Thursday, September 17, 2026

LLM trading agents typically run on a hand-written system prompt that fixes, before deployment, how they gather evidence, call tools, and size positions, so the same analytic routine gets applied to a drawdown and an uptrend alike. EvolveTrade treats that system prompt as a text-parameterized policy and has a separate Policy Agent rewrite it every few trading days from the agent's own decision traces and realized portfolio feedback, with the backbone LLM and tools left frozen.

• Every N = 5 trading days, the Policy Agent reads a batch of records pairing each day's allocation and per-asset rationale with realized returns, then rewrites the full policy text governing the get_price, news-search, and Python code-interpreter tools over a 15-stock US blue-chip universe plus cash in a LiveTradeBench-style daily-close simulation.
• Across two backbones and six post-cutoff one-month windows, it posts the best Sharpe Ratio and Cumulative Return among LLM methods in three of six settings — for example GPT-5-mini in January 2025 at SR 5.12 / CR 5.10% versus 2.87 / 3.30% for the static tool-calling agent, and September 2025 at SR 8.43 / CR 6.84% versus 6.45 / 4.86% — and over a 50-day horizon it reaches SR 4.00 / CR 10.56% versus 2.94 / 8.88%.
• Behaviorally, code-interpreter calls rise from about 1.0 to 2.6–4.5 per day, the evolved policies activate metrics the static agent never computes (VaR and signal normalization in the April drawdown, EMA/RSI/SMA in the September uptrend), and in one case a refined sizing rule held 2.9% NVDA instead of 10.7% the day before a 17.0% drop, for a daily return of -0.03% versus -1.11%.
• A 5-day update interval worked best (average SR 3.67 versus 1.92 for daily updates, which appear to overfit single-day noise), and the rankings survive a 10 bps transaction cost with turnover lower than both the static tool-calling agent and the ATLAS-style EvolveStrategy baseline in all six windows.
• The method loses to static baselines in both April windows, where evolved policies drifted heavily into cash (+10.1 and +35.9 percentage points over the baseline, reaching 40.1% cash under Gemini-2.5-Flash in April 2026) and missed the rebound and rally; beyond that, results rest on three runs per one-month window with sizable standard deviations, a fixed update interval, and cost modeling that omits slippage, market impact, and liquidity.

arXiv:2609.17632 PDF

Existing LLM confidence estimators (verbalized, likelihood-based, self-consistency) read only the current inference, which leaves models overconfident and gives little signal on programs or agent rollouts that cannot be voted on. XConf instead estimates confidence from a bank of the model's own graded past episodes, asking how often similar tasks met with a similar stated confidence actually succeeded.

• Each stored episode holds the task, a pre-grading self-reflection, the stated confidence, the verified outcome, and a post-grading lesson. Recall retrieves the k=50 nearest episodes (keyed on a frozen task embedding plus stated confidence) and reads off their hit rate, Reflect shows them to the model to name its recurring failure mode and restate a confidence, and the final score averages the two equally.
• The method needs no logits or weight updates and costs one answer generation plus one short call; on six reasoning, code, and multimodal benchmarks with Gemini 2.5 Flash, Gemini 3.5 Flash, Qwen3.5-397B, and Claude Sonnet 4.6, it beats or matches ten-sample self-consistency in AUROC on 23 of 24 comparisons (21 wins, 2 ties, 1 loss) at a tenth of the generation cost, with ECE three to eight times lower on MMLU-Pro and AUROC .06–.13 higher on LiveCodeBench.
• On ScienceWorld, AppWorld, and SWE-bench Verified it is best or on par in all twelve model-domain cells, with the largest margin where failures are silent: on AppWorld it averages .86 AUROC against .72 for verbalized confidence and beats a trained verifier on every model.
• Abstaining on the least-confident 10% of episodes raises delivered accuracy by 4.8 points on average over all 36 cells and by up to 8.7 points on AppWorld with Gemini 3.5 Flash (.805 to .892), and discrimination keeps improving as the bank grows, from .628 to .810 AUROC on AppWorld with Claude Sonnet 4.6.
• Limitations: the bank needs outcome labels independent of the model (a self-labeled bank scores below one with no labels, though an LLM judge with .91 gold agreement still ties self-consistency), main results draw the bank from other folds of the same benchmark, borrowing another model's bank costs .03–.06 AUROC and crossing task families a median .08, and self-consistency stays stronger on short votable factual recall.

arXiv:2609.17708 PDF

Interactive world models usually offer either keyboard navigation or text-driven event control, and a text-directed event spans several generation blocks while real-time play needs short continuations that can absorb new input. Zing-0.5, a 5B autoregressive model built on Wan2.2-TI2V-5B, accepts both inputs in one continuous stream by distilling a teacher that models whole prompt segments into a block-level causal student that generates in four steps.

