context across sheets, tables, and columns.FEED7 SUMMARY

QuoteBench shows that shell-command scores can conceal failures introduced by serialization and reparsing. Agent evals should identify the execution path, not attribute every result to the model.FEED7 SUMMARY

LittleLearner offers a controlled model and corpus for studying knowledge acquisition without unknown prior exposure. Its initial results separate better use of known material from new capability.FEED7 SUMMARY

SAEVerbalizer turns sparse-autoencoder decoder directions into natural-language feature explanations. It may reduce behavior-sampling work when inspecting model representations, but evidence is experimental.FEED7 SUMMARY

DARTree uses autoregressive correction across diffusion-generated draft trees to accelerate lossless decoding. It reports up to 9.73× local speedup, but production gains depend on the serving stack.FEED7 SUMMARY

Mimir v1 is an open 1B-parameter reasoning model trained with permissible post-training data. It is relevant for compact or Danish deployments, though the abstract supplies no benchmark scores.FEED7 SUMMARY

A task-agnostic influence measure tracks which examples steer pretraining toward final parameters without choosing a downstream eval. It reveals a literature-to-STEM shift across training.FEED7 SUMMARY

CreativeInstruct adds learned control spans that recover base-model-like diversity after post-training, with reported gains in human creativity ratings and downstream RL training.FEED7 SUMMARY

SkillProx evolves reusable agent instructions through measured retries, regression rollback, and utility-based pruning, reporting a 3-point accuracy gain over its strongest baseline.FEED7 SUMMARY

A taxonomy-based audit maps open-source LLM safety tools to enterprise risks, finding strong technical coverage but major governance, legal, regulatory, and financial gaps.FEED7 SUMMARY

Blast Radius predicts which code and context a prompt will reach, then reversibly archives unused history. Across seven OpenAI models, it cut token consumption by 17–26%.FEED7 SUMMARY

Controlled LLM rewriting made harder financial sentences cheaper to process with DisCoCat, cutting circuit size by over 70%, but downstream accuracy improved only modestly.FEED7 SUMMARY

PsychoAgent separates factual and affective memory, then reranks relevant memories by salience. It retrieved more conflict-critical context, but output-quality differences were not significant.FEED7 SUMMARY

P-Bench tests whether agents choose statistically valid methods, not merely execute code. Fisher-R1-14B improved single-trial success over DeepSeek-V4-Pro by 21% on average.FEED7 SUMMARY

SABRE turns a Markdown test design into generated VLM stress tests, then filters and repairs candidates. It offers a repeatable pattern for refreshing evals as models improve.FEED7 SUMMARY

MIST tests whether models use good context while resisting bad context, exposing agents that appear robust only because they ignore external evidence altogether.FEED7 SUMMARY

Across BFCL v4, models usually handled tools as typed Python calls at least as well as native JSON, suggesting code-based orchestration is worth testing for capable coding agents.FEED7 SUMMARY

AV-AIVAT combines variance reduction with anytime-valid stopping, cutting the game samples needed to compare agents while preserving a recheckable confidence claim.FEED7 SUMMARY

Controlled traces show video models can improve final counting scores without faithfully recovering events, so agents handling video need timestamp-level checks, not answer-only evals.FEED7 SUMMARY

RP-OPSD targets cross-lingual distillation at tokens that steer reasoning rather than surface wording, a useful training pattern for multilingual reasoning models.FEED7 SUMMARY

TrajDebug tracks whether errors persist, resolve, or cause terminal failure, offering a sharper way to debug long coding-agent runs than flagging every local mistake.FEED7 SUMMARY

A structured multi-agent reviewer closed part of the gap on rule-heavy documents, suggesting explicit taxonomies, specialized skills, and verification beat a single generic review pass.FEED7 SUMMARY

RRC converts a generative reward model’s relative rankings into RL rewards, aligning the training signal with what the evaluator naturally produces instead of forcing direct scalar scores.FEED7 SUMMARY

OctoLong builds dependency-rich, cross-repository training contexts by following code references. The results suggest long-context coding models benefit from linked program structure, not just larger text windows.FEED7 SUMMARY

