CS Paper Radar

When Is Language-Model Advice Worth Buying? A Matched-Budget Study of Predictive Gating in Reinforcement Learning

Ibne Farabi Shihab, Md Najmus Swaqeeb, Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Joyanta Jyoti Mondal

cs.LG ·submitted 2026-08-26 02:18 UTC · v2
theory

Abstract

Language-model advice can accelerate reinforcement learning, but every call has a price, and a returned suggestion may be unparsable, unsafe to execute, or no better than the agent's own action. We introduce certified predictive advice gating, a cost-aware controller that forecasts possible advisor responses, values the action actually executed after parsing, execution checks, and cache reuse, and requests advice only when a calibrated lower bound on that value covers the price. Our analysis separates the gap of executed advice, cumulative regret under an intervention-stable fallback, and myopic acquisition loss. On BabyAI GoToObj with Qwen2.5-1.5B and 7B advisors, the gate improves return over no advice by paired $0.029\pm0.016$ and $0.030\pm0.015$ while using more than $97\%$ fewer calls than always asking. Matched-budget comparisons, however, do not resolve an advantage over an early query schedule, so the gain reflects economical advice volume rather than demonstrated state selection. Groupwise calibration raises coverage on positive-value opportunities from $0.47$ to $0.85$, still below the $0.90$ target. A separate headroom-aware gate reports paired gains over an early schedule of $0.032\pm0.012$ and $0.029\pm0.011$ on GoToObj and $0.027\pm0.011$ and $0.028\pm0.012$ on GoToObjDoor, and adds BabyAI and ALFWorld success, decision-stratum coverage, and component timing. These results separate return gains, query placement, and execution screening; finite-sample deployment guarantees remain unestablished.

Comment
41 pages, 6 figures

Closest papers

Nearest neighbours by meaning across the whole index.

Measure First, Then Allocate: Calibrated Rollout Budgets for MCP Agents in Reinforcement Learning

cs.LG cs.CL ·2026-10-20 ·Hieu Tran, Zonghai Yao, Sunjae Kwon, Hong Yu
similarity 0.66
Abstract · 21 pages, 7 figures, 6 tables

Training tool-using language models requires costly interaction, yet identical rewards within a rollout group provide no group-relative learning signal. We use reward disagreement to guide harness evaluation and the allocation of rollouts for policy training in Model Context Protocol (MCP) environments. We first introduce Discord, which tests harness edits using exact paired inference and decision-preserving early stopping: only disagreeing paired outcomes favor one harness over another. This insight motivates prospective allocation in BRACE (Bank-calibrated Rollout Allocation under Correlated Episodes). BRACE combines historical task evidence, uncertainty in task success rates, and a correlated reward model to forecast mixed-reward groups before allocating generation budget. Controlled harness experiments favor Discord, while live MCP experiments expose a retention-cost trade-off. At three Qwen3-VL model scales, BRACE exceeds the strongest plotted baseline peaks with fewer new training rollouts; historical-bank collection is counted separately. It also leads the compared methods on BFCL and NESTFUL without further training. Ablations probe the bank, estimator, and allocation score; prospective evaluation checks forecast calibration. These results support forecasting reward contrast before generation.

0.17

Confidence-Calibrated Inference Expansion for Evaluator-Guided Test-Time Reasoning

cs.AI cs.CLcs.LG ·2026-10-14 ·Weida Liang, Kenji Kawaguchi
top venue similarity 0.66
Abstract · Accepted at NeurIPS 2026

Evaluator-guided reasoning presents a local control problem at every step: trust the current answer, return a fallback, or spend more inference. We formulate this choice as a compute-aware three-action controller, scoring terminal actions by predicted answer success and expansion by a rollout-based continuation-success surrogate, with compute cost priced explicitly. A local action-gap analysis characterizes when score errors can change a decision, while an expected regret bound accounts for errors along the realized expansion path. Guided by this analysis, we learn and calibrate action-specific confidence estimates, with local uncertainty penalties for ambiguous decisions. On held-out width expansion and revision tasks, the controller improves observed utility over calibration-selected fixed policies while reducing normalized online compute by about 30% and 16%, respectively. Across four Qwen3-8B seeds, compute is lower than fitted Q in both settings, while utility differences are small and mixed. The results frame evaluator-guided reasoning as resource allocation under uncertainty, rather than a uniformly stronger reasoner.