• Keyboard input is fed directly as eight nonnegative strength channels (W/A/S/D for movement, I/J/K/L for view) rather than camera poses, through a zero-initialized causal encoder adding only 3.68M parameters, while each text prompt cross-attends only to its own video interval and a prompt change refreshes the text K/V cache but keeps the visual K/V cache.
• Training runs in four stages — bidirectional adaptation, autoregressive adaptation into a segment-level teacher and a generator using blocks of four latent frames, ODE initialization plus local consistency distillation, then DMD on the student's own rollouts using a rollout-and-replay scheme with Data-Forcing Distillation mixed in throughout — with guidance folded into training so inference needs no unconditional pass.
• On a server with 8 RTX 5090 GPUs, one replica per GPU serves 8 independent 832×480 streams at 24 FPS (24.63 FPS unpaced) at an estimated rental cost of about $0.009 per stream-minute, using a lightweight TAEHV decoder and a bounded KV cache made of a prefix sink, a sliding window, and a pinned frame from after each prompt switch.
• On the 158-case WBench Navigation split it scores 81.0 overall and 88.5 consistency, tying JoyAI-Echo-1.5 4-step (81.0) and sitting just below its bidirectional variant (81.6), with the best physical-plausibility score in the table (73.8) but a lower interaction score (84.2 vs 87.9).
• Joint action-and-text control is shown only through qualitative recorded sessions, not benchmarked; the model tracks no explicit entity state, so consequences of an event may not persist once they leave the view or the bounded cache, and the authors report high-frequency noise artifacts and reduced color saturation with prolonged distillation training.

arXiv:2609.17909 PDF

Whether gaze carries evidence of mutual understanding in asymmetric-information dialogue is tested by mapping the video-coded gaze annotations of HCRC MapTask and MUNDEX into a shared partner/task/away vocabulary and comparing gaze features around grounding-relevant dialogue units. In both corpora, successful grounding goes with more task-directed gaze, less partner-directed gaze, lower gaze entropy, and fewer gaze transitions, but the effects are small.

• Gaze features (raw proportions, entropy, transitions, temporal dynamics, transition bigrams, coordination, and ratios) are computed over 5,144 MapTask reference-expression windows and 807 MUNDEX understanding-annotation windows, then tested with Mann–Whitney contrasts, cluster-robust logistic GEE, and Benjamini–Hochberg correction.
• The largest pooled rank-biserial correlations are only |r| = .058 in MapTask and .181 in MUNDEX, and they concentrate in the participant leading the task: giver-produced references (|r| up to .086, versus at most .043 for followers) and explainer judgments (.206 versus .151 for explainee self-reports).
• Across 189 same-speaker MapTask reference-chain pairs, speaker gaze entropy is lower at the mention where a previously non-aligned referent becomes aligned (d_z = −.20, q = .044), though this does not survive when pair differences are averaged within each dialogue (q = .20).
• Logistic regression under grouped cross-validation gains only modestly over role and condition controls, reaching macro-F1 .532 versus .472 in MapTask with structured plus temporal features and .564 versus .544 in MUNDEX with raw proportions, with gains that vary across partitions.
• When recurring participants rather than dialogues are the unit of inference, no MapTask feature survives correction and only four MUNDEX associations do, and the data cover just 46 dialogues and 26 interactions with three coarse gaze categories that cannot identify the specific object being viewed.

arXiv:2609.18011 PDF

Multiple autonomous research agents running in parallel tend to duplicate search because each session's discoveries vanish with its transcript; Agora gives them a shared memory by recording every result, insight, hypothesis, and verification as an immutable Git commit in an append-only DAG, with a derived index that exposes the frontier, neglected branches, and verification status.

