R2-LLMs: Adaptive Reasoning for LLMs

• R2-LLMs are an emerging family of models that integrate explicit reasoning over inference parameters such as output budget, token routing, and memory edits.
• They employ adaptive control policies to dynamically balance quality and cost, achieving notable efficiency improvements across diverse tasks.
• Implementations like R2-Router and R2R demonstrate significant gains in speed and accuracy by optimizing compute allocation and retrieval strategies.

Searching arXiv for papers on R2-LLMs and closely related usages of the “R2” label in LLM research. Searching arXiv for "R2-Router routing as reasoning". R2-LLMs is an emerging umbrella label for a cluster of LLM systems that replace fixed, one-shot model invocation with explicit reasoning over controllable structure at inference or training time. In current usage, the label spans budget-aware query routing, token-level small–large model routing, redundancy-reduced diffusion decoding, ranking-driven retrieval over heterogeneous memories for recommendation, fair reinforcement learning for long-horizon memory agents, and LLM-driven robot-to-everything coordination. Across these works, the shared pattern is not a single architecture or benchmark, but a shift from static selection toward adaptive control over model choice, output budget, retrieval actions, memory edits, or external sensing (Xue et al., 2 Feb 2026, Fu et al., 27 May 2025, Du et al., 21 Apr 2026, Li et al., 8 May 2026, Yan et al., 20 May 2026, Yang et al., 20 Mar 2026).

1. Scope of the term and its current usage

The literature does not present a single canonical expansion of “R2-LLMs.” Instead, papers use closely related “R2” formulations in domain-specific ways: “R2-Router” for routing with reasoning, “Roads to Rome” for token routing, “$R^2$-dLLM” for redundancy-reduced diffusion decoding, “Memory-R2” for fair long-horizon memory training, and “IndoorR2X” for robot-to-everything coordination. This suggests that “R2-LLMs” presently functions as a family-resemblance term for methods that add an extra reasoning or control layer on top of standard LLM inference rather than as a closed formal taxonomy.

System	Core controllable structure	Reported effect
R2-Router	Joint selection of model and output length budget	State-of-the-art routing at 4–5x lower cost
R2R	Token-level routing between SLM and LLM	2.8x wall-clock speedup over R1-32B
$R^2$-dLLM	Spatial and temporal redundancy in diffusion decoding	Up to 75% reduction in decoding steps
RRCM	Adaptive retrieval over collaborative and meta memories	Best results across all datasets and metrics
Memory-R2	Local and global credit assignment for memory edits	+7.46 F1 over Memory-R1
IndoorR2X	IoT-augmented semantic world modeling for coordination	92% success with lower steps, path length, and token usage

A unifying interpretation is that these systems do not treat the LLM as an isolated predictor. They instead expose latent operating variables—budget, token criticality, retrieval choice, memory state, or world state—and train a policy, router, or planner to reason over them. In that sense, the “R2” motif consistently denotes a movement from reactive generation toward structured decision-making.

2. Budget-aware routing and “routing as reasoning”

The clearest formalization of R2-LLMs appears in "R2-Router: A New Paradigm for LLM Routing with Reasoning" (Xue et al., 2 Feb 2026). Standard query-level routing selects an LLM $M_j$ for a query $x_i$ by balancing quality and cost with

$S=(1-\lambda)\cdot Q-\lambda \cdot C,$

where $\lambda$ is the user’s cost-sensitivity parameter. The paper argues that prior routers are built on a simplifying assumption: each LLM is assigned one fixed $(Q,C)$ pair per query. The proposed correction is to elevate output length budget $b$ to an explicit inference-time control variable and to optimize

$$(M^*, b^*) = \arg\max_{M \in \mathcal{M},\, b \in \mathcal{B}} \left((1-\lambda)\cdot Q(x,M,b)-\lambda\cdot C(b)\right).$$

This reformulation changes the search space. A reactive router searches only over fixed operating points,

$$\mathcal{S}_{reactive} = \{(M_i,\hat c_i)\mid M_i\in\mathcal{M}\},$$

whereas a reasoning-based router searches over the Cartesian product of models and budgets,

$R^2$0

The paper proves the dominance relation

$R^2$1

provided the reactive router’s predicted cost corresponds to a feasible budget. The central criticism of prior routing is stated succinctly: reactive routers reason over points, not curves.