0.60

Can Language Models Learn to Forecast Stock Prices

cs.CL ·2026-09-29 ·Jiacheng Guo, Suozhi Huang, Shuzhen Li, Yunlong Gao, Zerui Cheng +9 more
benchmark similarity 0.64
Abstract · 18 pages, 4 figures

Post-training has been shown to significantly improve language models' performance on tasks with verifiable outcomes, including mathematical reasoning, software engineering, and computer use. However, whether the same approach can improve forecasting in financial markets is much less clear. Compared with tasks with verifiable outcomes, not only are realized returns noisy, but even what constitutes a relevant information set for making effective predictions is not obvious a priori: the model must decide which observations to gather and then commit to a numerical judgment before the outcome is known. We study this question in a chronological stock-price sandbox, where a language model gathers price, volume, relative-performance, and market-context evidence and predicts a future return. We post-train Qwen3-4B with supervised fine-tuning (SFT) on tool-use demonstrations, then proximal policy optimization (PPO) with a terminal reward given by the forecast score against the realized return. The resulting AURA-4B more than doubles the starting direction--magnitude score, from 20.94 to 43.31, and is comparable to frontier language models on this benchmark. Conditional magnitude agreement rises from 33.3 to 66.2, while directional accuracy changes from 62.9 to 65.4. SFT expands tool use, and PPO further increases the share of ranking and market-context queries. These results show that post-training can substantially improve financial forecasting performance, together with changes in how the model investigates the market, on this outcome-selected benchmark.

0.27

Do Your Own Research: Learning to Forecast by Learning to Search

cs.LG ·2026-10-01 ·Yusuf Afifi, Artur Kiulian, Anton Polishko, Mykola Khandoga, Hamudi Naanaa +1 more
publisheddatasetcode/data similarity 0.64
Abstract · Accepted at the NeurIPS 2026 Workshop on Foundation Models for Temporal Systems (FMTS). 9 pages, 4 figures. Code and data: https://github.co…

Outcome-based reinforcement learning can train language models to forecast real-world events, but prior forecasting work either freezes research context before training or deploys agentic research only at test time, so the skill of gathering evidence is never shaped by the reward. We introduce an agentic forecasting environment, dataset, and harness built from 2,100+ resolved Polymarket questions; the agent acquires its own context at rollout time (web search, page reading, and financial time series, all restricted by layered leak filtering to information published before each question's cutoff), and we train Qwen3.5-35B-A3B (3B active parameters) on it with single-epoch GRPO under a Brier-score reward. Training changes how the agent interacts with information: calibration improves 30-40%, and search attempts fall from 3.8 to 2.25 per rollout as evidence discipline is learned. Evaluated in an identical harness against four frontier models, the trained policy also finishes ahead of every frontier model tested at evidence-based forecasting, including Claude Opus 4.5 (soft-Brier 0.254 vs. 0.256, n=265), at about 5% of the inference cost, and its margin is widest on the hardest questions, the ones the crowd itself had not decided. We release the environment, dataset, and per-rollout records as a reusable harness for temporal forecasting agents.

0.49

Better Behavioral Prediction, More Faithful Model Ablations? Evidence from Sequential Choice

cs.CL q-bio.NC ·2026-09-28 ·Hanbo Xie
hot direction similarity 0.63
Abstract

Using predictive models to explain cognition requires more than accurate behavioral predictions. Input ablations offer an appealing route: remove information from a model and interpret the resulting performance change as evidence of its importance for behavior. Yet this inference assumes that the model's dependence on information reflects the dependence of the process generating the behavior. We test it in two synthetic sequential bandit tasks with known generating policies, where past choices can remain informative when feedback is unavailable to a predictor. We compare GRUs and Transformers trained from scratch, a fine-tuned LLaMA model, and cognitive models across systematically varied reward contributions. Our analyses distinguish prediction after training without reward observations from the response of a fixed predictor to donor-reward replacement. Three findings emerge. First, in the restless task, neural models trained without rewards predict held-out choices better than four simple training-fitted behavioral baselines. Second, under matched donor replacement, accurate predictors can respond much less than the known generator. Third, at some reward weights, neural networks predict better than a pooled reinforcement-learning model but have less faithful changes in choice probabilities; the model ordering differs between the two tasks. These independent-test results separate information sufficient for prediction from response fidelity under a specified ablation in sequential choice. They motivate validating model-ablation responses independently of predictive performance before using them to infer how the observed behavior was generated.

