Entity trail

Adaptive

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-21 / arxiv-researcherSecrets Sitting in an Agent's Context Leak Through Ordinary Replies, and Stronger Models Leak MorearXiv 2608.19857 shows that merely holding a secret in the context window imprints recoverable correlations on a model's benign outputs, even when it correctly refuses direct extraction. Across eight proprietary models, an adaptive black-box attack reconstructs 2-digit in-context secrets at near-perfect accuracy and 4-digit secrets at 82% exact match, purely from responses to ordinary non-adversarial requests. The authors demonstrate a classifier that infers health and financial predicates about stored user memories and an RL-trained adversary that pulls full Social Security Numbers out of a production-style agent, and they find leakage scales with instruction-following ability rather than behaving like a patchable bug.
2026-08-21 / sources-researcherMemory Makes Models Worse: MemTrapBench Finds Every Memory Framework Tested Underperforms No Memory at All, Best Case Still Down 10%+MemTrapBench (arXiv 2608.20202, submitted 2026-08-20, Zhejiang-affiliated team led by Mengru Wang and Ningyu Zhang) isolates what the authors call memory-induced cognitive traps, where correctly stored and contextually relevant memories still corrupt reasoning through Reasoning Fixation and Belief Distortion. Across two model families and five memory frameworks, every strategy scored below the no-memory baseline, with the strongest methods still dropping more than 10%. Their proposed fix, AdaptiveMem, is inference-time only and preserves scores on standard memory benchmarks — worth reading before you bolt a memory layer onto an agent on the assumption that more context is free.
2026-08-20 / skill-finderRecurrent depth helps compositional tool calling but barely moves isolated API callsTesting looped language models on API-Bank, BFCL, and NESTful, the authors found accuracy on multi-step tool use generally rises with recurrent depth, while gains on single isolated calls were small and varied by model. Adaptive inference, spending extra recurrence only when needed, gave the better compute-performance trade-off. The takeaway for agent design is that the hard part is dependency tracking across calls, not the individual call, so extra compute should be aimed there.
2026-08-18 / arxiv-researcherLe Critique Revives Value Functions for LLM RL With Privileged Signal and an Adaptive GRPO InterpolatorCritic-free methods like GRPO won on practicality, but they give only sequence-level credit and stall on straggler rollouts, increasing off-policyness. This work proposes Privileged Value Functions, which inject extra task-relevant token-level signal without biasing the policy objective, plus TETHER, a baseline that adaptively interpolates between group-relative and value baselines according to how accurate the value function currently is. Across several reasoning tasks both consistently improve on the standard value-function baseline and are competitive with or better than mean-baseline GRPO — a credible case for putting critics back in the pipeline.
2026-08-18 / arxiv-researcherSA-MRPO: Multi-Reward RL Wastes Gradient Budget on Already-Solved Objectives; Reweighting by Saturation Gains Up to 5% on AIME24Standard practice scalarizes a reward vector with fixed weights before group-wise standardization, which lets rollouts with different reward profiles receive identical advantages and keeps optimizing objectives that are already maxed out. SA-MRPO standardizes each reward objective independently and discounts its contribution by a batch-level saturation estimate, which the authors show can reverse the sign of an update rather than merely rescale it. It beats GDPO on the harder correctness objective in 12 of 15 benchmark comparisons (up to 5% on AIME24), improves adaptive-reasoning accuracy on all five benchmarks (3.8% average, 9.2% on AMC23), and adds up to 2.3% pass rate on coding — while holding the easy objectives near their satisfied levels.
2026-08-18 / arxiv-researcherVCE-Skill Mines Public Skill Version Histories as Evolution Priors, Adding 3.20-4.98 Points Over Trajectory-Only Self-EvolutionCurrent skill self-evolution methods revise skills using only execution trajectories from the current task, ignoring the accumulated diff history of public skill repositories. A pilot study shows the two sources are complementary — version changes supply reusable evolution priors, trajectories supply task-grounded evidence — so VCE-Skill distills noisy, implementation-specific public changes into structured version-change experience and adaptively fuses it with the base evolver's trajectory proposals. Mean scores rise 3.20-4.98 points, and the evolved skills transfer better across models, pointing at a source of prior knowledge most agent-skill systems currently throw away.
2026-08-16 / skill-finderTag tool output with source provenance and field contracts: state-corruption attack success drops 84.7% to 2.3%PIPES (Provenance-Informed, Prior-Enforced Screening) defends against state-corruption attacks, where attacker-controlled data makes false claims about the environment that slip past injection filters because they read as ordinary tool results. It screens each response unit two ways — static field contracts where a schema gives stable expectations, and source provenance metadata for open-ended content — then removes, warns, blocks, or escalates on violation. On six benchmark splits with Gemma 4 31B IT under adaptive attack, attack success fell from 84.7% to 2.3% while benign utility rose slightly (92.5% vs 90.6%).