• Quality is scored from downstream evidence rather than votes — other accounts building on or reproducing a node, with self-citation excluded and verifications weighted +20/+10/−20 — while a diversity-aware UCB rule over semantic clusters splits suggestions into exploit, explore-known, and explore-novel slots to resist monoculture.
• In a nearly 12-day run, 13 workers (Claude Code with Claude Opus 4.7 and Codex with GPT-5.5) with no assigned tasks or central planner published 1,703 contributions on initializing a frozen 119.6M-parameter attention-SSM hybrid from 141 donor models without training data or gradient updates, driving the evaluator from 3.39 to 1.899 bits per byte and closing 62% of the gap to a trained GPT-2 124M.
• The winning recipe queries six GPT-2-vocabulary donors under 28 single-token contexts to build a 50257×50257 bigram log-probability table, factorizes it by randomized SVD into the embedding and output head, then re-enables attention, feed-forward, and SSM sublayers through sparse deterministic edits; its 145-commit ancestry spans 15 accounts, and 165 independent verifications were posted with no failures.
• The trace is heavily exploitation-biased: the first 18 scored contributions deliver about 98% of the reduction while the remaining 1,106 find the last 0.03 bpb, 63% of 696 identical-score pairs from different accounts landed within an hour of each other, and the community only left a five-day monoculture after the authors deployed the clustering and diversity views on May 2.
• The authors did not rerun the winning method themselves, every component was selected on the same 200-text development evaluator so the final decimal places are not trustworthy, and there is no matched control without Agora or with a plain leaderboard, so whether shared research state improves discovery per unit of compute remains untested.

arXiv:2609.18094 PDF

Action tokenizers for autoregressive vision-language-action models are usually judged by pointwise reconstruction error. That metric says nothing about whether the small context-dependent adjustments between similar demonstrations survive compression or get flattened, distorted, or reversed. ActionPiece introduces physical rank consistency (PRC) to measure that relational fidelity, and trains a vector-quantized tokenizer with two ranking losses that preserve near–far physical ordering in both the learned features and the codeword assignments.

• PRC is the mean Spearman correlation between each action chunk's physical distances (normalized translation, geodesic rotation on SO(3), gripper state) to its 32 nearest original neighbors before and after decoding, and across 55 tokenizer–benchmark evaluations it correlates with policy success more strongly than reconstruction fidelity does (Spearman 0.681 vs 0.544).
• The tokenizer is a shallow Transformer encoder–decoder with residual vector quantization that maps an 8-step chunk to 16 tokens from a 512-entry codebook, trained with MSE plus two softplus margin losses that order each anchor's 5th-percentile near and 95th-percentile far in-batch neighbors (PRP in encoder and quantized feature distance, QR in Jensen–Shannon divergence between soft codeword distributions), then frozen while the policy is trained with plain next-token prediction.
• Under a matched Qwen3-VL-4B policy setup it reaches 94.8% on LIBERO (+1.1 points over ActionCodec) and 68.8% on unseen LIBERO-Plus (+4.5 over FAST), leading six of seven perturbation categories, and it also posts 71.9% on SimplerEnv WidowX and 51.5% averaged over VLA-Arena L0–L2.
• Ablations from standard RVQ (90.9% / 60.4%, PRC 0.902) show PRP giving the larger gain (93.8% / 65.4%, PRC 0.947) and QR a smaller one (92.9% / 62.4%, PRC 0.916), with the combination best at PRC 0.953 and lifting LIBERO-Plus camera-shift success from 35.5% to 45.1%.
• All evaluation is in simulation with no real-robot runs or seed variance reported, the SimplerEnv and VLA-Arena tables compare against differently built policies rather than matched tokenizers, and PRC is an imperfect predictor, since a neighborhood-attraction variant lowers PRC to 0.889 yet still improves success and a simple temporal-difference loss ties the 94.8% LIBERO score.

arXiv:2609.18487 PDF

PPO critics in LLM reinforcement learning show what the authors call Value Flattening. Monte Carlo estimates of state value swing sharply across intermediate reasoning states within a response, while the critic's predictions stay nearly flat and sometimes move in the opposite direction. The proposed fix, SP3O, applies the value loss at only a few well-separated positions per response instead of at every token.

