InterleaveThinker: Multi-Agent Generation Framework

• InterleaveThinker is a multi-agent framework for interleaved text-image generation, decoupling global planning from step-wise local corrections.
• It integrates a planner, a black-box generator, and a critic to iteratively refine outputs, effectively mitigating issues like visual over-reliance and error accumulation.
• Its modular design enhances existing image generators and achieves strong benchmark performance, advancing multimodal reasoning systems.

to=arxiv_search.search สำนักเลขานุการองค์กร 天天中彩票nbajson {"^^^^^^^^3query3 OR all:\3^^^^^^^^"InterleaveThinker\"","max_results":^^^^3ti:\3query3 to=arxiv_search.search 񎔄json {"^^^^^^^^3query3 OR id:(&&&^^^^3ti:\3^^^^&&&) OR id:(&&&^^^^3 OR all:\3^^^^&&&) OR id:(Chu et al., 2014) OR id:(Chiu, 2019)","max_results":^^^^3ti:\3query3^^^^,"sort_by":"relevance"} InterleaveThinker is a multi-agent framework for interleaved generation, defined as the production of a coherent sequence of alternating text instructions and images, and introduced as a way to endow a frozen image generator with capabilities it does not natively possess (&&&^^^^3query3^^^^&&&). Its central design choice is to decouple global planning from local execution evaluation and correction: a planner agent organizes the required steps of an image–text trajectory, a black-box generator executes those steps, and a critic agent judges each intermediate result and refines the prompt when the output deviates from the plan (&&&^^^^3query3^^^^&&&). In adjacent lines of work, closely related systems interleave latent reasoning with text, speech, or streaming perception rather than executing all computation before any outward action, making InterleaveThinker part of a broader technical motif centered on staged, alternating inference (&&&^^^^3ti:\3^^^^&&&).

^^^^3ti:\3^^^^. Problem formulation and task scope

Interleaved generation refers to a setting in which the input and output are not a single prompt–image pair but an alternating text/image sequence. The motivating examples include step-by-step visual narratives, guided image editing, and embodied manipulation, all of which require a system to maintain coherence across a multi-step trajectory rather than optimize a single terminal image (&&&^^^^3query3^^^^&&&).

The framework is positioned against two limitations in existing systems. First, single-image generators, including models that support editing, are architecturally constrained to produce exactly one image given one text prompt, optionally with one image to edit. Second, Unified Multimodal Models such as JanusPro, Emu3.5, or Nano Banana Pro allow interleaved text/image tokens in principle but are reported to suffer from visual over-reliance and step-wise error accumulation in practice (&&&^^^^3query3^^^^&&&). The former denotes premature fixation on an intermediate image that appears close to the goal; the latter denotes the compounding of small early errors across subsequent steps.

InterleaveThinker addresses these failure modes by separating long-horizon structure from local correction. The planner specifies what should happen at each step, while the critic evaluates whether the generator actually achieved it and, if not, revises the prompt for regeneration. The paper states that the planner never conditions on intermediate images during planning, and that erroneous steps can be corrected immediately rather than propagated forward (&&&^^^^3query3^^^^&&&). This makes the framework a wrapper around any existing image generator rather than a replacement for the base generator itself.

^^^^3 OR all:\3^^^^. Planner–generator–critic architecture

At inference time, the system consists of three modules in a loop. The Planner Agent takes the user’s interleaved text/image sequence PRESERVED_PLACEHOLDER_^^^^3query3^^^^ and emits an PRESERVED_PLACEHOLDER_^^^^3ti:\3^^^^-step execution plan

PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^

where $u_i$ is a human-readable step description, $p_i$ is the initial model-facing prompt for the generator, and $a_i$ is auxiliary text such as a caption or knowledge snippet (&&&^^^^3query3^^^^&&&).

The Generator is treated as a black box. At step $i$ and iteration $t$, it consumes the current refined prompt $r_i^t$ and the previous image $I_{i-1}$, or a blank canvas when PRESERVED_PLACEHOLDER_^^^^3ti:\3query3^^^^, and produces PRESERVED_PLACEHOLDER_^^^^3ti:\3ti:\3^^^^ (&&&^^^^3query3^^^^&&&). This choice is significant because the framework does not require architectural modification of the generator.