Methodologically, R2-Router couples a quality-cost predictor with a decision maker. A shared encoder maps the query to $R^2$2. For each model $R^2$3, a set of $R^2$4 independent heads $R^2$5 predicts budget-conditioned quality at anchor budgets $R^2$6,

$R^2$7

with mean squared error supervision. To avoid a large number of heads, the method uses piecewise linear interpolation between anchors. After selecting $R^2$8, the system enforces the choice with a length-constrained instruction such as “Use at most $R^2$9 tokens.”

The benchmark contribution is equally important. R2-Bench is described as the first routing dataset that records how the same model behaves under multiple output length budgets for the same query. It contains 30,968 queries from GPQA, MuSR, MMLU-Pro, MATH, OpenHermes, and RAGBench, with responses collected at budgets including 10, 20, 30, 40, 50, 80, 100, 150, 200, 300, 500, 800, 1200, 2000, 4000, and default. Because open-ended answers are difficult to score automatically, the paper adopts an LLM-as-a-judge protocol and selects Qwen3-80B-Instruct after validation against human annotations from 30 experts, reporting Pearson correlation $M_j$0.

Empirically, R2-Bench raises the routing oracle upper bound relative to single-response datasets such as SPROUT, with AUDC improving from 0.85 to 0.98, QNC dropping from 0.18 to 0.04, and Peak Quality rising from 0.90 to 0.98. R2-Router then achieves state-of-the-art routing performance at 4–5x lower cost than prior routers, and in dynamic-pool integration with UniRouter it improves AUDC from 0.590 to 0.623 while reducing QNC by 80%. The paper also reports 20–30 minute training on a single GPU with negligible routing latency relative to generation. Within the R2-LLM landscape, this work establishes the archetype: the router reasons over inference-time degrees of freedom rather than merely selecting among fixed models.

3. Fine-grained compute allocation: token routing and redundancy reduction

A second major branch of R2-LLMs relocates the control problem from the query level to the token or decoding-step level. "R2R: Efficiently Navigating Divergent Reasoning Paths with Small-Large Model Token Routing" (Fu et al., 27 May 2025) studies greedy autoregressive decoding with a routing function

$M_j$1

where $M_j$2 chooses between an SLM and an LLM for token $M_j$3. The key empirical observation is that only a small fraction of tokens genuinely alter the reasoning path. Across 2,094 queries and 7.6M generated tokens, the 1.5B SLM differed from the 32B LLM on about 11% of tokens, while the divergent-token rate was 4.9%. By domain, divergent-token rates were 2.8% for math, 4.7% for code, and 9.7% for QA.

R2R labels token differences as identical, neutral, or divergent by continuing from alternative next-token choices and using Qwen2.5-72B as a verifier. A lightweight FFN router of around 56M parameters consumes top-100 SLM logits, predicted-token embeddings, and last-layer hidden states, and predicts whether the current token is divergent. Inference is online and immediate: the SLM proposes a token, the router inspects the local state, and the LLM is called only when divergence probability exceeds a threshold. In the reported DeepSeek-R1 setup, the method achieves about 5.6B average activated parameters per token, reaches about 92% of R1-32B accuracy while using only 11–15% of tokens from the LLM, surpasses the average accuracy of R1-7B by 1.6x, outperforms R1-14B, and delivers a 2.8x wall-clock speedup over R1-32B.

"$M_j$4-dLLM: Accelerating Diffusion LLMs via Spatio-Temporal Redundancy Reduction" (Du et al., 21 Apr 2026) addresses an analogous problem in diffusion LLMs, where latency is dominated by high NFE rather than autoregressive token routing. The paper identifies two forms of redundancy. Spatial redundancy arises from confidence clusters and positional ambiguity; temporal redundancy arises when stable tokens are repeatedly remasked and rediscovered. The training-free inference rules therefore aggregate local confidence, resolve token clusters by the most confident position, and finalize tokens that remain unchanged for $M_j$5 consecutive steps when confidence exceeds a threshold. Redundancy-aware supervised fine-tuning then aligns the model with low-redundancy trajectories.