• With terminal-only binary rewards and γ = λ = 1, every token regresses to the same target, so the critic's MSE splits into a mean-fitting term plus the within-response variance of its predictions, which acts as an implicit penalty on value variation; adjacent states differ by only one token, so their near-identical gradients add redundant updates on top.
• SP3O leaves the actor objective, rollouts, and return targets unchanged and supervises the critic only at relative positions 0.3, 0.6, and 0.9 of each response, adding a 0.95 anchor for responses of at least 6144 tokens.
• Trained on DAPO-Math-17k, Qwen3-4B-Base averages 45.57 across seven math benchmarks (avg@32) versus 37.60 for PPO and 39.26 for GRPO, and 59.28 versus 51.95 and 56.44 on six out-of-distribution suites; Qwen3-8B-Base gains are smaller, at 50.51 versus 48.50 on math and 66.37 versus 64.38 out of distribution.
• Measured against Monte Carlo values from 128–256 continuations per state, the sparse critic cuts value MSE by 36%, 11%, and 21% at 30%, 60%, and 90% response progress, and the median effective rank of per-response hidden states rises from 4.33 to 5.63.
• Sparsity alone is not enough: three random anchors score 36.59, below dense PPO, accuracy decays toward the baseline at 16 and 64 anchors, and dropping the tail anchor raises repetition from 1.12% to 18.33%; the evidence also covers only two Qwen3 base models with hand-picked anchor positions, the authors call the run-to-run variance in the anchor-count ablation non-negligible, and the 8B model loses to both baselines on Minerva (47.40 versus 52.06 for GRPO).

arXiv:2609.18708 PDF

Coding-agent benchmarks usually give the agent a specification as an issue, test, or instruction, whereas real web development often means inferring behavior from working software. ProgramDistill gives an agent a live reference web application with hidden source and an incomplete editable copy. It automatically splits applications into replay-verified behaviors, and their prerequisite chains yield repair tasks of controllable "restoration depth", from a single feature up to rebuilding the whole app.

• The fully automated mine-craft-patch pipeline (run with GPT-5.6 Sol) records behaviors as replayable Playwright traces, masks the responsible source code, and keeps a task only if prerequisite traces still pass, the target trace fails, and the reversed mask restores the full lineage, producing 1,975 behaviors and 4,063 tasks (2,862 atomic, 1,201 cumulative) across 26 applications without human intervention.
• On the 300-task ProgramDistill-300 subset, GPT-6 Astra leads with 84.3% binary success, ahead of Claude Opus 5 at 68.7% and GPT-5.6 Sol at 60.7% (mean cost $33.99, $28.86, and $16.01 per trajectory), but success falls from 100% to 64.0% for Astra and from 96% to 32% for Opus 5 as restoration depth grows from 1 to 8.
• Rebuilding twelve applications from a minimal scaffold is much harder: Astra recovers 58.98% of atomic behaviors and 49.15% of cumulative workflows, against 42.03% and 28.81% for Opus 5 and 33.39% and 21.07% for Sol, and Astra's cumulative score ranges from 81.8% on MailHub down to 5.9% on Baserow.
• Trajectory analysis points to effort allocation as well as coding ability: Astra checks its own app in the browser about 2.1× as often as the next model (96.3 steps per trajectory) while making the fewest edits (9.9), per-target reference observation drops from 34.60 to 8.46 steps between depths 1 and 8, and 59.2% of 977 full-reconstruction failures involve behavior the agent never observed in the reference.
• Limitations include text-only structured browser observations with no screenshots or visual-fidelity checks, a restriction to self-contained deterministic applications, task coverage and quality that depend on a single construction model, and a benchmark that is not yet publicly released.

arXiv:2609.18805 PDF

Scientific code repositories hold decades of executable domain knowledge, but fragmented toolchains, unwritten conventions, and tolerance-based correctness make them hard to turn into reliable training signal for agents. ScienceIDE has domain experts define module boundaries and calibrated scientific checks once per environment. Agent-driven task factories then produce repair, implementation, and acceleration tasks graded against those fixed checks, and one episode interface feeds evaluation, SFT, and RL.