0.33

A Query Is Not a Commitment: Learning to Correct Expert Answers in Online Deferral

cs.LG stat.ML ·2026-10-05 ·Yannis Montreuil, Axel Carlier, Lai Xing Ng, Wei Tsang Ooi
similarity 0.62
Abstract

An inaccurate expert can still provide useful information after correction. We study online learning to defer in which the learner chooses an expert and fixes a correction function before purchasing its answer, then applies that function to the answer received. The difficulty is that observed losses reflect both expert quality and an unfinished correction: early errors can discourage queries that would be valuable after learning. We propose ORUCB, which pools shared and expert-specific polynomial responses. A bound on cumulative response-learning error calibrates confidence-weighted risk regression and exploration, allowing the router to account for this error when deciding which answers to buy. Under bounded residuals and disagreements, a fixed feasible model of optimal responses, and linear models of free and optimal queried risk, the calibrated algorithm achieves high-probability pseudo-regret $O(\sqrt T\log(T+1))$ over $T$ rounds for fixed problem parameters. The guarantee permits singular answer distributions and misspecified shared responses; optimality is relative to the bounded response class. On four test streams, the selected cubic policy has lower fee-inclusive cost than seven baselines that deploy answers unchanged. Comparisons with a common correction learner examine routing, while six-price comparisons measure cost and query rates.

0.31

Good Judgment, Bad Advice: Decision-Model Controllers Raise Agent Cost Without Improving Success

cs.AI cs.LG ·2026-10-17 ·Naman Vats
code/data similarity 0.62
Abstract · 8 pages, 2 figures, 4 tables. Code: https://github.com/namanvats/decision-model-controllers

Agent scaffolds increasingly delegate small decisions (whether to stop, verify, or change approach) to cheaper models. We ask whether a typed decision model, Jev 1.13, consulted only at deterministic decision points, makes a Claude Sonnet 5.5 terminal agent cheaper without making it worse. In a pre-specified study on Terminal-Bench 2.1, we hold one agent loop fixed and vary only the controller: none, deterministic rules, a small LLM (Claude Haiku 5.5) answering the same questions, Jev, and a placebo that samples Jev's development-phase choice frequencies at the same moments. On 21 held-out tasks (315 trials), Jev raised cost per attempt by 19% (geometric-mean ratio 1.19, 95% CI 1.02-1.36) at an identical pass rate (73.0% vs. 73.0%). Jev was cheaper than the small-LLM controller (ratio 0.81) but 22% more expensive than the placebo, and no pass-rate difference between arms was detectable. Offline, Jev ranked finish requests better than the alternatives (AUROC 0.78 vs. 0.62 for rules and 0.60 for the small LLM) but was poorly calibrated: its mean probability that a run was complete was 0.50 when 87% of those runs passed, so it asked the agent to verify work that was already correct. No controller ever chose to abandon a run. Informative decision signals did not become useful interventions; we discuss what that would require.

0.53

Best Practice Critic Optimization

cs.LG cs.AIcs.CL ·2026-08-24 ·Penghui Qi, Xiangxin Zhou, Wee Sun Lee
code/datahighly cited similarity 0.61
Abstract

Group-based reinforcement learning methods such as GRPO for large language models avoid training a critic by sampling multiple responses for each prompt. A reliable critic could instead estimate token-level advantages from one response, but standard critic-based training recipes are often unstable. We study this instability and develop **Best Practice Critic Optimization (BPCO)**, a recipe that combines DPPO, value predictions bounded to the reward range, Monte Carlo value targets, unnormalized policy advantages, and length-adaptive generalized advantage estimation. Because the critic is used only during training, BPCO can also condition it on reward-defining information, such as a reference answer or grading rubric, that is hidden from the policy. Controlled experiments isolate the effect of each design choice. Across mathematical reasoning tasks with models ranging from 1.5B parameters to 30B-A3B mixtures of experts, BPCO improves a strong critic-based baseline consistently, and matches or exceeds a group-based baseline while sampling one response per prompt. The same recipe also improves learning with rubric-based rewards. These results show that a carefully designed critic provides a reliable alternative to group-relative advantage estimation. Code is available at https://github.com/QPHutu/golden_critic.

0.49