The Critic Agent receives PRESERVED_PLACEHOLDER_^^^^3ti:\3 OR all:\3^^^^ and returns three objects: a binary judgment PRESERVED_PLACEHOLDER_^^^^3ti:\33^^^^, a next refined prompt PRESERVED_PLACEHOLDER_^^^^3ti:\3^^^^4, and a textual reasoning trace PRESERVED_PLACEHOLDER_^^^^3ti:\3^^^^5 (&&&^^^^3query3^^^^&&&). If PRESERVED_PLACEHOLDER_^^^^3ti:\3^^^^6 and PRESERVED_PLACEHOLDER_^^^^3ti:\3^^^^7, the refined prompt is fed back to the generator; otherwise the image is accepted and the system advances to the next step.

Both planner and critic are implemented with Qwen3-VL-8B-Instruct, first under supervised cold-start and then, for the critic, under RL (&&&^^^^3query3^^^^&&&). The modular split is empirically important. The ablations report that a One-Agent variant that merges planning and criticism reaches an average of 54.5 on UEval, compared with 66.3 for the full system, and the paper attributes the degradation to visual over-reliance and myopia (&&&^^^^3query3^^^^&&&).

3. Data construction, cold-start, and single-step RL

The training pipeline is built around three specialized datasets.

Dataset	Role	Construction signal
Interleave-Planner-SFT-8^^^^3query3^^^^k	Planner cold-start	Interleaved text↔plan pairs
Interleave-Critic-SFT-^^^^3ti:\3ti:\3 OR all:\3^^^^k	Critic supervised training	Step-wise refinement episodes
Interleave-Critic-RL-^^^^3ti:\33^^^^k	Critic RL	Single-step rollouts

The Interleave-Planner-SFT-8^^^^3query3^^^^k corpus begins from procedurally generated ~4^^^^3query3^^^^K diverse “How-to” textual prompts spanning eight top-level categories and ~75 subcategories, expanded into vocabulary banks with Gemini ^^^^3 OR all:\3^^^^.5 Pro and templated into ~^^^^3ti:\3query3query3^^^^ instructional patterns (&&&^^^^3query3^^^^&&&). Full trajectories are synthesized with Gemini ^^^^3 OR all:\3^^^^.5 Pro and Nano Banana Pro, then interleaved with ground-truth final images and randomly truncated so that the planner predicts the remaining text steps. The format is a JSON list containing step number, step name, instruction PRESERVED_PLACEHOLDER_^^^^3ti:\3^^^^8, prompt PRESERVED_PLACEHOLDER_^^^^3ti:\3^^^^9, and auxiliary text PRESERVED_PLACEHOLDER_^^^^3 OR all:\3query3^^^^ (&&&^^^^3query3^^^^&&&).

The Interleave-Critic-SFT-^^^^3ti:\3ti:\3 OR all:\3^^^^k dataset decomposes synthesized trajectories into independent step-wise refinement episodes of the form

PRESERVED_PLACEHOLDER_^^^^3 OR all:\3ti:\3^^^^

Gemini ^^^^3 OR all:\3^^^^.5 Pro with an adapted ViEScore prompt assigns semantic-alignment and visual-quality scores from ^^^^3query3^^^^ to ^^^^3ti:\3query3^^^^; unstable or negative refinement trends are filtered out, stable upward-quality steps are retained, and high-variance versus low-variance subsets are split at a ^^^^3 OR all:\3^^^^:^^^^3ti:\3^^^^ ratio for RL versus SFT (&&&^^^^3query3^^^^&&&). The critic’s textual format is explicitly > …reasoning…</think> followed by <answer>[Judgment] [Refined Prompt]</answer>.

The Interleave-Critic-RL-^^^^3ti:\33^^^^k stage exists because a full interleaved trajectory may require ^^^^3 OR all:\3^^^^5+ generator calls, making end-to-end trajectory-level RL computationally impractical (&&&^^^^3query3^^^^&&&). Instead, the paper optimizes only the critic’s single-step behavior with GRPO under a KL-regularized objective. The reward design has three components: an accuracy reward based on agreement with the ground-truth binary judgment, a step-wise reward equal to the Gemini score improvement from PRESERVED_PLACEHOLDER_^^^^3 OR all:\3 OR all:\3^^^^ to PRESERVED_PLACEHOLDER_^^^^3 OR all:\33^^^^, and a format reward for well-formed JSON (&&&^^^^3query3^^^^&&&). The combined reward is

PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^4

A recurring misconception is that the framework performs global trajectory RL; in fact, the paper’s stated strategy is to use a frozen planner and optimize only local critic corrections.