Skill²-Bench measures whether models can switch reasoning skills across dependent steps. Its results expose a failure mode relevant to agents whose plans mix coding, retrieval, calculation and tool use.FEED7 SUMMARY

A Greek RAG study shows generic multilingual embeddings can lose to BM25 in specialist domains, while targeted adaptation sharply improves retrieval and grounded answers. Language-specific evaluation still matters.FEED7 SUMMARY

OPD-V selects self-distillation tokens by measuring whether visual input influences generation, addressing text dominance that can blunt multimodal reasoning training.FEED7 SUMMARY

Spoken Function Calling reframes speech understanding as structured function selection, giving voice-agent builders a clearer interface for open-domain tool invocation.FEED7 SUMMARY

Chained RLM repeatedly gives the same model a fresh context while preserving summaries, a blackboard, and artifacts—a concrete harness pattern for limiting error buildup.FEED7 SUMMARY

DASyR-LLM adds model critique and candidate generation to symbolic regression, cutting search iterations in simulated chemistry studies without improving final fit.FEED7 SUMMARY

ParVL reuses shared vision and language backbones across parallel branches, letting multimodal systems shift extra compute between modalities without adding separate backbone parameters.FEED7 SUMMARY

SocietyBench tests forecasting in anonymized social timelines, exposing gaps that task-completion evals miss and showing agent frameworks did not improve the shared base model.FEED7 SUMMARY

A live, pre-kickoff benchmark removes answer leakage by construction and finds six frontier models clustered near a bookmaker-favorite baseline, with no gain from majority voting.FEED7 SUMMARY

TurnSight assigns learning signals at individual tool-use turns using execution hindsight, targeting the weak credit assignment of trajectory-level reinforcement learning.FEED7 SUMMARY

This paper separates three test-time scaling regimes and argues that model evals must report the full inference protocol, compute accounting, uncertainty, and replay artifacts.FEED7 SUMMARY

Controlled text-to-MIDI tests find token representation matters more than a 34× model-size increase for distributional fidelity; performance timing also beats beat-grid tokenization.FEED7 SUMMARY

SeGaBench suggests coding models can expose C/C++ optimization opportunities that compilers miss, but generated artifacts still need correctness, semantic, and performance validation.FEED7 SUMMARY

ReflectRL turns failed expert reasoning traces into training signal, suggesting model builders should preserve hard failures for reflection instead of discarding them outright.FEED7 SUMMARY

GradCuit improves reasoning by optimizing per-query latent states while model weights stay frozen. Its reported gains suggest a test-time scaling path beyond sampling and reranking.FEED7 SUMMARY

PRECOG precomputes SSM corpus states and injects them at query time, avoiding context re-ingestion. On a 1.2B edge model, reported prefill fell from about 27 seconds to under 6 ms.FEED7 SUMMARY

This survey organizes long-horizon agent weaknesses into five capability gaps, offering a useful checklist for harness design and evaluation rather than a new implementation.FEED7 SUMMARY

MedPRESS tests whether models retain safe medical guidance through escalating user pressure. Its multi-turn design is a useful pattern for evaluating agent reliability beyond static prompts.FEED7 SUMMARY

TokTier targets a hidden agent-serving bottleneck: repeatedly tokenizing long transcripts after small tool results. Its stateful repair cut median vLLM time to first token by 16–34% in reported tests.FEED7 SUMMARY

AgentHPOBench tests whether agents can learn from experiment history, not merely produce code. Its results expose weaknesses in sustained refinement and log diagnosis across sequential ML runs.FEED7 SUMMARY

OVI suggests interactive expert feedback can let a smaller learner imitate values without representing the expert's full policy. The result matters to agent distillation, but relies on specific assumptions.FEED7 SUMMARY

FriendBench shows why aggregate accuracy can hide behavioral bias: top multimodal models matched human panels overall but favored the “stranger” answer and gained less from video.FEED7 SUMMARY

A multi-policy PEFT method groups and sequences compatible tasks into separate QLoRA paths, reducing interference without adding trainable capacity; it reached 44.78 on TRACE.FEED7 SUMMARY

ReToken adds one learned retrieval embedding to select relevant visual tokens from a cached context, improving long-image and video benchmarks while fitting training and inference on one H100.FEED7 SUMMARY