The reported efficiency gains are substantial. The abstract states up to 75% reduction in decoding steps relative to existing strategies. For LLaDA-Instruct-8B on GSM8K, latency drops from $M_j$6 to $M_j$7, NFE from $M_j$8 to $M_j$9, and accuracy changes from $x_i$0 to $x_i$1. On MBPP, latency drops from $x_i$2 to $x_i$3, NFE from $x_i$4 to $x_i$5, and accuracy rises from $x_i$6 to $x_i$7. In this branch of R2-LLMs, the central idea is again adaptive control: compute is concentrated on tokens or steps that matter for trajectory fidelity rather than distributed uniformly across the generation process.

4. Retrieval-and-reasoning over heterogeneous memories

In recommendation, R2-LLMs appear as adaptive context-construction systems. "RRCM: Ranking-Driven Retrieval over Collaborative and Meta Memories for LLM Recommendation" (Li et al., 8 May 2026) starts from the claim that current LLM recommenders often use fixed prompt templates, static retrieval pipelines, or handcrafted collaborative-filtering injections. RRCM instead begins from a lightweight user-history context,

$x_i$8

and learns whether to recommend directly, retrieve collaborative evidence, retrieve item metadata, or interleave both through reasoning.

The framework constructs a unified memory corpus $x_i$9 with two textual memory types. Collaborative memory stores anonymized user histories; meta memory stores item-side metadata such as author, genres, director, price, brand, or categories. Both are accessed through a single natural-language retrieval interface $S=(1-\lambda)\cdot Q-\lambda \cdot C,$0. The policy $S=(1-\lambda)\cdot Q-\lambda \cdot C,$1 operates in a multi-turn loop: reason over current evidence, either retrieve and append a tool response to the state or terminate, then output a free-text recommendation title. The generated title is grounded back to the catalog by sentence-encoder retrieval using cosine similarity.

The reward is ranking-driven rather than retrieval-driven: $S=(1-\lambda)\cdot Q-\lambda \cdot C,$2 with $S=(1-\lambda)\cdot Q-\lambda \cdot C,$3, $S=(1-\lambda)\cdot Q-\lambda \cdot C,$4, and $S=(1-\lambda)\cdot Q-\lambda \cdot C,$5. Optimization uses Group Relative Policy Optimization with reward standardization within sampled groups. The experimental setup uses Goodreads, MovieLens, and Amazon CDs & Vinyl; Qwen3-1.7B is the default backbone; retrieval uses Flat E5; tool-calling uses SGLang; and the maximum number of turns is 5.

The reported results show that RRCM is best across all datasets and metrics. On Goodreads, HR@5 is 0.0223 and HR@10 is 0.0340, with improvements over the prior best around 9–16% depending on metric. On MovieLens, HR@5 is 0.0130 and HR@10 is 0.0188. On Amazon CDs & Vinyl, HR@5 is 0.0102 and HR@10 is 0.0129. Ablations show that all three components matter: collaborative memory, meta memory, and explicit reasoning. Within R2-LLMs, this work exemplifies a shift from static retrieval augmentation to decision-aware retrieval policies optimized directly for the downstream ranking objective.

5. Fair credit assignment for long-horizon memory agents

A more explicitly reinforcement-learning-oriented branch is represented by "Memory-R2: Fair Credit Assignment for Long-Horizon Memory-Augmented LLM Agents" (Yan et al., 20 May 2026). The problem is that, in multi-session environments, memory makes the agent’s own past actions part of its future environment. Once different rollouts have written, updated, or deleted different memories, later trajectory-level comparisons no longer occur under the same effective state. The paper argues that this violates a key assumption behind group-relative methods such as GRPO and makes trajectory-level reward unfair for memory operations.