4. Empirical performance and ablations

On UEval—described as measuring eight diverse tasks—InterleaveThinker paired with FLUX.^^^^3 OR all:\3^^^^-klein-9B reaches an average of 66.3, compared with 66.^^^^3query3^^^^ for Nano Banana, 65.^^^^3 OR all:\3^^^^ for GPT-5-Instant, and 49.^^^^3ti:\3^^^^ for Emu3.5 (&&&^^^^3query3^^^^&&&). The reported interpretation is that the wrapper can raise an open generator to performance comparable to frontier proprietary systems on interleaved generation benchmarks.

On CoMM Task 3/4, the same pairing scores 9.3/9.6 on style consistency, 9.^^^^3 OR all:\3^^^^/9.6 on entity consistency, 9.^^^^3ti:\3^^^^/9.5 on trend alignment, 9.^^^^3ti:\3^^^^/9.6 on completeness, 9.7/9.8 on image quality, and 5.^^^^3 OR all:\3^^^^/8.^^^^3 OR all:\3^^^^ on text-image alignment for interleaved versus pure-text input settings (&&&^^^^3query3^^^^&&&). The alignment number is notably lower than the image-quality figures, indicating that not all aspects of the task are equally saturated.

The framework also improves reasoning-based benchmarks. On WISE, FLUX.^^^^3 OR all:\3^^^^-klein-9B alone scores ^^^^3query3^^^^.47 overall, while +InterleaveThinker reaches ^^^^3query3^^^^.73; Qwen-Image-Edit moves from ^^^^3query3^^^^.6^^^^3query3 to ^^^^3query3^^^^.7^^^^3 OR all:\3^^^^ (&&&^^^^3query3^^^^&&&). On RISE, FLUX.^^^^3 OR all:\3^^^^-klein-9B moves from ^^^^3ti:\33^^^^.3 to ^^^^3 OR all:\3^^^^8.9, and Qwen-Image-Edit from ^^^^3ti:\3^^^^9.4 to 3^^^^3query3^^^^.^^^^3query3 (&&&^^^^3query3^^^^&&&). The paper explicitly notes that these gains occur without explicit reasoning training for the planner or critic.

The ablations isolate the contribution of each component. Using FLUX.^^^^3 OR all:\3^^^^-klein-9B on UEval, the average rises from ^^^^3ti:\3^^^^8.^^^^3 OR all:\3^^^^ for the baseline generator alone, to 48.^^^^3ti:\3^^^^ with zero-shot Qwen3-VL planner+critic, to 6^^^^3query3^^^^.5 with Planner-SFT, to 64.5 with Full-SFT, and finally to 66.3 with Full-RL (&&&^^^^3query3^^^^&&&). Removing either RL reward is damaging: RL w/o step reward gives 65.^^^^3 OR all:\3^^^^, and w/o accuracy reward gives 65.^^^^3ti:\3^^^^. Increasing the refinement budget also matters: PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^5 yields 6^^^^3query3^^^^.^^^^3 OR all:\3^^^^, PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^6 yields 65.3, and PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^7 yields 66.3 (&&&^^^^3query3^^^^&&&). The paper’s own reading is that the planner is essential for bypassing visual over-reliance, the supervised critic improves image quality, and the dual-reward RL gives the final increment.

5. Related interleaved reasoning systems

InterleaveThinker belongs to a wider family of systems that alternate hidden computation and outward generation rather than enforcing a strict think-then-act schedule.

System	Interleaving object	Representative result
SpiralThinker	Latent vectors and text	56.56 / 99.4^^^^3query3^^^^ / 63.3^^^^3 OR all:\3^^^^ accuracy
AsyncReasoning	Prompt, thinker, writer streams	TTFT PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^8 s; 6–^^^^3ti:\3ti:\3^^^^× lower delay
SHANKS	Hearing and unspoken reasoning	37.^^^^3ti:\3^^^^ pp gain in valid-interrupt accuracy
InterRS	Reasoning and spoken answer segments	49.33% avg, instant response

SpiralThinker performs iterative updates over latent representations PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^9 interleaved with textual reasoning markers such as <bol>, <eol>, <bot>, and <eot>, and anchors the latent trajectory to an explicit-reasoning model through a progressively weighted alignment loss (&&&^^^^3 OR all:\3^^^^7&&&). Its reported accuracy is 56.56 on GSM8K-Aug, 99.4^^^^3query3^^^^ on ProsQA, and 63.3^^^^3 OR all:\3^^^^ on StrategyQA, with ablations showing that alignment alone helps, iteration alone can be unstable, and both together are necessary (&&&^^^^3 OR all:\3^^^^7&&&).