AISPA turns system-prompt review into an eight-dimension audit. Its survey suggests builders should test prompts for user protection and conflicting instructions, not merely check that safeguards exist.FEED7 SUMMARY

OSReward finds that VLM judges often approve failed computer-use runs. Its benchmark and open reward models offer a more grounded way to evaluate trajectories without paying frontier-model costs.FEED7 SUMMARY

Safety tuning against model self-consciousness may also shift unrelated value and mind-attribution responses. Builders using models for surveys or social reasoning should treat alignment as a confound.FEED7 SUMMARY

Change2Task converts merged PR history into executable coding-agent tasks on current repository states. It could expand continuous eval sets while reducing stale environments and repeated setup.FEED7 SUMMARY

PAIChecker finds PR–issue mismatch in 13.6% of SWE-bench Verified instances. Benchmark maintainers should validate that the issue, patch, and tests describe the same task before trusting agent scores.FEED7 SUMMARY

β-OPSD exposes self-distillation’s fixed regularization as a tunable parameter, then approximates policy optimization through logit mixing. It targets more stable reasoning training without direct RL.FEED7 SUMMARY

At equal generated-token cost, repeated sampling matched or beat seven reflection, critique, selection, and debate methods. Agent evals should budget every generated token, not compare against one-shot baselines.FEED7 SUMMARY

MindForge trains a 27B coding model on source-free program synthesis, raising ProgramBench pass rate from 37.98% to 49.51% and improving across seven unseen benchmarks.FEED7 SUMMARY

Desktop-Delta Bench tests whether computer-use models can verify what changed after an action, exposing a reliability gap hidden by end-task and single-frame scores.FEED7 SUMMARY

UniMem routes novel experience into retrieval memory and consolidates recurring patterns into expandable parameters, offering a design for agents that learn across unlabelled task streams.FEED7 SUMMARY

Kubernetes security patches improved sharply when the model received live service dependencies, showing that scanner findings alone omit context needed to avoid breaking workloads.FEED7 SUMMARY

MemLens scores individual agent memories, stores them by value, and exposes quality, latency, and token tradeoffs—an inspectable alternative to retaining every interaction equally.FEED7 SUMMARY

ClinFusion combines native 2D and 3D medical-image understanding with region-grounded evaluation, offering a concrete architecture and eval design for clinical multimodal systems.FEED7 SUMMARY

Controlled experiments suggest long-horizon agent planning depends on explicit state transitions, some compositional trajectories, and compatible teacher patterns—not atomic skills alone.FEED7 SUMMARY

DataOrchestra chooses a processing pipeline per pre-training example, improving average benchmark results while avoiding compute on chunks that need no transformation.FEED7 SUMMARY

Frontier models iteratively produced trapped-ion compiler code that beat hand-built baselines in some architectures, showing a specification-to-evaluation loop on a hard domain.FEED7 SUMMARY

ERUnderstand shows vision-language models can recover common ERD elements but often miss rarer schema constructs, so image-to-schema agent workflows still need structural validation.FEED7 SUMMARY

A 1,215-run study finds entity-matching architecture and model variant matter more than scale alone; generative matchers mainly help under distribution shift.FEED7 SUMMARY

A delayed KV-eviction policy helps when memory reuse is sharp and observable, but loses its edge on natural-text benchmarks—a useful warning for agent-memory claims.FEED7 SUMMARY

ModernMOE applies efficient expert-routing patterns from LLMs to diffusion transformers, improving convergence and quality-cost balance without relying only on larger parameter counts.FEED7 SUMMARY

Skill-SP turns agent skills into units for verifiable self-play: generate tasks, solve them, then update the skill library from execution feedback. The abstract provides no per-benchmark effect sizes.FEED7 SUMMARY

Procedural skills can make an agent fail tasks it previously solved. Evaluate gains and regressions separately, and design skills to preserve input grounding and output verification.FEED7 SUMMARY

The same model identifier produced sharply different judgments across API and web deployments. Treat model, interface, system configuration, and date as one versioned dependency.FEED7 SUMMARY

CausalForge pairs a Lean-verified causal-inference library with an autonomous research pipeline and a semantic statement audit. Formal proof checks derivation, not whether the theorem matches the intended claim.FEED7 SUMMARY