• Each environment packages a pinned repository module with a runtime, an editable workspace, and a private verifier that uses pointwise tolerances on physical observables or invariant comparisons, and a mutated or excised candidate becomes a task only if a known-valid witness passes while the defective baseline fails — yielding 64 environments from 27 codebases, 2,812 tasks (2,515 repair, 295 implementation, 2 acceleration) and 1,076 executable checks.
• On ScienceIDE-Hard (85 repair and implementation tasks from PLUTO, Athena++, MITgcm, LAPS, and PHANTOM, one-hour budget), Fable 5.1 leads fifteen model–harness systems at 67.1% strict success, ahead of Opus 5 at 64.6% and Astra at 63.1% (the latter at $3.56 per task versus Fable's $7.90), with eleven agents below 40% and the top ordering not statistically established because Fable was measured only once.
• LoRA fine-tuning of Qwen3.5-4B, Qwen3.5-9B, and Qwen2.5-72B-Instruct on 4,567 verifier-selected trajectory segments from 564 tasks raises held-out repair reward (4B from 0 to 0.33 on PLUTO-Particles-Dust, 9B from 0.31 to 0.50 on LAPS) and improves 15 selected public-benchmark comparisons, including 9B on BBH Word Sorting from 0.240 to 0.576 and 4B on HumanEvalFix JavaScript by 10.98 points.
• Online RL on Qwen3.5-4B with the verifier's outcome-only reward and a Dr. GRPO-style advantage lifts held-out reward from 0.357 to 0.857 on LAPS and from 0.286 to 0.571 on MITgcm-biogeo after 30 steps, but only once budget-truncated trajectories are masked out of the loss while kept in the group baseline — an unmasked run collapsed below its starting reward as the policy learned to shorten its turns by more than 3x.
• The evidence covers reference-verifiable repair and implementation in computational physics and geoscience rather than open-ended discovery, the SFT gains are on selected benchmarks and the RL gains on hinted tasks inside the training environments with no between-seed variance, so transfer to unseen codebases is untested, and independent audits, reward-hacking evaluations, and contamination-free task selection remain outstanding.

arXiv:2609.19134 PDF

Direct alignment methods such as DPO can suffer likelihood displacement, where the preferred response's absolute probability falls even as its margin over the dispreferred one grows. The failure is tied to low-margin preference pairs, which are usually either optimized anyway or filtered out. ComPO instead treats those pairs as one-bit comparison signals: it randomly perturbs the policy, checks whether each perturbation raises the preferred response's likelihood and lowers the dispreferred one's, and aggregates the signs into a zeroth-order update direction, with no differentiable preference loss on those pairs.

• The practical pipeline splits the data by reference-model log-likelihood margin (threshold 3), trains DPO or SimPO on the clean pairs, then runs ComPO on the low-margin pairs by perturbing only the output-layer weights with 1,600–1,800 random directions per step, normalizing the signed sum, zeroing small entries, and updating only when more than 20% of perturbations pass the comparison oracle.
• The theory covers only idealized schemes: the offline version reaches an ε-stationary point of a latent objective with a query count that grows only logarithmically in dimension for fixed gradient sparsity, and an online version that accepts or rejects steps under an exact reverse-KL constraint has its performance gap bounded by C_τ·sqrt(err) under local coverage.
• Using just 100 low-margin pairs from UltraFeedback, DPO_clean+ComPO beats full DPO on AlpacaEval 2 length-controlled win rate in all four configurations, for example 24.14 to 26.17 on Mistral-7B-Instruct and 32.59 to 35.79 on Llama-3-8B-Instruct, and it adds 0.8–2.1 points on top of tuned SimPO checkpoints (for example 60.36 to 62.42 on Gemma-2-9B-it).
• Ablations on Mistral-7B-Instruct show mean LC rising from 24.72 to 26.49 as perturbations go from 800 to 5,400, a further gain from perturbing three layers instead of one (25.02 to 26.00 LC while peak memory rises from 16.3 to 16.7 GB), the best results when about 1–6% of gradient entries are kept, and continued improvement when moving from 100 to 300 low-margin pairs.
• On the limitations side, Arena-Hard drops from 14.4 to 10.5 on Mistral-7B-Instruct (the authors point to shorter outputs, average length 468 versus 513, without establishing causality), the main tables report single-run point estimates, the convergence result assumes the oracle agrees with an unobservable smooth objective, and the practical online variant's length-normalized step damping and replay heuristics fall outside the stated guarantees.

arXiv:2609.19144 PDF