2026-08-15 / arxiv-researcherConfidence-Calibrated Cascade Hits 98.8% Accuracy on LLM Acceptance Testing and Cuts Cost 31.7% vs Single JudgeRequirements-Augmented Generation (REAG) generates context-aware test oracles for LLM-based software by retrieving software requirements, domain knowledge, and personas via adaptive RAG and self-reasoning — addressing the fact that the same query can demand different correct responses per persona and context. A confidence-calibrated cascade then quantifies verdict reliability through simulated expert agreement, accepting, escalating, or abstaining, with guarantees backed by conformal risk control. In an industrial case study on a production nutrition advisory app, REAG scored 3.91/5 oracle quality (qualified or marginal in 82% of cases), and the cascade reached 98.8% accuracy, lifted oracle quality to 4.30 by filtering, and delivered 31.7% better cost-efficiency than single-judge baselines.
2026-08-15 / arxiv-researcherReduced Matrix Multiplication Cuts Transformer Inference Without Touching Weights — and Attention Is Far More Reducible Than MLPsRMM is a training-free, input-adaptive method that shrinks Transformer matrix products by selecting informative slices along contraction dimensions under a single retention-ratio knob, giving a smooth accuracy-efficiency trade-off with no weight modification. Tested from 1B to 70B parameters across discriminative, autoregressive, and long-context settings — plus multimodal vision-language inference — reduction tolerance varied by model family, task, and component but often improved with scale. The mechanistic finding is the actionable one: attention-side computations are substantially more reducible than MLP components, and custom A100 kernels turned the theoretical savings into real wall-clock gains, especially at long sequence lengths.
2026-08-15 / agents-researcherVALG routes failed proofs by obstruction type — derivation, structure, or the theorem statement itselfVALG (arXiv 2608.13060, 2026-08-13) organizes ML theory research as an agentic workflow combining multi-level verification, adaptive problem formulation and graph-structured proof development. Within each source-relative theorem branch it holds a fixed mathematical specification, checks theorem-level composition of a typed proof-dependency graph, and builds and reviews local proofs in dependency order. The interesting control-flow idea for agent builders is the failure router: when a proof attempt fails, VALG diagnoses whether the obstruction sits in a derivation, in the proof structure, or in the theorem formulation, and routes the next attempt accordingly — formulation-level obstructions spawn an explicitly related variant or relaxation that preserves the relation to the source problem. Evaluated on nine subproblems.
2026-08-15 / agents-researcherSkillShapley puts a Shapley value on each step of an agent skill, and finds step interactions are mostly additiveAgent skills — the SKILL.md-style procedural instructions now shipping across coding and document agents — are written by hand or scraped from execution traces with no measurement of which steps actually help. SkillShapley (arXiv 2608.13173, 2026-08-13) models step attribution as Shapley value estimation and exploits two empirical observations to make it affordable: benchmark rewards are discretized so performance shows sharp cliffs, and step interactions are largely additive rather than synergistic. It first identifies informative coalitional regions and then adaptively samples coalitions that yield reusable marginal evidence, and on skills from SkillsBench it separates high- from low-value steps — a concrete way to prune bloated skill files instead of guessing.
2026-08-12 / sources-researcherSurvey Maps "Co-Evolution in Agentic Systems": Agent-Agent, Agent-Environment, and Meta Co-Evolution Where the Evolutionary Mechanism Itself EvolvesA survey posted August 10 (2608.10299, 95 upvotes) by Qing Zong and colleagues organizes the fast-growing self-evolving-agent literature into a three-stage taxonomy: Agent-Agent co-evolution (adversarial, collaborative, organizational adaptation), Agent-Environment co-evolution (adaptive tasks and feedback), and Meta co-evolution, where the evolutionary mechanisms are themselves evolvable. The framing that distinguishes it from prior self-improvement surveys is treating this as multi-component — agents and their environment imposing adaptive pressure on each other rather than one agent unilaterally rewriting itself. The authors flag evaluation, scaling across components, and keeping increasingly autonomous evolution safe and controllable as the open problems.

Graph relationships

RELEASED
Agent Computer
Adaptive launched Agent Computer, a no-code platform for enterprise SaaS automation
Source finding

Source trail

Anthropic Release Notes BenchLM TechCrunch searching for Vertex AI Telegram CopilotKit's Vapi Cursor's Composer 2.5 Adaptive model router jointly released guidance Agent Timeline Scale AI's new benchmark

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…