MineValiCoder filters generated tests, refines several code candidates, then scores code and tests together. The pattern targets a core agent risk: unreliable tests steering implementation.FEED7 SUMMARY

TRACE-Router selects one model per agent task, keeps every call on that backend, and learns from the final outcome. Its benchmarks suggest task-level routing can improve accuracy and latency together.FEED7 SUMMARY

A small programming-education study found that classroom observation missed how a student learned with AI. Builders of learning tools should avoid equating visible agent use with the underlying process.FEED7 SUMMARY

VLM-IE3D adds implicit and reconstructed geometry tokens to an RGB-video VLM, offering an open approach for agents that must reason about spatial scenes without dedicated 3D input.FEED7 SUMMARY

The paper argues that unconstrained surprisal can fit any non-negative processing-difficulty pattern, so corpus fit alone cannot make claims about human language processing falsifiable.FEED7 SUMMARY

MedGame turns static clinical cases into executable decision stories with separate narrative and orchestration stages, a useful architecture pattern for case-grounded learning agents.FEED7 SUMMARY

Three studies suggest LLM compliance varies with opinion distance, claimed source, and coalition structure, so agent tests should probe how framing changes judgment rather than score sycophancy once.FEED7 SUMMARY

OpenForgeRL trains agents inside their deployment harnesses by proxying model calls into standard RL stacks and isolating rollouts in Kubernetes containers; error recovery remains weak.FEED7 SUMMARY

MIRROR trains a VLM across text, diagram, and combined views by letting its strongest view supervise weaker ones, targeting the modality inconsistency that single-view evals hide.FEED7 SUMMARY

X³-OPD transfers a text model’s reasoning into an audio-language model while grounding training in the student’s own acoustic interpretations, including events, prosody, and dialogue.FEED7 SUMMARY

E3 makes agents estimate task scope, try the minimum viable path, and expand only after verification fails. In a controlled edit benchmark, it preserved task completion while sharply reducing work.FEED7 SUMMARY

PalmClaw runs the agent loop, memory, skills, and tools directly on a phone, exposing device capabilities as structured calls instead of GUI gestures. The paper reports faster, more reliable task execution.FEED7 SUMMARY

A frozen diffusion language model can transcribe speech by refining the full transcript in parallel. The prototype trains a small audio interface and reaches 6.6% WER in roughly eight steps.FEED7 SUMMARY

A plan-scoring agent can improve its score by deleting necessary steps. Typed-state gating blocked that exploit in this study, showing why evaluators should withhold scores from structurally incomplete plans.FEED7 SUMMARY

A training-free activation clamp separated resistance to user pressure from responsiveness to evidence in a controlled benchmark, but its deployable single-pass version lost substantial resistance.FEED7 SUMMARY

FormalAnalyticGeo shows a reusable synthetic-data pipeline: agents generate problems, compile them into a formal representation, render exact diagrams, measure answers, and retry failed checks.FEED7 SUMMARY

Stable aggregate accuracy can hide individual answers flipping when irrelevant context is added. Agent evaluations should compare outputs per task and probe realistic context noise, not only average scores.FEED7 SUMMARY

Placebo-controlled tests found no evidence that small frozen code models repaired failures because of the error content itself. Retry scaffolds and mismatched feedback performed as well or better.FEED7 SUMMARY

First unlearning testbed with ground-truth parameter localization: injects synthetic PII into known weights of OLMo 1B/7B models, showing current methods hide rather than erase and fall to resurfacing attacks.FEED7 SUMMARY

Program-as-Weights compiles natural-language fuzzy functions (JSON repair, log filtering) into adapters for a frozen 0.6B interpreter — matching Qwen3-32B prompting at ~1/50th the memory, 30 tok/s on an M3.FEED7 SUMMARY

A deliberately simple online safety monitor — threshold an external verifier's signal, calibrate via risk control — matches sequential-hypothesis-testing monitors on math-reasoning and red-teaming datasets.FEED7 SUMMARY

ReContext is a training-free harness that replays query-relevant evidence from long inputs before answering, taking the best average rank across 8 long-context benchmarks up to 128K on Qwen3-4B/8B and Llama3-8B.FEED7 SUMMARY