Memory-R2 addresses this with LoGo-GRPO, combining global and local group-relative optimization. Globally, the agent is still trained end-to-end on long-horizon rewards derived from the final memory state. Locally, the algorithm samples selected sessions, retrieves a cached pre-session memory state, and rerolls that session multiple times from the same anchor memory. This makes comparisons fairer because local rerollouts differ primarily in the current session’s memory operations rather than in inherited memory corruption.

The agent factorizes memory construction into fact extraction and memory management: $S=(1-\lambda)\cdot Q-\lambda \cdot C,$6 with shared parameters across the two roles via role-specific prompts. Rewards combine QA performance with a compression penalty so that memory is both useful and compact. To avoid bias toward long generations, the paper introduces step-level length normalization for importance ratios and advantages. Training is stabilized with a progressive curriculum from 8 to 16 to 32 sessions.

The reported gains are large. In the backbone-controlled LoCoMo setting, Memory-R2 achieves F1 $S=(1-\lambda)\cdot Q-\lambda \cdot C,$7, BLEU-1 $S=(1-\lambda)\cdot Q-\lambda \cdot C,$8, and judge score $S=(1-\lambda)\cdot Q-\lambda \cdot C,$9, compared with Memory-R1 at F1 $\lambda$0, BLEU-1 $\lambda$1, and $\lambda$2. Replacing LoGo-GRPO with standard GRPO lowers F1 from 49.67 to 46.62 and raises M-Fail from 6.72 to 10.20. Removing the curriculum causes F1 to collapse to 24.12 and M-Fail to rise to 46.50. In the R2-LLM context, the significance is that “reasoning” is extended from inference-time deliberation to reinforcement-learning credit assignment over persistent state.

6. Embodied coordination, misconceptions, and broader significance

The embodied extension of this line appears in "IndoorR2X: Indoor Robot-to-Everything Coordination with LLM-Driven Planning" (Yang et al., 20 Mar 2026). The benchmark and simulation framework is presented as the first indoor benchmark for LLM-driven multi-robot planning with robot-to-everything perception and communication. It uses 85 multi-room indoor environments built on AI2-THOR—10 ArchitecTHOR homes and 75 RoboTHOR modular apartments—and augments robot observations with static IoT devices, mainly CCTV-like cameras, configured so that roughly 50% of the house area is covered. A central coordination hub fuses robot observations, IoT observations, and prior state into a global semantic state

$\lambda$3

which an LLM planner uses to produce dependency-structured multi-agent plans.

The quantitative comparison clarifies the effect of richer shared state. With GPT-4.1 and three robots, isolated robots achieve 66% success, 186 action steps, 137 m path length, and 54,572 LLM tokens; R2R reaches 92% success, 116 action steps, 99 m path length, and 47,875 tokens; R2X preserves 92% success while lowering cost further to 108 action steps, 88 m path length, and 42,438 tokens. The paper also shows sensitivity to stale or incorrect IoT information: 5-step and 10-step delays reduce success to 90% and 87%, and incorrect semantic status is more harmful than missing detections.

Several misconceptions are corrected by this body of work. First, these systems are not reducible to prompting the model to be concise. R2-Router explicitly reports that concise-output prompts given to reactive baselines do not recover the gains of joint model–budget reasoning. Second, they are not equivalent to static retrieval augmentation; RRCM is trained to retrieve only when retrieval improves final top-$\lambda$4 recommendation quality. Third, token-level routing is not merely speculative decoding; R2R explicitly allows neutral token differences rather than requiring exact equality. Fourth, the label “R2” should not be conflated with the statistical coefficient $\lambda$5, which denotes the multiple correlation coefficient in an unrelated asymptotic inference literature (Dedecker et al., 2024).

Taken together, these works indicate that R2-LLMs are best understood as systems-level LLM designs in which reasoning is delegated not only to the base model’s latent chain of thought, but also to explicit control layers over compute, budget, retrieval, memory, or world-state acquisition. A plausible implication is that the term will remain useful so long as LLM deployment problems continue to be governed by hidden operating curves—quality versus cost, retrieval versus context budget, memory utility versus compression, or observability versus coordination overhead—rather than by fixed model comparisons alone.