AsyncReasoning is training-free and uses RoPE ^^^^^^^^3query3 rotation to maintain three logical streams—Prompt, Thinker, and Writer—inside one shared KV cache, allowing a model to think and write concurrently without re-encoding the cache (&&&^^^^3ti:\3^^^^&&&). On Qwen3-3^^^^3 OR all:\3^^^^B, the MATH setup reduces TTFT from 59^^^^3 OR all:\3^^^^ s for synchronous thinking to ^^^^3 OR all:\3^^^^.49 s for AsyncReasoning (Q-Continue), with overall real-time delays reduced by 6–^^^^3ti:\3ti:\3^^^^× while preserving most reasoning accuracy (&&&^^^^3ti:\3^^^^&&&).

SHANKS interleaves streamed audio chunks with hidden reasoning blocks in spoken LLMs. Incoming speech is segmented into fixed-duration chunks marked with [EOPA] or [EOA], and the model emits <think>… blocks between chunks to support interruption and early tool invocation (&&&3^^^^3ti:\3^^^^&&&). On incorrect math solutions, Shanks-E^^^^3 OR all:\3^^^^E reaches 84.8% interruption ratio and 63.9% valid interrupt ratio versus ^^^^3ti:\33^^^^.8% and ^^^^3 OR all:\3^^^^6.8% for a no-thinking baseline, a 37.^^^^3ti:\3^^^^ pp improvement in valid-interrupt accuracy; in tool-augmented dialogue it completes 56.9% of calls before the user finishes speaking (&&&3^^^^3ti:\3^^^^&&&).

InterRS, introduced under the “thinking-while-speaking” paradigm, alternates <|thinking|> and <|answer|> states inside a unified decoder and trains with interleaved SFT plus RL using a TA-Balance Reward and a Linguistic Quality Reward (&&&^^^^3 OR all:\3^^^^&&&). On SATA-Bench, KK-audio, SMQA, and Meta-audio, InterRS reports 49.33% average accuracy, a ^^^^3ti:\3 OR all:\3^^^^.4 pp gain over Fast CoT–SFT+RL at 36.9^^^^3 OR all:\3^^^^%, while preserving instant response (&&&^^^^3 OR all:\3^^^^&&&).

Taken together, these systems indicate that InterleaveThinker is not an isolated architecture but one instance of a broader shift from monolithic generation to alternating internal and external trajectories.

6. Broader technical lineage and open directions

Outside multimodal generation, interleaving has long denoted a structural method for distributing local failures across a larger process. In coding theory, interleaved polar (i-polar) codes insert interleavers between intermediate stages of the polar encoder, preserving polarization while changing the code’s weight structure; the reported concatenated design with $u_i$^^^^3query3^^^^ outperforms a CRC-aided i-polar code with $u_i$^^^^3ti:\3^^^^ of the same length and rate at high SNR and allows multiple decoders to operate in parallel for URLLC (Chiu, 2019). In terrestrial FSO, data interleaving on a 4.6 km urban testbed yields a two-order-of-magnitude reduction in outage probability, with the achievable trade-off explicitly tied to turbulence strength, interleaver depth, and data rate (Gümüş et al., 15 Jun 2026). In short-packet OFDM for industrial environments, frame-level QPP and S-random interleavers improve PER by ^^^^3query3^^^^.5–^^^^3ti:\3 order of magnitude under moderate impulse interference (Zhan et al., 2022).

In formal methods, the combination of partial order reduction with state interpolation is presented as a way to reason about interleavings in concurrent program verification, pruning search while retaining soundness (Chu et al., 2014). In collaborative editing, FugueMax introduces maximal non-interleaving as a correctness criterion for replicated lists so that concurrent text insertions remain contiguous rather than being merged into unreadable character-level mixtures (Weidner et al., 2023). These are distinct from InterleaveThinker’s multimodal objective, but they show that interleaving has an established technical meaning: controlled rearrangement of dependency structure to reduce a specific mode of failure.

For the multimodal InterleaveThinker itself, the paper identifies several immediate extensions: video or 3D interleaved generation, integration with real-time embodied agents, larger or specialized vision-language backbones for both planner and critic, more sophisticated global planning objectives, and richer critic feedback such as region-level critiques (&&&^^^^3query3^^^^&&&). A plausible implication is that the framework’s lasting contribution is less a single benchmark number than a systems pattern: freeze a capable base generator, externalize trajectory planning, and use local corrective reasoning to stabilize long-horizon interleaved behavior.