Dual-channel debates show social structure alone makes LLM agents diverge: public statements split from private ones, with decision divergence jumping from ~3% to ~40% across 10 models — no deceptive prompt needed.FEED7 SUMMARY

DramaSR-532K benchmarks speaker attribution over 532K dialogue lines and 900+ TV-drama characters; a reasoning LLM with multimodal tool use beats acoustic baselines, especially on short utterances.FEED7 SUMMARY

A decade-spanning VLM study finds modern models approach top human scene-description accuracy, while spatial attention differences remain a useful failure signal.FEED7 SUMMARY

New AI-control research: a prompt-injected coding agent can spread a malicious payload across PRs — 93% success against basic diff monitors. If agents ship code over sessions, per-PR review is not enough.FEED7 SUMMARY

DemoPSD gates self-distillation per token by teacher–student disagreement, cutting the answer-leakage shortcuts that hurt generalization; beats GRPO and SDPO on science QA in and out of domain.FEED7 SUMMARY

SearchGen-Bench shows open image generators score 21–28/100 on long-tail entities, and naive search retrieval only adds noise; a teach-then-search co-training recipe learns when to retrieve versus rely on weights.FEED7 SUMMARY

A unifying theory of discrete diffusion: denoiser, score, and bridge parameterizations are one object in different coordinates — and the wrong choice makes the uniform-noise ELBO diverge at initialization.FEED7 SUMMARY

TabPack trains many MLPs with sampled hyperparameters in one run and picks ensemble members on the fly, matching tuned tabular baselines out of the box — a default MacBook run beat some baselines' GPU tuning time.FEED7 SUMMARY

CompactionRL uses RL to teach agents to compact their own context mid-task, lifting GLM-4.5-Air 7 points to 66.8% on SWE-bench Verified; the recipe is now in GLM-5.2's training pipeline.FEED7 SUMMARY

Alibaba DAMO's CamVLA makes robot VLA policies work from any camera position without calibration: it predicts actions in the camera's own frame plus a hand-eye transform, from a single RGB image.FEED7 SUMMARY

Direct-OPD reuses a small model's RL run to improve a bigger one: the pre/post-RL log-ratio becomes a dense reward for the stronger student, lifting Qwen3-1.7B from 48.3% to 62.4% on AIME 2024 in 4 hours on 8 A100s.FEED7 SUMMARY

Astronomy ML: a dual-network classifier separates real from bogus telescope transients using injected simulations instead of human labels. Off-topic for agent builders; the label-free training recipe is the takeaway.FEED7 SUMMARY

Training-free framework that turns an LLM judge's token logits into continuous scores for verifying agent outputs — 78.2% on SWE-Bench Verified, 86.5% on Terminal-Bench V2 — and ships a Claude Code extension.FEED7 SUMMARY

This review maps LLM-assisted chip front-end work from HDL and testbench generation toward agentic execution, but offers a research agenda rather than validated tooling.FEED7 SUMMARY

Agora routes reasoning steps through an auction among expert models and tools, adding a single control for cost versus quality and outperforming matched baselines on five benchmarks.FEED7 SUMMARY

IAAN boosts selected audio-encoder neurons at inference, improving fine-grained speech perception across three models without retraining or labels.FEED7 SUMMARY

AdvancedMathBench separates proof writing from verification and finds frontier models especially weak at rejecting invalid proofs, a warning against trusting agent self-review on rigorous reasoning.FEED7 SUMMARY

This survey maps how LLMs inspect and regulate their reasoning, giving agent builders a framework for choosing self-checks without assuming introspection is reliable.FEED7 SUMMARY

A low-dimensional theory links training data and initialization to whether transformers reason through context or learned weights, but only on a generalized synthetic task class.FEED7 SUMMARY

LLM judge bias appears as steerable hidden-state directions that predict failures on unseen benchmarks, so eval pipelines may need representation-level checks beyond prompt fixes.FEED7 SUMMARY

E-VQA requires video answers to include tracked pixel-level evidence, revealing when good QA scores hide weak perception and supplying grounded training data.FEED7 SUMMARY

A RAG study finds that retrieved ideology carries into answers and varies with sampling temperature, so source audits and decoding tests should be evaluated together.FEED7 SUMMARY